Stage 2 завершён: problem-first ответы и follow-up continuity - ассистент переведён от entity-heavy логики к problem-first ответам с problem-unit слоем, удержанием контекста в follow-up и очисткой пользовательского ответа от сырых технических ссылок.

llm_normalizer/backend/dist/config.js

@@ -3,7 +3,7 @@ var __importDefault = (this && this.__importDefault) || function (mod) {
     return (mod && mod.__esModule) ? mod : { "default": mod };
 };
 Object.defineProperty(exports, "__esModule", { value: true });
-exports.ARCH_EXPORT_2020_DIR = exports.SCHEMAS_DIR = exports.EVAL_DATASETS_DIR = exports.REPORTS_DIR = exports.PROMPTS_DIR = exports.ASSISTANT_SESSIONS_DIR = exports.EVAL_CASES_DIR = exports.PRESETS_DIR = exports.TRACES_DIR = exports.DATA_DIR = exports.FEATURE_ASSISTANT_ACCOUNTANT_EVAL_V1 = exports.FEATURE_ASSISTANT_ANSWER_POLICY_V11 = exports.FEATURE_ASSISTANT_ANTI_GENERIC_RANKING_GUARD_V1 = exports.FEATURE_ASSISTANT_MIN_EVIDENCE_GATE_V1 = exports.FEATURE_ASSISTANT_BROAD_GUARD_V1 = exports.FEATURE_ASSISTANT_EVIDENCE_ENRICHMENT_V1 = exports.FEATURE_ASSISTANT_STATE_FOLLOWUP_BINDING_V1 = exports.FEATURE_ASSISTANT_CONTRACTS_V11 = exports.FEATURE_ASSISTANT_INVESTIGATION_STATE_V1 = exports.DEFAULT_PROMPT_VERSION = exports.DEFAULT_MAX_OUTPUT_TOKENS = exports.DEFAULT_TEMPERATURE = exports.DEFAULT_MODEL = exports.DEFAULT_OPENAI_BASE_URL = exports.TIMEZONE = exports.PORT = exports.MODULE_ROOT = exports.BACKEND_ROOT = void 0;
+exports.ARCH_EXPORT_2020_DIR = exports.SCHEMAS_DIR = exports.EVAL_DATASETS_DIR = exports.REPORTS_DIR = exports.PROMPTS_DIR = exports.ASSISTANT_SESSIONS_DIR = exports.EVAL_CASES_DIR = exports.PRESETS_DIR = exports.TRACES_DIR = exports.DATA_DIR = exports.FEATURE_ASSISTANT_STAGE2_EVAL_V1 = exports.FEATURE_ASSISTANT_PROBLEM_UNIT_CONTINUITY_V1 = exports.FEATURE_ASSISTANT_PROBLEM_CENTRIC_ANSWER_V1 = exports.FEATURE_ASSISTANT_PROBLEM_UNITS_V1 = exports.FEATURE_ASSISTANT_ACCOUNTANT_EVAL_V1 = exports.FEATURE_ASSISTANT_ANSWER_POLICY_V11 = exports.FEATURE_ASSISTANT_ANTI_GENERIC_RANKING_GUARD_V1 = exports.FEATURE_ASSISTANT_MIN_EVIDENCE_GATE_V1 = exports.FEATURE_ASSISTANT_BROAD_GUARD_V1 = exports.FEATURE_ASSISTANT_EVIDENCE_ENRICHMENT_V1 = exports.FEATURE_ASSISTANT_STATE_FOLLOWUP_BINDING_V1 = exports.FEATURE_ASSISTANT_CONTRACTS_V11 = exports.FEATURE_ASSISTANT_INVESTIGATION_STATE_V1 = exports.DEFAULT_PROMPT_VERSION = exports.DEFAULT_MAX_OUTPUT_TOKENS = exports.DEFAULT_TEMPERATURE = exports.DEFAULT_MODEL = exports.DEFAULT_OPENAI_BASE_URL = exports.TIMEZONE = exports.PORT = exports.MODULE_ROOT = exports.BACKEND_ROOT = void 0;
 const path_1 = __importDefault(require("path"));
 exports.BACKEND_ROOT = path_1.default.resolve(__dirname, "..");
 exports.MODULE_ROOT = path_1.default.resolve(exports.BACKEND_ROOT, "..");
@@ -30,6 +30,10 @@ exports.FEATURE_ASSISTANT_MIN_EVIDENCE_GATE_V1 = toBooleanFlag(process.env.FEATU
 exports.FEATURE_ASSISTANT_ANTI_GENERIC_RANKING_GUARD_V1 = toBooleanFlag(process.env.FEATURE_ASSISTANT_ANTI_GENERIC_RANKING_GUARD_V1, true);
 exports.FEATURE_ASSISTANT_ANSWER_POLICY_V11 = toBooleanFlag(process.env.FEATURE_ASSISTANT_ANSWER_POLICY_V11, false);
 exports.FEATURE_ASSISTANT_ACCOUNTANT_EVAL_V1 = toBooleanFlag(process.env.FEATURE_ASSISTANT_ACCOUNTANT_EVAL_V1, true);
+exports.FEATURE_ASSISTANT_PROBLEM_UNITS_V1 = toBooleanFlag(process.env.FEATURE_ASSISTANT_PROBLEM_UNITS_V1, false);
+exports.FEATURE_ASSISTANT_PROBLEM_CENTRIC_ANSWER_V1 = toBooleanFlag(process.env.FEATURE_ASSISTANT_PROBLEM_CENTRIC_ANSWER_V1, false);
+exports.FEATURE_ASSISTANT_PROBLEM_UNIT_CONTINUITY_V1 = toBooleanFlag(process.env.FEATURE_ASSISTANT_PROBLEM_UNIT_CONTINUITY_V1, false);
+exports.FEATURE_ASSISTANT_STAGE2_EVAL_V1 = toBooleanFlag(process.env.FEATURE_ASSISTANT_STAGE2_EVAL_V1, false);
 exports.DATA_DIR = process.env.DATA_DIR ?? path_1.default.resolve(exports.MODULE_ROOT, "data");
 exports.TRACES_DIR = path_1.default.resolve(exports.DATA_DIR, "traces");
 exports.PRESETS_DIR = path_1.default.resolve(exports.DATA_DIR, "presets");

llm_normalizer/backend/dist/services/answerComposer.js

@@ -10,21 +10,103 @@ function fallbackFromSummary(routeSummary) {
 function uniqueStrings(values, limit = 6) {
     return Array.from(new Set(values.map((item) => item.trim()).filter(Boolean))).slice(0, limit);
 }
+const UUID_PATTERN = /\b[0-9a-f]{8}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{12}\b/gi;
+const LONG_HEX_PATTERN = /\b[0-9a-f]{24,}\b/gi;
+const RAW_REF_BLOB_PATTERN = /\bevidence_source_ref_v1\|[^\s,;]+/gi;
+const RAW_REF_TOKEN_PATTERN = /\b(?:source_ref|canonical_ref|entity_id|fragment_id|guid|uuid)\b/gi;
+function looksLikeMojibake(value) {
+    const text = String(value ?? "");
+    if (!text.trim()) {
+        return false;
+    }
+    if (/(?:Р.|С.){5,}/u.test(text)) {
+        return true;
+    }
+    if (/[ЃѓЂђЌќЎў]/u.test(text)) {
+        return true;
+    }
+    return false;
+}
+function looksLikeTechnicalIdentifier(value) {
+    const text = String(value ?? "").trim();
+    if (!text) {
+        return false;
+    }
+    if (UUID_PATTERN.test(text)) {
+        UUID_PATTERN.lastIndex = 0;
+        return true;
+    }
+    UUID_PATTERN.lastIndex = 0;
+    if (LONG_HEX_PATTERN.test(text)) {
+        LONG_HEX_PATTERN.lastIndex = 0;
+        return true;
+    }
+    LONG_HEX_PATTERN.lastIndex = 0;
+    return /(?:evidence_source_ref_v1\||cmp%3a|batch_refresh_then_store:|^cmp:)/i.test(text);
+}
+function scrubRawTechnicalRefs(value) {
+    const raw = String(value ?? "").trim();
+    if (!raw) {
+        return "";
+    }
+    return raw
+        .replace(RAW_REF_BLOB_PATTERN, "linked source")
+        .replace(UUID_PATTERN, "[id]")
+        .replace(LONG_HEX_PATTERN, "[id]")
+        .replace(RAW_REF_TOKEN_PATTERN, "reference")
+        .replace(/\(\s*\[id\]\s*\)/g, "")
+        .replace(/\[\s*id\s*\](?:\s*,\s*\[\s*id\s*\])+/g, "[id]")
+        .replace(/\s{2,}/g, " ")
+        .trim();
+}
+function sanitizeUserFacingReply(value) {
+    return scrubRawTechnicalRefs(value)
+        .replace(/[ \t]+\n/g, "\n")
+        .replace(/\n{3,}/g, "\n\n")
+        .trim();
+}
+function sanitizeUserText(value) {
+    const normalized = scrubRawTechnicalRefs(String(value ?? "").replace(/\s+/g, " ").trim());
+    if (!normalized) {
+        return null;
+    }
+    if (looksLikeMojibake(normalized)) {
+        return null;
+    }
+    return normalized;
+}
+function sanitizeUserLines(values, limit = 6) {
+    const cleaned = values
+        .map((item) => sanitizeUserText(item))
+        .filter((item) => Boolean(item));
+    return uniqueStrings(cleaned, limit);
+}
 function formatList(items) {
     if (items.length === 0) {
         return "";
     }
     return items.map((item) => `- ${item}`).join("\n");
 }
+function formatSafeItemLine(entity, sourceId, riskScore) {
+    const entityLabel = sanitizeUserText(String(entity ?? "")) ?? "Record";
+    const idRaw = String(sourceId ?? "").trim();
+    const exposeId = idRaw.length > 0 && !looksLikeTechnicalIdentifier(idRaw);
+    const subject = exposeId ? `${entityLabel} (${idRaw})` : entityLabel;
+    if (riskScore !== undefined) {
+        return `${subject} - risk ${String(riskScore)}.`;
+    }
+    return `${subject}.`;
+}
 function extractTopFacts(results) {
     const lines = [];
     for (const result of results.filter((item) => item.status === "ok").slice(0, 3)) {
         if (result.result_type === "chain") {
             const top = result.items.slice(0, 3).map((item) => {
-                const counterparty = String(item.counterparty_id ?? "не указан");
+                const counterparty = String(item.counterparty_id ?? "").trim();
                 const operations = String(item.operations_count ?? "0");
                 const docs = String(item.document_refs_count ?? "0");
-                return `Контрагент ${counterparty}: операций ${operations}, документов в связке ${docs}.`;
+                const counterpartyLabel = counterparty.length > 0 && !looksLikeTechnicalIdentifier(counterparty) ? `Counterparty ${counterparty}` : "Counterparty";
+                return `${counterpartyLabel}: operations ${operations}, linked docs ${docs}.`;
             });
             lines.push(...top);
             continue;
@@ -32,41 +114,41 @@ function extractTopFacts(results) {
         if (result.result_type === "ranking") {
             const top = result.items
                 .slice(0, 5)
-                .map((item) => `${item.rank ?? "•"}. ${String(item.entity ?? "Сущность")} — ${String(item.records_count ?? 0)}.`);
+                .map((item) => `${item.rank ?? "*"}. ${String(item.entity ?? "Entity")} - ${String(item.records_count ?? 0)}.`);
             lines.push(...top);
             continue;
         }
         if (result.result_type === "list") {
             const top = result.items.slice(0, 5).map((item) => {
                 if (item.risk_score !== undefined) {
-                    return `${String(item.source_entity ?? "Запись")} (${String(item.source_id ?? "")}) — риск ${String(item.risk_score)}.`;
+                    return formatSafeItemLine(item.source_entity ?? "Record", item.source_id ?? "", item.risk_score);
                 }
-                return `${String(item.source_entity ?? "Запись")} (${String(item.source_id ?? "")}).`;
+                return formatSafeItemLine(item.source_entity ?? "Record", item.source_id ?? "");
             });
             lines.push(...top);
             continue;
         }
         const top = result.items
             .slice(0, 3)
-            .map((item) => `${String(item.source_entity ?? "Запись")} (${String(item.source_id ?? "")}).`);
+            .map((item) => formatSafeItemLine(item.source_entity ?? "Record", item.source_id ?? ""));
         lines.push(...top);
     }
-    return lines;
+    return sanitizeUserLines(lines, 8);
 }
 function extractWhyIncluded(results) {
-    return uniqueStrings(results.flatMap((item) => item.why_included));
+    return sanitizeUserLines(results.flatMap((item) => item.why_included));
 }
 function extractSelectionReasons(results) {
-    return uniqueStrings(results.flatMap((item) => item.selection_reason));
+    return sanitizeUserLines(results.flatMap((item) => item.selection_reason));
 }
 function extractRiskFactors(results) {
-    return uniqueStrings(results.flatMap((item) => item.risk_factors));
+    return sanitizeUserLines(results.flatMap((item) => item.risk_factors));
 }
 function extractBusinessInterpretation(results) {
-    return uniqueStrings(results.flatMap((item) => item.business_interpretation));
+    return sanitizeUserLines(results.flatMap((item) => item.business_interpretation));
 }
 function extractLimitations(results) {
-    return uniqueStrings(results.flatMap((item) => item.limitations));
+    return sanitizeUserLines(results.flatMap((item) => item.limitations), 10);
 }
 function summaryValue(result, key) {
     const summary = result.summary ?? {};
@@ -79,6 +161,55 @@ function summaryString(result, key) {
     const value = summaryValue(result, key);
     return typeof value === "string" ? value : null;
 }
+function summaryNumber(result, key) {
+    const value = summaryValue(result, key);
+    return typeof value === "number" && Number.isFinite(value) ? value : null;
+}
+function summaryStringArray(result, key) {
+    const value = summaryValue(result, key);
+    if (!Array.isArray(value)) {
+        return [];
+    }
+    return sanitizeUserLines(value.map((item) => String(item)), 6);
+}
+function buildFallbackWhyIncluded(results) {
+    const lines = [];
+    for (const result of results.slice(0, 2)) {
+        const routeFocus = summaryString(result, "route_focus");
+        const sourceRecords = summaryNumber(result, "source_records");
+        const filteredRecords = summaryNumber(result, "filtered_records_after_narrowing");
+        const checkedRecords = summaryNumber(result, "checked_records");
+        if (routeFocus) {
+            lines.push(`Проверка выполнена по профилю ${routeFocus}.`);
+        }
+        if (sourceRecords !== null && filteredRecords !== null && filteredRecords < sourceRecords) {
+            lines.push(`Применено сужение выборки: ${filteredRecords} из ${sourceRecords} записей.`);
+        }
+        if (checkedRecords !== null) {
+            lines.push(`Проверено записей в текущем проходе: ${checkedRecords}.`);
+        }
+    }
+    return sanitizeUserLines(lines, 4);
+}
+function buildFallbackSelectionReasons(results) {
+    const lines = [];
+    for (const result of results.slice(0, 2)) {
+        if (summaryBoolean(result, "semantic_narrowing_applied")) {
+            lines.push("Отбор выполнен по семантическому сужению предметной области.");
+        }
+        const rankingBasis = summaryStringArray(result, "ranking_basis");
+        if (rankingBasis.length > 0) {
+            lines.push(`Ранжирование основано на: ${rankingBasis.join(", ")}.`);
+        }
+        if (summaryBoolean(result, "broad_guard_applied")) {
+            lines.push("Применен broad-query guard для контроля ложной точности.");
+        }
+    }
+    if (lines.length === 0) {
+        lines.push("Отбор выполнен по совпадению предметных сигналов и доступной evidence-опоры.");
+    }
+    return sanitizeUserLines(lines, 4);
+}
 function suggestNextStep(requirements, coverage) {
     const next = [];
     if (coverage.clarification_needed_for.length > 0) {
@@ -95,9 +226,131 @@ function suggestNextStep(requirements, coverage) {
     }
     return next;
 }
+const PROBLEM_HEAVY_TYPES = new Set([
+    "document_conflict",
+    "broken_chain_segment",
+    "lifecycle_anomaly_node",
+    "unresolved_settlement_cluster",
+    "period_risk_cluster",
+    "cross_branch_inconsistency_cluster"
+]);
 function flattenEvidence(results) {
     return results.flatMap((item) => item.evidence);
 }
+function flattenProblemUnits(results) {
+    const units = [];
+    for (const result of results) {
+        if (!Array.isArray(result.problem_units)) {
+            continue;
+        }
+        units.push(...result.problem_units);
+    }
+    const byId = new Map();
+    for (const unit of units) {
+        byId.set(unit.problem_unit_id, unit);
+    }
+    return Array.from(byId.values());
+}
+function selectProblemUnitSummary(results) {
+    let selected = null;
+    for (const result of results) {
+        if (!result.problem_unit_summary) {
+            continue;
+        }
+        if (!selected || result.problem_unit_summary.units_total > selected.units_total) {
+            selected = result.problem_unit_summary;
+        }
+    }
+    return selected;
+}
+function formatAffectedScope(unit) {
+    const scopeParts = [];
+    if (unit.affected_accounts.length > 0) {
+        scopeParts.push(`счета: ${unit.affected_accounts.slice(0, 2).join(", ")}`);
+    }
+    if (unit.affected_counterparties.length > 0) {
+        scopeParts.push(`контрагенты: ${unit.affected_counterparties.slice(0, 2).join(", ")}`);
+    }
+    if (unit.affected_documents.length > 0) {
+        scopeParts.push(`документы: ${unit.affected_documents.slice(0, 2).join(", ")}`);
+    }
+    if (scopeParts.length === 0 && unit.affected_entities.length > 0) {
+        scopeParts.push(`объекты: ${unit.affected_entities.slice(0, 2).join(", ")}`);
+    }
+    if (scopeParts.length === 0) {
+        return "затронутый контур требует уточнения";
+    }
+    return scopeParts.join("; ");
+}
+function buildProblemCentricActions(input) {
+    const actions = [];
+    const unitTypes = new Set(input.units.map((item) => item.problem_unit_type));
+    if (unitTypes.has("broken_chain_segment")) {
+        actions.push("Проверьте связку выписка -> документ -> проводка по проблемным участкам цепочки.");
+    }
+    if (unitTypes.has("unresolved_settlement_cluster")) {
+        actions.push("Сверьте хвосты по расчетам: закрылся ли документ оплаты корректным закрывающим документом.");
+    }
+    if (unitTypes.has("period_risk_cluster")) {
+        actions.push("Оцените влияние дефекта на закрытие периода и корректность регламентных операций.");
+    }
+    if (unitTypes.has("cross_branch_inconsistency_cluster")) {
+        actions.push("Сверьте противоречия между документами, проводками и регистрами по НДС/межконтурным связям.");
+    }
+    if (unitTypes.has("lifecycle_anomaly_node")) {
+        actions.push("Проверьте lifecycle объекта: ожидаемый этап не должен оставаться в partially_linked состоянии.");
+    }
+    if (input.mode === "clarification_required") {
+        if (input.missingAnchors.period) {
+            actions.push("Уточните период проверки, чтобы зафиксировать границы проблемного контура.");
+        }
+        if (input.missingAnchors.account) {
+            actions.push("Уточните счет или группу счетов для предметной локализации дефекта.");
+        }
+        if (input.missingAnchors.documentOrObject) {
+            actions.push("Укажите конкретный документ или объект трассировки для проверки механизма отклонения.");
+        }
+        if (input.missingAnchors.counterparty) {
+            actions.push("Укажите контрагента/договор, чтобы проверить хвосты и разрывы на конкретной связке.");
+        }
+    }
+    if (input.coverageReport.requirements_uncovered.length > 0) {
+        actions.push(`Закройте непокрытые требования: ${input.coverageReport.requirements_uncovered.join(", ")}.`);
+    }
+    return uniqueStrings(actions, 6);
+}
+function buildProblemCentricClarifications(input) {
+    if (input.mode !== "clarification_required") {
+        return [];
+    }
+    const questions = [];
+    const unitTypes = new Set(input.units.map((item) => item.problem_unit_type));
+    if (input.missingAnchors.period) {
+        questions.push("Уточните период (например, 2020-06), в котором нужно проверить проблемный кластер.");
+    }
+    if (input.missingAnchors.account) {
+        questions.push("Уточните счет или связку счетов (например, 51/60), где вы ожидаете дефект.");
+    }
+    if (input.missingAnchors.documentOrObject) {
+        questions.push("Укажите документ/объект, от которого нужно строить проверку цепочки.");
+    }
+    if (input.missingAnchors.counterparty) {
+        questions.push("Укажите контрагента или договор, по которому проверить незакрытую экспозицию.");
+    }
+    if (unitTypes.has("broken_chain_segment")) {
+        questions.push("Уточните участок цепочки: выписка, платежный документ или проводка.");
+    }
+    if (unitTypes.has("period_risk_cluster")) {
+        questions.push("Уточните, какой этап закрытия периода критичен: начисление, закрытие счетов или НДС-блок.");
+    }
+    if (unitTypes.has("unresolved_settlement_cluster")) {
+        questions.push("Уточните, интересуют хвосты поставщиков, покупателей или оба направления.");
+    }
+    if (input.coverageReport.clarification_needed_for.length > 0) {
+        questions.push(`Закройте уточнения для требований: ${input.coverageReport.clarification_needed_for.join(", ")}.`);
+    }
+    return uniqueStrings(questions, 6);
+}
 function buildClaimEvidenceLinks(results) {
     const byClaim = new Map();
     for (const evidence of flattenEvidence(results)) {
@@ -246,7 +499,7 @@ function buildClarificationQuestions(input) {
 function buildRecommendedActions(input) {
     const actions = [];
     if (input.mode === "focused_grounded") {
-        actions.push("Проверьте 1-2 ключевые записи по source_ref и зафиксируйте итог в рабочем файле проверки.");
+        actions.push("Проверьте 1-2 ключевые записи в учетной базе и зафиксируйте итог в рабочем файле проверки.");
     }
     if (input.mode === "broad_partial") {
         actions.push("Сузьте запрос до периода + счета или периода + документа и повторите проверку.");
@@ -270,7 +523,7 @@ function buildRecommendedActions(input) {
         actions.push("Проверьте source mapping для связей document/register по указанным ref.");
     }
     if (input.sourceRefs.length > 0) {
-        actions.push(`Начните проверку с source_ref: ${input.sourceRefs.slice(0, 2).join(", ")}.`);
+        actions.push(`Начните проверку с ${input.sourceRefs.length} подтвержденных записей и сверьте их с первичными документами.`);
     }
     return uniqueStrings(actions, 6);
 }
@@ -403,6 +656,122 @@ function buildDirectAnswer(input) {
     }
     return "Не удалось сформировать обоснованный ответ; нужно уточнение запроса.";
 }
+function buildProblemCentricAnswerSummary(input) {
+    if (input.mode === "clarification_required") {
+        return "Выявлены проблемные кластеры, но для надежного вывода требуется предметное уточнение фокуса.";
+    }
+    if (input.weakUnits) {
+        return "Сформирован problem-centric срез с ограниченной опорой; вывод предварительный и требует до-проверки.";
+    }
+    if (input.summary?.units_total && input.summary.units_total > 1) {
+        return `Сформирован problem-centric срез: выделено ${input.summary.units_total} проблемных кластера с приоритетами.`;
+    }
+    return "Сформирован problem-centric срез: выделен ключевой проблемный кластер и затронутый контур.";
+}
+function buildProblemCentricDirectAnswer(input) {
+    const lead = input.mode === "clarification_required"
+        ? "Обнаружены проблемные зоны, но без уточнения якорей сильный factual-вывод преждевременен."
+        : input.weakUnits
+            ? "Выделены проблемные зоны с ограниченной надежностью; вывод дан в ограниченном режиме."
+            : "Выделены ключевые проблемные зоны и их влияние на учетный контур.";
+    const unitLines = input.units.map((unit) => {
+        const scope = formatAffectedScope(unit);
+        return `- ${unit.title}: ${unit.business_defect_class}; ${scope}; severity=${unit.severity.grade}, confidence=${unit.confidence.grade}.`;
+    });
+    if (unitLines.length === 0) {
+        return `${lead}\nПроблемные кластеры не удалось детализировать в текущем срезе.`;
+    }
+    return [lead, "Проблемные кластеры:", ...unitLines].join("\n");
+}
+function buildProblemCentricAnswerStructure(input) {
+    const weakUnits = input.selectedUnits.every((item) => item.confidence.grade === "low");
+    const unitMechanismNotes = uniqueStrings(input.selectedUnits
+        .map((item) => item.mechanism_summary)
+        .filter((item) => typeof item === "string" && item.trim().length > 0), 6);
+    const sourceRefs = uniqueStrings(input.evidenceItems
+        .map((item) => item.source_ref?.canonical_ref)
+        .filter((item) => typeof item === "string" && item.trim().length > 0), 6);
+    const evidenceIds = uniqueStrings(input.evidenceItems.map((item) => item.evidence_id), 10);
+    const mechanismStatus = unitMechanismNotes.length === 0
+        ? "unresolved"
+        : weakUnits || input.limitationReasonCodes.includes("missing_mechanism")
+            ? "limited"
+            : "grounded";
+    const problemSpecificLimitations = [];
+    if (weakUnits) {
+        problemSpecificLimitations.push("Problem units remain weak-confidence; conclusions are intentionally limited.");
+    }
+    if (input.problemSummary?.duplicate_collapses && input.problemSummary.duplicate_collapses > 0) {
+        problemSpecificLimitations.push("Part of the problem signal was merged due to duplicate collapse.");
+    }
+    const limitations = uniqueStrings([
+        ...problemSpecificLimitations,
+        ...input.limitationReasonCodes.map((code) => limitationReasonToText(code)),
+        ...extractLimitations(input.retrievalResults),
+        ...input.groundingCheck.reasons
+    ], 10);
+    const openUncertainties = uniqueStrings([
+        ...input.groundingCheck.missing_requirements,
+        ...(input.mode === "clarification_required" && input.missingAnchors.period ? ["missing_anchor:period"] : []),
+        ...(input.mode === "clarification_required" && input.missingAnchors.account ? ["missing_anchor:account"] : []),
+        ...(input.mode === "clarification_required" && input.missingAnchors.documentOrObject
+            ? ["missing_anchor:document_or_object"]
+            : []),
+        ...(input.mode === "clarification_required" && input.missingAnchors.counterparty ? ["missing_anchor:counterparty"] : [])
+    ], 8);
+    return {
+        schema_version: "answer_structure_v1_1",
+        answer_summary: buildProblemCentricAnswerSummary({
+            mode: input.mode,
+            weakUnits,
+            summary: input.problemSummary
+        }),
+        direct_answer: buildProblemCentricDirectAnswer({
+            mode: input.mode,
+            units: input.selectedUnits,
+            weakUnits
+        }),
+        mechanism_block: {
+            status: mechanismStatus,
+            mechanism_notes: unitMechanismNotes,
+            limitation_reason_codes: input.limitationReasonCodes
+        },
+        evidence_block: {
+            evidence_ids: evidenceIds,
+            source_refs: sourceRefs,
+            mechanism_notes: unitMechanismNotes,
+            coverage_note: input.coverageReport.requirements_total > 0 &&
+                input.coverageReport.requirements_total === input.coverageReport.requirements_covered &&
+                input.coverageReport.requirements_uncovered.length === 0 &&
+                input.coverageReport.requirements_partially_covered.length === 0
+                ? "coverage_full_or_near_full"
+                : "coverage_partial_or_limited",
+            ...(input.claimEvidenceLinks.length > 0
+                ? {
+                    claim_evidence_links: input.claimEvidenceLinks
+                }
+                : {})
+        },
+        uncertainty_block: {
+            open_uncertainties: openUncertainties,
+            limitations
+        },
+        next_step_block: {
+            recommended_actions: buildProblemCentricActions({
+                units: input.selectedUnits,
+                mode: input.mode,
+                missingAnchors: input.missingAnchors,
+                coverageReport: input.coverageReport
+            }),
+            clarification_questions: buildProblemCentricClarifications({
+                units: input.selectedUnits,
+                missingAnchors: input.missingAnchors,
+                coverageReport: input.coverageReport,
+                mode: input.mode
+            })
+        }
+    };
+}
 function renderPolicyReply(structure) {
     const mechanismLines = [`status=${structure.mechanism_block.status}`];
     if (structure.mechanism_block.mechanism_notes.length > 0) {
@@ -414,18 +783,18 @@ function renderPolicyReply(structure) {
     if (structure.mechanism_block.status === "unresolved" && structure.mechanism_block.mechanism_notes.length === 0) {
         mechanismLines.push("mechanism_note is intentionally omitted due to weak or missing mechanism evidence");
     }
+    const sourceRefCount = Array.isArray(structure.evidence_block.source_refs) ? structure.evidence_block.source_refs.length : 0;
+    const claimLinkCount = Array.isArray(structure.evidence_block.claim_evidence_links)
+        ? structure.evidence_block.claim_evidence_links.length
+        : 0;
     const evidenceLines = [
         `coverage=${structure.evidence_block.coverage_note}`,
-        `evidence_ids=${structure.evidence_block.evidence_ids.length > 0 ? structure.evidence_block.evidence_ids.join(", ") : "none"}`
+        `supporting_evidence_count=${structure.evidence_block.evidence_ids.length}`,
+        `supporting_source_count=${sourceRefCount}`,
+        `claim_support_links=${claimLinkCount}`
     ];
-    if (Array.isArray(structure.evidence_block.source_refs) && structure.evidence_block.source_refs.length > 0) {
-        evidenceLines.push(`source_refs=${structure.evidence_block.source_refs.join(", ")}`);
-    }
-    if (Array.isArray(structure.evidence_block.claim_evidence_links) && structure.evidence_block.claim_evidence_links.length > 0) {
-        const compactLinks = structure.evidence_block.claim_evidence_links
-            .slice(0, 4)
-            .map((item) => `${item.claim_ref}:${item.evidence_ids.join("|")}`);
-        evidenceLines.push(`claim_evidence_links=${compactLinks.join("; ")}`);
+    if (sourceRefCount > 0) {
+        evidenceLines.push("Detailed source references are available in debug payload.");
     }
     const uncertaintyLines = [
         ...structure.uncertainty_block.open_uncertainties.map((item) => `open: ${item}`),
@@ -441,7 +810,7 @@ function renderPolicyReply(structure) {
     if (nextStepLines.length === 0) {
         nextStepLines.push("No additional action is required for this scoped answer.");
     }
-    return [
+    return sanitizeUserFacingReply([
         `Answer summary: ${structure.answer_summary}`,
         `Direct answer:\n${structure.direct_answer}`,
         `Mechanism block:\n${formatList(mechanismLines)}`,
@@ -450,7 +819,7 @@ function renderPolicyReply(structure) {
         `Next step block:\n${formatList(nextStepLines)}`
     ]
         .filter(Boolean)
-        .join("\n\n");
+        .join("\n\n"));
 }
 function composeAssistantAnswerV11(input) {
     const fallbackType = fallbackFromSummary(input.routeSummary);
@@ -467,8 +836,15 @@ function composeAssistantAnswerV11(input) {
     const mechanismNotes = uniqueStrings(evidenceItems
         .map((item) => item.mechanism_note)
         .filter((item) => typeof item === "string" && item.trim().length > 0), 6);
+    const problemUnits = flattenProblemUnits(input.retrievalResults);
+    const problemUnitSummary = selectProblemUnitSummary(input.retrievalResults);
+    const problemHeavyUnits = problemUnits.filter((item) => PROBLEM_HEAVY_TYPES.has(item.problem_unit_type));
+    const selectedProblemUnits = problemHeavyUnits.slice(0, 4);
     const claimEvidenceLinks = buildClaimEvidenceLinks(input.retrievalResults);
     const aggregateEvidenceConfidence = aggregateConfidence(input.retrievalResults, evidenceItems);
+    const lowConfidenceSignals = evidenceItems.filter((item) => item.confidence === "low").length;
+    const lowConfidenceShare = evidenceItems.length > 0 ? lowConfidenceSignals / evidenceItems.length : 0;
+    const lowConfidenceConcentration = lowConfidenceShare >= 0.6;
     const hasSupport = okResults.length > 0 ||
         partialResults.length > 0 ||
         evidenceItems.length > 0 ||
@@ -501,6 +877,45 @@ function composeAssistantAnswerV11(input) {
         policySignals
     });
     const missingAnchors = detectMissingAnchors(input.userMessage);
+    const hasProblemWeakSignal = policySignals.narrowing_strength !== "strong" ||
+        policySignals.minimum_evidence_failed ||
+        limitationReasonCodes.includes("missing_mechanism") ||
+        limitationReasonCodes.includes("weak_source_mapping") ||
+        aggregateEvidenceConfidence === "low" ||
+        lowConfidenceConcentration;
+    const hardBlockedMode = decision.mode === "out_of_scope" || decision.mode === "route_mismatch" || decision.mode === "backend_error";
+    const problemCentricModeEligible = decision.mode === "broad_partial" ||
+        decision.mode === "clarification_required" ||
+        (decision.mode === "focused_grounded" && hasProblemWeakSignal);
+    const shouldUseProblemCentricAnswer = Boolean(input.enableProblemCentricAnswerV1) &&
+        !hardBlockedMode &&
+        problemCentricModeEligible &&
+        (!focusedStrong || hasProblemWeakSignal) &&
+        selectedProblemUnits.length > 0;
+    if (shouldUseProblemCentricAnswer) {
+        const problemCentricStructure = buildProblemCentricAnswerStructure({
+            mode: decision.mode,
+            selectedUnits: selectedProblemUnits,
+            problemSummary: problemUnitSummary,
+            evidenceItems,
+            claimEvidenceLinks,
+            limitationReasonCodes,
+            groundingCheck: input.groundingCheck,
+            retrievalResults: input.retrievalResults,
+            missingAnchors,
+            coverageReport: input.coverageReport
+        });
+        return {
+            assistant_reply: renderPolicyReply(problemCentricStructure),
+            fallback_type: decision.fallback_type,
+            reply_type: decision.reply_type,
+            answer_structure_v11: problemCentricStructure,
+            problem_centric_answer_applied: true,
+            problem_units_used_count: selectedProblemUnits.length,
+            problem_answer_mode: "stage2_problem_centric_v1",
+            problem_unit_ids_used: selectedProblemUnits.map((item) => item.problem_unit_id)
+        };
+    }
     const clarificationQuestions = buildClarificationQuestions({
         mode: decision.mode,
         missingAnchors,
@@ -577,13 +992,18 @@ function composeAssistantAnswerV11(input) {
         assistant_reply: renderPolicyReply(answerStructure),
         fallback_type: decision.fallback_type,
         reply_type: decision.reply_type,
-        answer_structure_v11: answerStructure
+        answer_structure_v11: answerStructure,
+        problem_centric_answer_applied: false,
+        problem_units_used_count: 0,
+        problem_answer_mode: "stage1_policy_v11"
     };
 }
 function composeExplainableAnswer(input, scopeLabel) {
     const facts = extractTopFacts(input.retrievalResults);
-    const whyIncluded = extractWhyIncluded(input.retrievalResults);
-    const selectionReasons = extractSelectionReasons(input.retrievalResults);
+    const whyIncludedRaw = extractWhyIncluded(input.retrievalResults);
+    const selectionReasonsRaw = extractSelectionReasons(input.retrievalResults);
+    const whyIncluded = whyIncludedRaw.length > 0 ? whyIncludedRaw : buildFallbackWhyIncluded(input.retrievalResults);
+    const selectionReasons = selectionReasonsRaw.length > 0 ? selectionReasonsRaw : buildFallbackSelectionReasons(input.retrievalResults);
     const riskFactors = extractRiskFactors(input.retrievalResults);
     const interpretation = extractBusinessInterpretation(input.retrievalResults);
     const limitations = uniqueStrings([...extractLimitations(input.retrievalResults), ...input.groundingCheck.reasons]);

llm_normalizer/backend/dist/services/assistantService.js

@@ -638,17 +638,17 @@ function hasAccountingSignal(text) {
     if (/(?:^|[\s,;:])\d{2}(?:\.\d{2})?(?=$|[\s,.;:])/i.test(lower)) {
         return true;
     }
-    return /(проводк|документ|реализац|поступлен|взаиморасчет|сальдо|остатк|счет|ндс|амортиз|рбп|ос|контрагент|поставщик|покупател|оплат|банк|выписк|склад|товар|материал|counterparty|supplier|invoice|posting|ledger|account|anomaly|risk)/i.test(lower);
+    return /(РїСЂРѕРІРѕРґРє|документ|реализац|поступлен|взаиморасчет|сальдо|остатк|счет|РЅРґСЃ|амортиз|СЂР±Рї|РѕСЃ|контрагент|поставщик|покупател|оплат|банк|выписк|склад|товар|материал|проводк|документ|реализац|поступлен|взаиморасчет|сальдо|остатк|счет|счёт|ндс|амортиз|рбп|контрагент|поставщик|покупател|оплат|банк|выписк|склад|товар|материал|закрыти|период|postavshchik|kontragent|schet|schetu|period|counterparty|supplier|invoice|posting|ledger|account|anomaly|risk)/i.test(lower);
 }
 function hasFollowupMarker(text) {
     const compact = compactWhitespace(text.toLowerCase());
-    return /^(и|а еще|а ещё|еще|ещё|добав|уточн|продолж|также|plus|also|dobav|utochn|prodolzh)/i.test(compact);
+    return /^(Рё|Р° еще|Р° ещё|еще|ещё|добав|уточн|продолж|также|и|а если|а еще|а ещё|еще|ещё|добав|уточн|продолж|также|plus|also|dobav|utochn|prodolzh)/i.test(compact);
 }
 function hasReferentialPointer(text) {
-    return /(по этому|по тому|это же|этой|этим|тому|same thing|that one|po etomu|po tomu)/i.test(text.toLowerCase());
+    return /(РїРѕ этому|РїРѕ тому|это же|этой|этим|тому|по этому|по тому|это же|этой|этим|этому|из этого|в этом|тот же|same thing|that one|po etomu|po tomu)/i.test(text.toLowerCase());
 }
 function hasSmallTalkSignal(text) {
-    return /(привет|как дела|спасибо|thanks|thank you|hello|hi)\b/i.test(text.toLowerCase());
+    return /(привет|как дела|спасибо|привет|как дела|спасибо|благодарю|thanks|thank you|hello|hi)\b/i.test(text.toLowerCase());
 }
 function countTokens(text) {
     return compactWhitespace(text)
@@ -658,6 +658,37 @@ function countTokens(text) {
 function hasPeriodLiteral(text) {
     return /\b(20\d{2}(?:[-/.](?:0[1-9]|1[0-2]))?)\b/.test(text);
 }
+function extractNormalizedPeriodLiteral(text) {
+    const monthly = text.match(/\b(20\d{2})[-/.](0[1-9]|1[0-2])\b/);
+    if (monthly) {
+        return `${monthly[1]}-${monthly[2]}`;
+    }
+    const yearly = text.match(/\b(20\d{2})\b/);
+    if (yearly) {
+        return yearly[1];
+    }
+    return null;
+}
+function hasStrongFollowupAnchors(userMessage, state) {
+    const explicitPeriod = extractNormalizedPeriodLiteral(userMessage);
+    if (explicitPeriod && state.focus.period && explicitPeriod !== state.focus.period) {
+        const periodLooksLikeFollowupRefinement = hasFollowupMarker(userMessage) || hasReferentialPointer(userMessage);
+        if (!periodLooksLikeFollowupRefinement) {
+            return true;
+        }
+    }
+    const explicitAccounts = extractAccountTokens(userMessage);
+    if (explicitAccounts.length > 0) {
+        const knownAccounts = new Set(state.focus.primary_accounts.map((item) => item.trim()));
+        if (knownAccounts.size === 0) {
+            return true;
+        }
+        if (explicitAccounts.some((item) => !knownAccounts.has(item))) {
+            return true;
+        }
+    }
+    return false;
+}
 function routeFromInvestigationState(state) {
     const rawDomain = compactWhitespace(state.focus.domain ?? "");
     if (!rawDomain) {
@@ -699,7 +730,15 @@ function buildFollowupStateBinding(input) {
     const referentialPointer = hasReferentialPointer(userMessage);
     const shortPrompt = countTokens(userMessage) <= 10;
     const smallTalkSignal = hasSmallTalkSignal(userMessage);
-    const shouldBind = !smallTalkSignal && (followupMarker || referentialPointer || (!strongSignal && shortPrompt));
+    const problemState = input.investigationState.problem_unit_state;
+    const problemContinuityAvailable = config_1.FEATURE_ASSISTANT_PROBLEM_UNIT_CONTINUITY_V1 &&
+        Boolean(problemState) &&
+        ((problemState?.active_problem_units.length ?? 0) > 0 || (problemState?.focus_problem_types.length ?? 0) > 0);
+    const strongNewAnchorDetected = hasStrongFollowupAnchors(userMessage, input.investigationState);
+    const periodRefinementFollowup = hasPeriodLiteral(userMessage) && problemContinuityAvailable;
+    const shouldBind = !smallTalkSignal &&
+        !strongNewAnchorDetected &&
+        (followupMarker || referentialPointer || periodRefinementFollowup || (!strongSignal && shortPrompt));
     if (!shouldBind) {
         return {
             normalizedQuestion: userMessage,
@@ -710,6 +749,7 @@ function buildFollowupStateBinding(input) {
     const context = {
         ...(input.payloadContext ?? {})
     };
+    const hasExplicitExpectedRoute = Boolean(input.payloadContext?.expected_route);
     const expectedRouteFromState = !context?.expected_route ? routeFromInvestigationState(input.investigationState) : null;
     const periodHintFromState = !context?.period_hint ? input.investigationState.focus.period : null;
     if (expectedRouteFromState) {
@@ -720,6 +760,8 @@ function buildFollowupStateBinding(input) {
     }
     const subject = withCappedLength(compactWhitespace(input.investigationState.focus.active_query_subject ?? ""), FOLLOWUP_SUBJECT_MAX);
     const businessContextPatch = ["followup_state_binding_v1"];
+    let problemContinuityApplied = false;
+    let problemContinuitySkippedReason = null;
     if (input.investigationState.focus.period) {
         businessContextPatch.push("active_period");
     }
@@ -729,6 +771,21 @@ function buildFollowupStateBinding(input) {
     if (input.investigationState.focus.primary_accounts.length > 0) {
         businessContextPatch.push(`focus_accounts:${input.investigationState.focus.primary_accounts.join(",")}`);
     }
+    if (problemContinuityAvailable) {
+        if (hasExplicitExpectedRoute) {
+            problemContinuitySkippedReason = "explicit_expected_route";
+        }
+        else {
+            const focusTypes = (problemState?.focus_problem_types ?? []).slice(0, 3);
+            const activeCount = problemState?.active_problem_units.length ?? 0;
+            businessContextPatch.push("problem_unit_continuity_v1");
+            if (focusTypes.length > 0) {
+                businessContextPatch.push(`problem_focus_types:${focusTypes.join(",")}`);
+            }
+            businessContextPatch.push(`problem_active_count:${activeCount}`);
+            problemContinuityApplied = true;
+        }
+    }
     const mergedBusinessContext = mergeBusinessContext(context?.business_context, businessContextPatch);
     if (mergedBusinessContext) {
         context.business_context = mergedBusinessContext;
@@ -743,6 +800,9 @@ function buildFollowupStateBinding(input) {
         if (periodHintFromState && !hasPeriodLiteral(userMessage)) {
             appendParts.push(`Период фокуса: ${periodHintFromState}`);
         }
+        if (problemContinuityApplied && (problemState?.focus_problem_types.length ?? 0) > 0) {
+            appendParts.push(`Problem focus types: ${(problemState?.focus_problem_types ?? []).slice(0, 3).join(", ")}`);
+        }
         const appendBlock = withCappedLength(compactWhitespace(appendParts.join("; ")), FOLLOWUP_QUESTION_APPEND_MAX);
         normalizedQuestion = `${userMessage}\n${appendBlock}`.trim();
     }
@@ -762,7 +822,11 @@ function buildFollowupStateBinding(input) {
                 period_hint_from_state: Boolean(periodHintFromState),
                 expected_route_from_state: Boolean(expectedRouteFromState),
                 business_context_from_state: Boolean(mergedBusinessContext),
-                question_augmented: shouldAugmentQuestion
+                question_augmented: shouldAugmentQuestion,
+                problem_continuity_available: problemContinuityAvailable,
+                problem_continuity_applied: problemContinuityApplied,
+                problem_continuity_skipped_reason: problemContinuityApplied ? null : problemContinuitySkippedReason,
+                strong_new_anchor_detected: strongNewAnchorDetected
             }
         }
     };
@@ -903,7 +967,8 @@ class AssistantService {
             requirements: coverageEvaluation.requirements,
             coverageReport: coverageEvaluation.coverage,
             groundingCheck,
-            enableAnswerPolicyV11: config_1.FEATURE_ASSISTANT_ANSWER_POLICY_V11
+            enableAnswerPolicyV11: config_1.FEATURE_ASSISTANT_ANSWER_POLICY_V11,
+            enableProblemCentricAnswerV1: config_1.FEATURE_ASSISTANT_PROBLEM_CENTRIC_ANSWER_V1
         });
         const answerStructureV11 = config_1.FEATURE_ASSISTANT_CONTRACTS_V11
             ? config_1.FEATURE_ASSISTANT_ANSWER_POLICY_V11 && composition.answer_structure_v11
@@ -952,6 +1017,14 @@ class AssistantService {
             answer_grounding_check: groundingCheck,
             dropped_intent_segments: extractDiscardedIntentSegments(normalized.normalized),
             ...(followupBinding.usage ? { followup_state_usage: followupBinding.usage } : {}),
+            problem_centric_answer_applied: composition.problem_centric_answer_applied ?? false,
+            problem_units_used_count: composition.problem_units_used_count ?? 0,
+            problem_answer_mode: composition.problem_answer_mode ?? "stage1_policy_v11",
+            ...(Array.isArray(composition.problem_unit_ids_used) && composition.problem_unit_ids_used.length > 0
+                ? {
+                    problem_unit_ids_used: composition.problem_unit_ids_used
+                }
+                : {}),
             answer_structure_v11: answerStructureV11,
             investigation_state_snapshot: investigationStateSnapshot,
             normalized: normalized.normalized
@@ -1007,6 +1080,14 @@ class AssistantService {
                 clarification_target: coverageEvaluation.coverage.clarification_needed_for,
                 dropped_intent_segments: extractDiscardedIntentSegments(normalized.normalized),
                 ...(followupBinding.usage ? { followup_state_usage: followupBinding.usage } : {}),
+                problem_centric_answer_applied: composition.problem_centric_answer_applied ?? false,
+                problem_units_used_count: composition.problem_units_used_count ?? 0,
+                problem_answer_mode: composition.problem_answer_mode ?? "stage1_policy_v11",
+                ...(Array.isArray(composition.problem_unit_ids_used) && composition.problem_unit_ids_used.length > 0
+                    ? {
+                        problem_unit_ids_used: composition.problem_unit_ids_used
+                    }
+                    : {}),
                 answer_structure_v11: answerStructureV11,
                 investigation_state_snapshot: investigationStateSnapshot,
                 fallback_type: composition.fallback_type,

llm_normalizer/backend/dist/services/evalService.js

@@ -9,6 +9,7 @@ const path_1 = __importDefault(require("path"));
 const nanoid_1 = require("nanoid");
 const config_1 = require("../config");
 const stage1Contracts_1 = require("../types/stage1Contracts");
+const stage2EvalContracts_1 = require("../types/stage2EvalContracts");
 const http_1 = require("../utils/http");
 const assistantService_1 = require("./assistantService");
 const assistantSessionStore_1 = require("./assistantSessionStore");
@@ -214,6 +215,20 @@ function isDecisionStateConsistent(decision) {
 const DEFAULT_ASSISTANT_STAGE1_SUITE_FILE = "assistant_stage1_canonical_v0_1.json";
 const ASSISTANT_STAGE1_RUN_SCHEMA_VERSION = "assistant_stage1_eval_run_v0_1";
 const ASSISTANT_STAGE1_COMPARISON_SCHEMA_VERSION = "assistant_stage1_eval_comparison_v0_1";
+const DEFAULT_ASSISTANT_STAGE2_SUITE_FILE = "assistant_stage2_canonical_v0_1.json";
+const ASSISTANT_STAGE2_RUN_SCHEMA_VERSION = "assistant_stage2_eval_run_v0_1";
+const ASSISTANT_STAGE2_COMPARISON_SCHEMA_VERSION = "assistant_stage2_eval_comparison_v0_1";
+const KNOWN_PROBLEM_UNIT_TYPES = [
+    "document_conflict",
+    "broken_chain_segment",
+    "lifecycle_anomaly_node",
+    "unresolved_settlement_cluster",
+    "period_risk_cluster",
+    "cross_branch_inconsistency_cluster"
+];
+function toProblemUnitType(value) {
+    return KNOWN_PROBLEM_UNIT_TYPES.includes(value) ? value : null;
+}
 function round2(value) {
     return Number(value.toFixed(2));
 }
@@ -258,6 +273,44 @@ function rubricBandForMetric(metric, value) {
     const score = rateToBandScore(metric, value);
     return stage1Contracts_1.ACCOUNTANT_SCORING_RUBRIC_V01[metric].find((item) => item.score === score) ?? null;
 }
+function rateToBandScoreStage2(metric, value) {
+    if (metric === "problem_unit_precision" || metric === "problem_unit_recall_proxy" || metric === "problem_first_answer_rate") {
+        if (value >= 0.75)
+            return 5;
+        if (value >= 0.45)
+            return 3;
+        return 0;
+    }
+    if (metric === "duplicate_collapse_rate") {
+        if (value >= 0.2)
+            return 5;
+        if (value >= 0.08)
+            return 3;
+        return 0;
+    }
+    if (metric === "entity_leakage_rate") {
+        if (value <= 0.2)
+            return 5;
+        if (value <= 0.4)
+            return 3;
+        return 0;
+    }
+    if (metric === "mechanism_coherence_score" || metric === "problem_clarity_score") {
+        if (value >= 4)
+            return 5;
+        if (value >= 2.5)
+            return 3;
+        return 0;
+    }
+    return 0;
+}
+function rubricBandForMetricStage2(metric, value) {
+    if (value === null) {
+        return null;
+    }
+    const score = rateToBandScoreStage2(metric, value);
+    return stage2EvalContracts_1.ASSISTANT_STAGE2_SCORING_RUBRIC_V01[metric].find((item) => item.score === score) ?? null;
+}
 function buildFeatureProfileSnapshot() {
     return {
         FEATURE_ASSISTANT_ACCOUNTANT_EVAL_V1: config_1.FEATURE_ASSISTANT_ACCOUNTANT_EVAL_V1,
@@ -266,7 +319,11 @@ function buildFeatureProfileSnapshot() {
         FEATURE_ASSISTANT_MIN_EVIDENCE_GATE_V1: process.env.FEATURE_ASSISTANT_MIN_EVIDENCE_GATE_V1 ?? null,
         FEATURE_ASSISTANT_ANTI_GENERIC_RANKING_GUARD_V1: process.env.FEATURE_ASSISTANT_ANTI_GENERIC_RANKING_GUARD_V1 ?? null,
         FEATURE_ASSISTANT_INVESTIGATION_STATE_V1: process.env.FEATURE_ASSISTANT_INVESTIGATION_STATE_V1 ?? null,
-        FEATURE_ASSISTANT_STATE_FOLLOWUP_BINDING_V1: process.env.FEATURE_ASSISTANT_STATE_FOLLOWUP_BINDING_V1 ?? null
+        FEATURE_ASSISTANT_STATE_FOLLOWUP_BINDING_V1: process.env.FEATURE_ASSISTANT_STATE_FOLLOWUP_BINDING_V1 ?? null,
+        FEATURE_ASSISTANT_PROBLEM_UNITS_V1: process.env.FEATURE_ASSISTANT_PROBLEM_UNITS_V1 ?? String(config_1.FEATURE_ASSISTANT_PROBLEM_UNITS_V1),
+        FEATURE_ASSISTANT_PROBLEM_CENTRIC_ANSWER_V1: process.env.FEATURE_ASSISTANT_PROBLEM_CENTRIC_ANSWER_V1 ?? String(config_1.FEATURE_ASSISTANT_PROBLEM_CENTRIC_ANSWER_V1),
+        FEATURE_ASSISTANT_PROBLEM_UNIT_CONTINUITY_V1: process.env.FEATURE_ASSISTANT_PROBLEM_UNIT_CONTINUITY_V1 ?? String(config_1.FEATURE_ASSISTANT_PROBLEM_UNIT_CONTINUITY_V1),
+        FEATURE_ASSISTANT_STAGE2_EVAL_V1: process.env.FEATURE_ASSISTANT_STAGE2_EVAL_V1 ?? String(config_1.FEATURE_ASSISTANT_STAGE2_EVAL_V1)
     };
 }
 function buildCodeVersionMarker() {
@@ -331,6 +388,41 @@ function parseAssistantSuiteFile(inputPath) {
     }
     return parsed;
 }
+function parseAssistantStage2SuiteFile(inputPath) {
+    const filePath = resolveReadablePath(inputPath ?? DEFAULT_ASSISTANT_STAGE2_SUITE_FILE);
+    const raw = fs_1.default.readFileSync(filePath, "utf-8").replace(/^\uFEFF/, "");
+    const parsed = JSON.parse(raw);
+    if (!parsed || typeof parsed !== "object") {
+        throw new Error(`Invalid assistant stage2 suite format: ${filePath}`);
+    }
+    if (!Array.isArray(parsed.cases)) {
+        throw new Error(`Assistant stage2 suite cases[] is required: ${filePath}`);
+    }
+    if (!Array.isArray(parsed.case_ids)) {
+        throw new Error(`Assistant stage2 suite case_ids[] is required: ${filePath}`);
+    }
+    if (typeof parsed.suite_id !== "string" || !parsed.suite_id.trim()) {
+        throw new Error(`Assistant stage2 suite_id is required: ${filePath}`);
+    }
+    if (typeof parsed.suite_version !== "string" || !parsed.suite_version.trim()) {
+        throw new Error(`Assistant stage2 suite_version is required: ${filePath}`);
+    }
+    if (parsed.scenario_count !== parsed.cases.length) {
+        throw new Error(`Assistant stage2 scenario_count mismatch: ${filePath}`);
+    }
+    const declaredIds = [...parsed.case_ids].sort();
+    const actualIds = parsed.cases.map((item) => item.case_id).sort();
+    const idsMatch = declaredIds.length === actualIds.length && declaredIds.every((item, index) => item === actualIds[index]);
+    if (!idsMatch) {
+        throw new Error(`Assistant stage2 case_ids do not match cases[]: ${filePath}`);
+    }
+    for (const item of parsed.cases) {
+        if (!Array.isArray(item.turns) || item.turns.length === 0) {
+            throw new Error(`Assistant stage2 case ${item.case_id} must include at least one turn.`);
+        }
+    }
+    return parsed;
+}
 function hasDomainAnchors(text) {
     const source = String(text ?? "");
     if (!source.trim()) {
@@ -342,6 +434,16 @@ function hasDomainAnchors(text) {
     const hits = [hasPeriod, hasAccountingObject, hasAccountCode].filter(Boolean).length;
     return hits >= 2;
 }
+function detectEntityLeakage(text) {
+    const source = String(text ?? "");
+    if (!source.trim()) {
+        return false;
+    }
+    const uuidHits = source.match(/\b[0-9a-f]{8}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{12}\b/gi)?.length ?? 0;
+    const guidHits = source.match(/\b(?:guid|uuid|entity_id|source_ref|canonical_ref|fragment_id)\b/gi)?.length ?? 0;
+    const longHexHits = source.match(/\b[0-9a-f]{24,}\b/gi)?.length ?? 0;
+    return uuidHits > 0 || guidHits > 1 || longHexHits > 0;
+}
 function extractTextList(value) {
     if (!Array.isArray(value)) {
         return [];
@@ -411,6 +513,50 @@ function buildAssistantEvalMarkdownReport(report) {
         ""
     ].join("\n");
 }
+function buildAssistantStage2EvalMarkdownReport(report) {
+    const metrics = (report.metrics ?? {}).raw ?? {};
+    const bands = (report.rubric_bands ?? {});
+    const subsets = (report.subsets ?? {});
+    const scenarioSummary = (report.scenario_summary ?? {});
+    const rows = Object.keys(metrics)
+        .map((key) => {
+        const rawValue = metrics[key];
+        const band = bands[key];
+        const rawPrintable = rawValue === null || rawValue === undefined ? "n/a" : String(rawValue);
+        const bandPrintable = band ? `${String(band.score)} (${String(band.label)})` : "n/a";
+        return `| ${key} | ${rawPrintable} | ${bandPrintable} |`;
+    })
+        .join("\n");
+    return [
+        `# ${String(report.report_title ?? "Assistant Stage 2 Eval Run")}`,
+        "",
+        `- run_id: ${String(report.run_id ?? "")}`,
+        `- eval_target: ${String(report.eval_target ?? "")}`,
+        `- run_timestamp: ${String(report.run_timestamp ?? "")}`,
+        `- suite_id: ${String(report.suite_id ?? "")}`,
+        `- suite_version: ${String(report.suite_version ?? "")}`,
+        `- cases_total: ${String(report.cases_total ?? 0)}`,
+        "",
+        "## Raw Metrics and Rubric Bands",
+        "",
+        "| Metric | Raw | Rubric band |",
+        "|---|---:|---|",
+        rows || "| n/a | n/a | n/a |",
+        "",
+        "## Subsets",
+        "",
+        `- expected_problem_cases_total: ${String(subsets.expected_problem_cases_total ?? 0)}`,
+        `- followup_cases_total: ${String(subsets.followup_cases_total ?? 0)}`,
+        `- candidate_cases_total: ${String(subsets.candidate_cases_total ?? 0)}`,
+        "",
+        "## Scenario Summary",
+        "",
+        `- improved_or_strong: ${String(scenarioSummary.improved_or_strong ?? 0)}`,
+        `- unchanged_or_mixed: ${String(scenarioSummary.unchanged_or_mixed ?? 0)}`,
+        `- weak_or_regressed: ${String(scenarioSummary.weak_or_regressed ?? 0)}`,
+        ""
+    ].join("\n");
+}
 function buildAssistantComparisonMarkdownReport(report) {
     const metrics = (report.metric_deltas ?? {});
     const summary = (report.scenario_notes_summary ?? {});
@@ -442,6 +588,37 @@ function buildAssistantComparisonMarkdownReport(report) {
         ""
     ].join("\n");
 }
+function buildAssistantStage2ComparisonMarkdownReport(report) {
+    const metrics = (report.metric_deltas ?? {});
+    const summary = (report.scenario_notes_summary ?? {});
+    const rows = Object.keys(metrics)
+        .map((key) => {
+        const row = metrics[key];
+        return `| ${key} | ${String(row.baseline ?? "n/a")} | ${String(row.current ?? "n/a")} | ${String(row.delta ?? "n/a")} | ${String(row.trend ?? "n/a")} |`;
+    })
+        .join("\n");
+    return [
+        `# ${String(report.report_title ?? "Assistant Stage 2 Baseline vs Current")}`,
+        "",
+        `- comparison_id: ${String(report.comparison_id ?? "")}`,
+        `- baseline_run_id: ${String(report.baseline_run_id ?? "")}`,
+        `- current_run_id: ${String(report.current_run_id ?? "")}`,
+        `- suite_version: ${String(report.suite_version ?? "")}`,
+        "",
+        "## Metric Deltas",
+        "",
+        "| Metric | Baseline | Current | Delta | Trend |",
+        "|---|---:|---:|---:|---|",
+        rows || "| n/a | n/a | n/a | n/a | n/a |",
+        "",
+        "## Scenario Notes Summary",
+        "",
+        `- improved: ${String(summary.improved ?? 0)}`,
+        `- unchanged: ${String(summary.unchanged ?? 0)}`,
+        `- weakened: ${String(summary.weakened ?? 0)}`,
+        ""
+    ].join("\n");
+}
 class EvalService {
     normalizerService;
     constructor(normalizerService) {
@@ -806,6 +983,148 @@ class EvalService {
             uncertainty_limitations_count: uncertaintyLimitationsCount
         };
     }
+    collectAssistantStage2Signals(finalResponse, turnResponses) {
+        const base = this.collectAssistantSignals(finalResponse, turnResponses);
+        const debug = finalResponse.debug;
+        const retrievalResults = Array.isArray(debug?.retrieval_results) ? debug.retrieval_results : [];
+        const typeSet = new Set();
+        const mechanismSummaries = new Set();
+        let candidateEvidenceTotal = 0;
+        let problemUnitsTotal = 0;
+        let duplicateCollapsesTotal = 0;
+        for (const result of retrievalResults) {
+            const candidates = Array.isArray(result.candidate_evidence) ? result.candidate_evidence : [];
+            candidateEvidenceTotal += candidates.length;
+            const problemUnits = Array.isArray(result.problem_units) ? result.problem_units : [];
+            problemUnitsTotal += problemUnits.length;
+            for (const unit of problemUnits) {
+                const unitType = toProblemUnitType(unit.problem_unit_type);
+                if (unitType) {
+                    typeSet.add(unitType);
+                }
+                const mechanismSummary = String(unit.mechanism_summary ?? "").trim();
+                if (mechanismSummary) {
+                    mechanismSummaries.add(mechanismSummary);
+                }
+            }
+            if (result.problem_unit_summary && typeof result.problem_unit_summary.duplicate_collapses === "number") {
+                duplicateCollapsesTotal += Number(result.problem_unit_summary.duplicate_collapses);
+            }
+        }
+        const answerMode = typeof debug?.problem_answer_mode === "string" ? debug.problem_answer_mode : null;
+        const unitsUsedCount = Number(debug?.problem_units_used_count ?? 0);
+        const unitIdsUsed = Array.isArray(debug?.problem_unit_ids_used)
+            ? debug.problem_unit_ids_used
+                .map((item) => String(item ?? "").trim())
+                .filter(Boolean)
+            : [];
+        const problemCentricApplied = debug?.problem_centric_answer_applied === true || answerMode === "stage2_problem_centric_v1";
+        return {
+            ...base,
+            candidate_evidence_total: candidateEvidenceTotal,
+            problem_units_total: problemUnitsTotal,
+            problem_unit_types: [...typeSet],
+            problem_mechanism_summaries: [...mechanismSummaries],
+            duplicate_collapses_total: duplicateCollapsesTotal,
+            problem_centric_answer_applied: problemCentricApplied,
+            problem_units_used_count: unitsUsedCount,
+            problem_answer_mode: answerMode,
+            problem_unit_ids_used: unitIdsUsed,
+            entity_leakage_detected: detectEntityLeakage(String(finalResponse.assistant_reply ?? ""))
+        };
+    }
+    getExpectedProblemUnitTypes(suiteCase) {
+        const expected = Array.isArray(suiteCase.expected_hints?.expected_problem_unit_types)
+            ? suiteCase.expected_hints?.expected_problem_unit_types
+            : [];
+        const output = new Set();
+        for (const value of expected ?? []) {
+            const mapped = toProblemUnitType(value);
+            if (mapped) {
+                output.add(mapped);
+            }
+        }
+        return [...output];
+    }
+    computeProblemUnitPrecision(expectedTypes, detectedTypes) {
+        const uniqueExpected = [...new Set(expectedTypes)];
+        const uniqueDetected = [...new Set(detectedTypes)];
+        if (uniqueDetected.length === 0) {
+            return uniqueExpected.length === 0 ? 1 : 0;
+        }
+        if (uniqueExpected.length === 0) {
+            return 0;
+        }
+        const matchedDetected = uniqueDetected.filter((item) => uniqueExpected.includes(item)).length;
+        return round2(matchedDetected / uniqueDetected.length);
+    }
+    computeProblemUnitRecallProxy(expectedTypes, detectedTypes) {
+        const uniqueExpected = [...new Set(expectedTypes)];
+        const uniqueDetected = [...new Set(detectedTypes)];
+        if (uniqueExpected.length === 0) {
+            return null;
+        }
+        if (uniqueDetected.length === 0) {
+            return 0;
+        }
+        const matchedExpected = uniqueExpected.filter((item) => uniqueDetected.includes(item)).length;
+        return round2(matchedExpected / uniqueExpected.length);
+    }
+    computeDuplicateCollapseRate(candidateTotal, duplicateCollapses) {
+        if (candidateTotal <= 0) {
+            return null;
+        }
+        return round2(Math.min(1, Math.max(0, duplicateCollapses / candidateTotal)));
+    }
+    computeMechanismCoherenceScore(finalResponse, signals) {
+        const mechanismBlock = finalResponse.debug?.answer_structure_v11?.mechanism_block;
+        const mechanismStatus = mechanismBlock?.status;
+        const mechanismNotes = extractTextList(mechanismBlock?.mechanism_notes);
+        const hasProblemMechanism = signals.problem_mechanism_summaries.length > 0;
+        let score = 0;
+        if (mechanismStatus === "grounded" && hasProblemMechanism && mechanismNotes.length > 0) {
+            score = 5;
+        }
+        else if ((mechanismStatus === "limited" || mechanismStatus === "unresolved") && (hasProblemMechanism || mechanismNotes.length > 0)) {
+            score = 3;
+        }
+        else if (hasProblemMechanism || mechanismNotes.length > 0) {
+            score = 2;
+        }
+        if (mechanismStatus === "grounded" && !hasProblemMechanism) {
+            score = Math.min(score, 2);
+        }
+        if (signals.limitation_reason_codes.includes("missing_mechanism")) {
+            score -= 1;
+        }
+        return clampScore(score);
+    }
+    computeProblemClarityScore(finalResponse, signals) {
+        const structure = finalResponse.debug?.answer_structure_v11;
+        const answerSummary = String(structure?.answer_summary ?? "").trim();
+        const directAnswer = String(structure?.direct_answer ?? finalResponse.assistant_reply ?? "").trim();
+        const recommendedActions = extractTextList(structure?.next_step_block?.recommended_actions);
+        const clarificationQuestions = extractTextList(structure?.next_step_block?.clarification_questions);
+        const uncertaintyLimitations = extractTextList(structure?.uncertainty_block?.limitations);
+        let score = 0;
+        if (answerSummary.length > 20)
+            score += 1;
+        if (directAnswer.length > 20)
+            score += 1;
+        if (hasDomainAnchors(`${answerSummary} ${directAnswer}`))
+            score += 1;
+        if (recommendedActions.length > 0 || clarificationQuestions.length > 0)
+            score += 1;
+        if (signals.problem_units_total > 0 || signals.problem_centric_answer_applied)
+            score += 1;
+        if ((signals.minimum_evidence_failed || signals.degraded_to === "clarification") && uncertaintyLimitations.length === 0) {
+            score -= 1;
+        }
+        if (signals.entity_leakage_detected) {
+            score -= 1;
+        }
+        return clampScore(score);
+    }
     computeAssistantMetrics(input) {
         const diagnostics = input.diagnostics;
         const total = Math.max(1, diagnostics.length);
@@ -855,6 +1174,68 @@ class EvalService {
             signature_counts: signatureCounter
         };
     }
+    computeAssistantStage2Metrics(input) {
+        const diagnostics = input.diagnostics;
+        const signatureCounter = diagnostics.reduce((acc, item) => {
+            acc[item.signature] = (acc[item.signature] ?? 0) + 1;
+            return acc;
+        }, {});
+        const precisionValues = diagnostics
+            .map((item) => item.problem_unit_precision)
+            .filter((item) => typeof item === "number");
+        const recallValues = diagnostics
+            .map((item) => item.problem_unit_recall_proxy)
+            .filter((item) => typeof item === "number");
+        const collapseValues = diagnostics
+            .map((item) => item.duplicate_collapse_rate)
+            .filter((item) => typeof item === "number");
+        const mechanismValues = diagnostics.map((item) => item.mechanism_coherence_score);
+        const clarityValues = diagnostics.map((item) => item.problem_clarity_score);
+        const firstApplicable = diagnostics.filter((item) => item.problem_first_answer_applied !== null);
+        const firstApplied = firstApplicable.filter((item) => item.problem_first_answer_applied === true).length;
+        const leakageCases = diagnostics.filter((item) => item.entity_leakage).length;
+        const followupCases = diagnostics.filter((item) => item.suite_case.question_type === "followup" || item.turn_count > 1);
+        const candidateCases = diagnostics.filter((item) => item.signals.candidate_evidence_total > 0);
+        const expectedProblemCases = diagnostics.filter((item) => item.expected_problem_first);
+        const average = (values) => {
+            if (values.length === 0)
+                return null;
+            return round2(values.reduce((acc, item) => acc + item, 0) / values.length);
+        };
+        const raw = {
+            problem_unit_precision: average(precisionValues),
+            problem_unit_recall_proxy: average(recallValues),
+            duplicate_collapse_rate: average(collapseValues),
+            mechanism_coherence_score: average(mechanismValues),
+            problem_clarity_score: average(clarityValues),
+            problem_first_answer_rate: firstApplicable.length > 0 ? round2(firstApplied / firstApplicable.length) : null,
+            entity_leakage_rate: diagnostics.length > 0 ? round2(leakageCases / diagnostics.length) : null
+        };
+        const rubric_bands = {
+            problem_unit_precision: rubricBandForMetricStage2("problem_unit_precision", raw.problem_unit_precision),
+            problem_unit_recall_proxy: rubricBandForMetricStage2("problem_unit_recall_proxy", raw.problem_unit_recall_proxy),
+            duplicate_collapse_rate: rubricBandForMetricStage2("duplicate_collapse_rate", raw.duplicate_collapse_rate),
+            mechanism_coherence_score: rubricBandForMetricStage2("mechanism_coherence_score", raw.mechanism_coherence_score),
+            problem_clarity_score: rubricBandForMetricStage2("problem_clarity_score", raw.problem_clarity_score),
+            problem_first_answer_rate: rubricBandForMetricStage2("problem_first_answer_rate", raw.problem_first_answer_rate),
+            entity_leakage_rate: rubricBandForMetricStage2("entity_leakage_rate", raw.entity_leakage_rate)
+        };
+        return {
+            raw,
+            rubric_bands,
+            denominators: {
+                cases_total: diagnostics.length,
+                expected_problem_cases_total: expectedProblemCases.length,
+                followup_cases_total: followupCases.length,
+                candidate_cases_total: candidateCases.length,
+                precision_cases_total: precisionValues.length,
+                recall_cases_total: recallValues.length,
+                duplicate_collapse_cases_total: collapseValues.length,
+                problem_first_applicable_cases_total: firstApplicable.length
+            },
+            signature_counts: signatureCounter
+        };
+    }
     buildAssistantComparisonReport(input) {
         const baselinePath = resolveReadablePath(input.baselineReportFile);
         const baselineReport = JSON.parse(fs_1.default.readFileSync(baselinePath, "utf-8"));
@@ -959,6 +1340,122 @@ class EvalService {
             }
         };
     }
+    buildAssistantStage2ComparisonReport(input) {
+        const baselinePath = resolveReadablePath(input.baselineReportFile);
+        const baselineReport = JSON.parse(fs_1.default.readFileSync(baselinePath, "utf-8"));
+        const currentReport = input.currentReport;
+        const metricKeys = [
+            "problem_unit_precision",
+            "problem_unit_recall_proxy",
+            "duplicate_collapse_rate",
+            "mechanism_coherence_score",
+            "problem_clarity_score",
+            "problem_first_answer_rate",
+            "entity_leakage_rate"
+        ];
+        const lowerIsBetter = new Set(["entity_leakage_rate"]);
+        const baselineRaw = (baselineReport.metrics ?? {}).raw ?? {};
+        const currentRaw = (currentReport.metrics ?? {}).raw ?? {};
+        const deltas = {};
+        for (const metric of metricKeys) {
+            const baseline = typeof baselineRaw[metric] === "number" ? Number(baselineRaw[metric]) : null;
+            const current = typeof currentRaw[metric] === "number" ? Number(currentRaw[metric]) : null;
+            const delta = baseline !== null && current !== null ? round2(current - baseline) : null;
+            let trend = "n/a";
+            if (baseline !== null && current !== null) {
+                const improved = lowerIsBetter.has(metric) ? current < baseline - 0.01 : current > baseline + 0.01;
+                const weakened = lowerIsBetter.has(metric) ? current > baseline + 0.01 : current < baseline - 0.01;
+                trend = improved ? "improved" : weakened ? "weakened" : "unchanged";
+            }
+            deltas[metric] = { baseline, current, delta, trend };
+        }
+        const baselineResults = Array.isArray(baselineReport.results) ? baselineReport.results : [];
+        const currentResults = Array.isArray(currentReport.results) ? currentReport.results : [];
+        const baselineByCase = new Map();
+        for (const row of baselineResults) {
+            baselineByCase.set(String(row.case_id ?? ""), row);
+        }
+        const improvedNotes = [];
+        const unchangedNotes = [];
+        const weakenedNotes = [];
+        const toComposite = (row) => {
+            if (!row || typeof row !== "object")
+                return null;
+            const metricSubscores = row.metric_subscores;
+            if (!metricSubscores)
+                return null;
+            const clarity = typeof metricSubscores.problem_clarity_score === "number" ? Number(metricSubscores.problem_clarity_score) : null;
+            const mechanism = typeof metricSubscores.mechanism_coherence_score === "number" ? Number(metricSubscores.mechanism_coherence_score) : null;
+            const firstRate = typeof metricSubscores.problem_first_answer_rate === "number" ? Number(metricSubscores.problem_first_answer_rate) : null;
+            const leakageRate = typeof metricSubscores.entity_leakage_rate === "number" ? Number(metricSubscores.entity_leakage_rate) : null;
+            if (clarity === null || mechanism === null || firstRate === null || leakageRate === null) {
+                return null;
+            }
+            return round2((clarity + mechanism + firstRate * 5 + (1 - leakageRate) * 5) / 4);
+        };
+        for (const row of currentResults) {
+            const caseId = String(row.case_id ?? "");
+            const currentComposite = toComposite(row);
+            const baselineComposite = toComposite(baselineByCase.get(caseId));
+            if (currentComposite === null || baselineComposite === null) {
+                continue;
+            }
+            const delta = round2(currentComposite - baselineComposite);
+            const note = `${caseId}: composite ${baselineComposite} -> ${currentComposite} (delta ${delta})`;
+            if (delta > 0.25) {
+                improvedNotes.push(note);
+            }
+            else if (delta < -0.25) {
+                weakenedNotes.push(note);
+            }
+            else {
+                unchangedNotes.push(note);
+            }
+        }
+        const comparisonId = `assistant-stage2-compare-${(0, nanoid_1.nanoid)(8)}`;
+        const comparisonReport = {
+            schema_version: ASSISTANT_STAGE2_COMPARISON_SCHEMA_VERSION,
+            comparison_id: comparisonId,
+            run_timestamp: new Date().toISOString(),
+            baseline_run_id: baselineReport.run_id ?? null,
+            current_run_id: currentReport.run_id ?? null,
+            eval_target: "assistant_stage2",
+            suite_id: currentReport.suite_id ?? baselineReport.suite_id ?? null,
+            suite_version: currentReport.suite_version ?? baselineReport.suite_version ?? null,
+            baseline_report_file: baselinePath,
+            current_report_file: currentReport.artifacts && typeof currentReport.artifacts === "object"
+                ? currentReport.artifacts.run_report_json_path ?? null
+                : null,
+            metric_deltas: deltas,
+            scenario_notes_summary: {
+                improved: improvedNotes.length,
+                unchanged: unchangedNotes.length,
+                weakened: weakenedNotes.length
+            },
+            scenario_notes: {
+                improved: improvedNotes,
+                unchanged: unchangedNotes,
+                weakened: weakenedNotes
+            },
+            known_limitations: currentReport.known_limitations ?? [
+                "Stage 2 comparison remains run-to-run and depends on stable feature profile.",
+                "Metrics are Stage 2 Wave 5 heuristics, not final product scorecards."
+            ],
+            report_title: "Assistant Stage 2 Baseline vs Current"
+        };
+        (0, files_1.ensureDir)(config_1.REPORTS_DIR);
+        const jsonPath = path_1.default.resolve(config_1.REPORTS_DIR, `${comparisonId}.json`);
+        const mdPath = path_1.default.resolve(config_1.REPORTS_DIR, `${comparisonId}.md`);
+        (0, files_1.writeJsonFile)(jsonPath, comparisonReport);
+        fs_1.default.writeFileSync(mdPath, buildAssistantStage2ComparisonMarkdownReport(comparisonReport), "utf-8");
+        return {
+            ...comparisonReport,
+            artifacts: {
+                comparison_report_json_path: jsonPath,
+                comparison_report_md_path: mdPath
+            }
+        };
+    }
     async runAssistantStage1(payload) {
         if (!config_1.FEATURE_ASSISTANT_ACCOUNTANT_EVAL_V1) {
             throw new http_1.ApiError("ASSISTANT_STAGE1_EVAL_DISABLED", "Assistant Stage 1 eval target is disabled by FEATURE_ASSISTANT_ACCOUNTANT_EVAL_V1.", 409);
@@ -1290,6 +1787,278 @@ class EvalService {
         }
         return report;
     }
+    async runAssistantStage2(payload) {
+        if (!config_1.FEATURE_ASSISTANT_STAGE2_EVAL_V1) {
+            throw new http_1.ApiError("ASSISTANT_STAGE2_EVAL_DISABLED", "Assistant Stage 2 eval target is disabled by FEATURE_ASSISTANT_STAGE2_EVAL_V1.", 409);
+        }
+        const suite = parseAssistantStage2SuiteFile(payload.caseSetFile);
+        const suiteCases = suite.cases.filter((item) => !payload.caseIds || payload.caseIds.includes(item.case_id));
+        const runId = `assistant-stage2-${(0, nanoid_1.nanoid)(10)}`;
+        const assistantService = new assistantService_1.AssistantService(this.normalizerService, new assistantSessionStore_1.AssistantSessionStore());
+        const diagnostics = [];
+        let requestsTotal = 0;
+        for (const suiteCase of suiteCases) {
+            const sessionId = `${runId}-${suiteCase.case_id}`;
+            const turnResponses = [];
+            const notes = [];
+            const limitations = [];
+            const expectedProblemUnitTypes = this.getExpectedProblemUnitTypes(suiteCase);
+            const expectedProblemFirst = suiteCase.expected_hints?.expected_problem_first ?? (suiteCase.broadness_level !== "low" || suiteCase.question_type !== "direct");
+            try {
+                for (const turn of suiteCase.turns) {
+                    const response = await assistantService.handleMessage({
+                        session_id: sessionId,
+                        user_message: turn.user_message,
+                        message: turn.user_message,
+                        mode: "assistant",
+                        apiKey: payload.normalizeConfig.apiKey,
+                        model: payload.normalizeConfig.model,
+                        baseUrl: payload.normalizeConfig.baseUrl,
+                        temperature: payload.normalizeConfig.temperature,
+                        maxOutputTokens: payload.normalizeConfig.maxOutputTokens,
+                        promptVersion: payload.normalizeConfig.promptVersion,
+                        systemPrompt: payload.normalizeConfig.systemPrompt,
+                        developerPrompt: payload.normalizeConfig.developerPrompt,
+                        domainPrompt: payload.normalizeConfig.domainPrompt,
+                        fewShotExamples: payload.normalizeConfig.fewShotExamples,
+                        useMock: payload.useMock
+                    });
+                    turnResponses.push(response);
+                    requestsTotal += 1;
+                }
+            }
+            catch (error) {
+                const errorMessage = error instanceof Error ? error.message : String(error);
+                diagnostics.push({
+                    suite_case: suiteCase,
+                    session_id: sessionId,
+                    trace_id: null,
+                    final_reply_type: "backend_error",
+                    turn_count: turnResponses.length,
+                    signature: `backend_error|${suiteCase.scenario_tag}`,
+                    expected_problem_unit_types: expectedProblemUnitTypes,
+                    expected_problem_first: expectedProblemFirst,
+                    problem_unit_precision: 0,
+                    problem_unit_recall_proxy: expectedProblemUnitTypes.length > 0 ? 0 : null,
+                    duplicate_collapse_rate: null,
+                    mechanism_coherence_score: 0,
+                    problem_clarity_score: 0,
+                    problem_first_answer_applied: expectedProblemFirst ? false : null,
+                    entity_leakage: false,
+                    signals: {
+                        broad_query_detected: suiteCase.broadness_level !== "low",
+                        broad_result_flag: false,
+                        narrowing_strength: null,
+                        minimum_evidence_failed: true,
+                        degraded_to: "clarification",
+                        evidence_confidence: "low",
+                        limitation_reason_codes: [],
+                        mechanism_status: null,
+                        source_refs: [],
+                        routes: [],
+                        followup_state_applied: false,
+                        uncertainty_limitations_count: 0,
+                        candidate_evidence_total: 0,
+                        problem_units_total: 0,
+                        problem_unit_types: [],
+                        problem_mechanism_summaries: [],
+                        duplicate_collapses_total: 0,
+                        problem_centric_answer_applied: false,
+                        problem_units_used_count: 0,
+                        problem_answer_mode: null,
+                        problem_unit_ids_used: [],
+                        entity_leakage_detected: false
+                    },
+                    limitations: [errorMessage],
+                    notes: [`Case execution failed: ${errorMessage}`]
+                });
+                continue;
+            }
+            const finalResponse = turnResponses[turnResponses.length - 1];
+            const signals = this.collectAssistantStage2Signals(finalResponse, turnResponses);
+            const problemUnitPrecision = this.computeProblemUnitPrecision(expectedProblemUnitTypes, signals.problem_unit_types);
+            const problemUnitRecallProxy = this.computeProblemUnitRecallProxy(expectedProblemUnitTypes, signals.problem_unit_types);
+            const duplicateCollapseRate = this.computeDuplicateCollapseRate(signals.candidate_evidence_total, signals.duplicate_collapses_total);
+            const mechanismCoherenceScore = this.computeMechanismCoherenceScore(finalResponse, signals);
+            const problemClarityScore = this.computeProblemClarityScore(finalResponse, signals);
+            const problemFirstAnswerApplied = expectedProblemFirst ? signals.problem_centric_answer_applied && signals.problem_units_used_count > 0 : null;
+            if (signals.problem_units_total === 0 && expectedProblemUnitTypes.length > 0) {
+                limitations.push("missing_problem_units");
+            }
+            if (signals.problem_centric_answer_applied && signals.problem_units_used_count <= 0) {
+                limitations.push("problem_mode_without_units");
+            }
+            limitations.push(...signals.limitation_reason_codes.map((item) => `limitation_reason:${item}`));
+            if (signals.entity_leakage_detected) {
+                limitations.push("entity_leakage_detected");
+            }
+            if (problemFirstAnswerApplied === false)
+                notes.push("problem_first_not_applied");
+            if (signals.problem_units_total === 0)
+                notes.push("problem_units_missing");
+            if (signals.problem_unit_types.length > 0)
+                notes.push(`problem_types:${signals.problem_unit_types.join(",")}`);
+            if (signals.entity_leakage_detected)
+                notes.push("entity_leakage");
+            if (signals.degraded_to === "clarification")
+                notes.push("clarification_degraded");
+            diagnostics.push({
+                suite_case: suiteCase,
+                session_id: sessionId,
+                trace_id: finalResponse.debug?.trace_id ?? null,
+                final_reply_type: finalResponse.reply_type,
+                turn_count: suiteCase.turns.length,
+                signature: [
+                    finalResponse.reply_type,
+                    signals.problem_answer_mode ?? "unknown",
+                    signals.problem_unit_types.sort().join(","),
+                    signals.degraded_to ?? "none"
+                ].join("|"),
+                expected_problem_unit_types: expectedProblemUnitTypes,
+                expected_problem_first: expectedProblemFirst,
+                problem_unit_precision: problemUnitPrecision,
+                problem_unit_recall_proxy: problemUnitRecallProxy,
+                duplicate_collapse_rate: duplicateCollapseRate,
+                mechanism_coherence_score: mechanismCoherenceScore,
+                problem_clarity_score: problemClarityScore,
+                problem_first_answer_applied: problemFirstAnswerApplied,
+                entity_leakage: signals.entity_leakage_detected,
+                signals,
+                limitations: Array.from(new Set(limitations)),
+                notes
+            });
+        }
+        const metrics = this.computeAssistantStage2Metrics({ diagnostics });
+        const caseRecords = diagnostics.map((item) => {
+            const caseMetricVector = {
+                problem_unit_precision: item.problem_unit_precision,
+                problem_unit_recall_proxy: item.problem_unit_recall_proxy,
+                duplicate_collapse_rate: item.duplicate_collapse_rate,
+                mechanism_coherence_score: round2(item.mechanism_coherence_score),
+                problem_clarity_score: round2(item.problem_clarity_score),
+                problem_first_answer_rate: item.problem_first_answer_applied === null ? null : item.problem_first_answer_applied ? 1 : 0,
+                entity_leakage_rate: item.entity_leakage ? 1 : 0
+            };
+            return {
+                schema_version: stage2EvalContracts_1.ASSISTANT_STAGE2_EVAL_RECORD_SCHEMA_VERSION,
+                created_at: new Date().toISOString(),
+                case_id: item.suite_case.case_id,
+                scenario_tag: item.suite_case.scenario_tag,
+                session_id: item.session_id,
+                trace_id: item.trace_id,
+                question_type: item.suite_case.question_type,
+                broadness_level: item.suite_case.broadness_level,
+                expected_problem_unit_types: item.expected_problem_unit_types,
+                expected_problem_first: item.expected_problem_first,
+                problem_units_detected: item.signals.problem_units_total,
+                candidate_evidence_detected: item.signals.candidate_evidence_total,
+                duplicate_collapses_detected: item.signals.duplicate_collapses_total,
+                metric_subscores: caseMetricVector,
+                raw_signals: {
+                    final_reply_type: item.final_reply_type,
+                    turn_count: item.turn_count,
+                    broad_query_detected: item.signals.broad_query_detected,
+                    broad_result_flag: item.signals.broad_result_flag,
+                    narrowing_strength: item.signals.narrowing_strength,
+                    minimum_evidence_failed: item.signals.minimum_evidence_failed,
+                    degraded_to: item.signals.degraded_to,
+                    evidence_confidence: item.signals.evidence_confidence,
+                    limitation_reason_codes: item.signals.limitation_reason_codes,
+                    mechanism_status: item.signals.mechanism_status,
+                    source_refs: item.signals.source_refs,
+                    routes: item.signals.routes,
+                    followup_state_applied: item.signals.followup_state_applied,
+                    problem_units_total: item.signals.problem_units_total,
+                    candidate_evidence_total: item.signals.candidate_evidence_total,
+                    problem_unit_types: item.signals.problem_unit_types,
+                    duplicate_collapses_total: item.signals.duplicate_collapses_total,
+                    problem_centric_answer_applied: item.signals.problem_centric_answer_applied,
+                    problem_units_used_count: item.signals.problem_units_used_count,
+                    problem_answer_mode: item.signals.problem_answer_mode,
+                    problem_unit_ids_used: item.signals.problem_unit_ids_used,
+                    entity_leakage_detected: item.signals.entity_leakage_detected
+                },
+                limitations: item.limitations,
+                notes: item.notes
+            };
+        });
+        const strongestSignals = Object.entries(metrics.rubric_bands)
+            .filter(([, band]) => band?.score === 5)
+            .map(([name]) => name);
+        const weakestSignals = Object.entries(metrics.rubric_bands)
+            .filter(([, band]) => band?.score === 0)
+            .map(([name]) => name);
+        const runTimestamp = new Date().toISOString();
+        const report = {
+            schema_version: ASSISTANT_STAGE2_RUN_SCHEMA_VERSION,
+            run_id: runId,
+            run_timestamp: runTimestamp,
+            eval_target: "assistant_stage2",
+            mode: payload.mode,
+            use_mock: Boolean(payload.useMock),
+            prompt_version: payload.normalizeConfig.promptVersion ?? null,
+            suite_id: suite.suite_id,
+            suite_version: suite.suite_version,
+            suite_schema_version: suite.schema_version ?? null,
+            scenario_count: suite.scenario_count,
+            case_ids: suiteCases.map((item) => item.case_id),
+            cases_total: caseRecords.length,
+            feature_profile_snapshot: buildFeatureProfileSnapshot(),
+            code_version: buildCodeVersionMarker(),
+            metrics: {
+                raw: metrics.raw,
+                denominators: metrics.denominators
+            },
+            rubric_bands: metrics.rubric_bands,
+            subsets: {
+                expected_problem_cases_total: metrics.denominators.expected_problem_cases_total,
+                followup_cases_total: metrics.denominators.followup_cases_total,
+                candidate_cases_total: metrics.denominators.candidate_cases_total
+            },
+            budget: {
+                requests_total: requestsTotal
+            },
+            results: caseRecords,
+            scenario_summary: {
+                improved_or_strong: caseRecords.filter((item) => {
+                    const clarity = Number(item.metric_subscores.problem_clarity_score ?? 0);
+                    const mechanism = Number(item.metric_subscores.mechanism_coherence_score ?? 0);
+                    return clarity >= 4 && mechanism >= 3;
+                }).length,
+                unchanged_or_mixed: caseRecords.filter((item) => {
+                    const clarity = Number(item.metric_subscores.problem_clarity_score ?? 0);
+                    return clarity >= 2.5 && clarity < 4;
+                }).length,
+                weak_or_regressed: caseRecords.filter((item) => Number(item.metric_subscores.problem_clarity_score ?? 0) < 2.5).length
+            },
+            improvement_hints: {
+                strongest_signals: strongestSignals.length > 0 ? strongestSignals.join(", ") : "none",
+                weakest_signals: weakestSignals.length > 0 ? weakestSignals.join(", ") : "none"
+            },
+            known_limitations: [
+                "Stage 2 eval remains heuristic and scoped to problem-unit baseline (no graph/lifecycle/investigation runtime scoring).",
+                "problem_unit_recall_proxy uses suite expected types as lightweight proxy, not full ground-truth labeling.",
+                "Comparison quality depends on stable feature profile and reproducible mock/runtime setup."
+            ],
+            report_title: "Assistant Stage 2 Eval Run"
+        };
+        (0, files_1.ensureDir)(config_1.REPORTS_DIR);
+        const runJsonPath = path_1.default.resolve(config_1.REPORTS_DIR, `${runId}.json`);
+        const runMdPath = path_1.default.resolve(config_1.REPORTS_DIR, `${runId}.md`);
+        (0, files_1.writeJsonFile)(runJsonPath, report);
+        fs_1.default.writeFileSync(runMdPath, buildAssistantStage2EvalMarkdownReport(report), "utf-8");
+        report.artifacts = {
+            run_report_json_path: runJsonPath,
+            run_report_md_path: runMdPath
+        };
+        if (payload.compareWithReportFile) {
+            report.comparison = this.buildAssistantStage2ComparisonReport({
+                currentReport: report,
+                baselineReportFile: payload.compareWithReportFile
+            });
+        }
+        return report;
+    }
     async run(payload) {
         const mode = payload.mode ?? "standard";
         const evalTarget = payload.evalTarget ?? "normalizer";
@@ -1303,6 +2072,16 @@ class EvalService {
                 compareWithReportFile: payload.compareWithReportFile
             });
         }
+        if (evalTarget === "assistant_stage2") {
+            return this.runAssistantStage2({
+                normalizeConfig: payload.normalizeConfig,
+                caseIds: payload.caseIds,
+                useMock: payload.useMock,
+                mode,
+                caseSetFile: payload.caseSetFile,
+                compareWithReportFile: payload.compareWithReportFile
+            });
+        }
         const promptVersion = String(payload.normalizeConfig.promptVersion ?? "").toLowerCase();
         const schemaVersion = String(payload.normalizeConfig.schemaVersion ?? "").toLowerCase();
         const isV2 = promptVersion.startsWith("normalizer_v2") || schemaVersion === "v2" || schemaVersion === "v2_0_1" || schemaVersion === "v2_0_2";

llm_normalizer/backend/dist/services/investigationState.js

@@ -4,6 +4,7 @@ exports.cloneInvestigationState = cloneInvestigationState;
 exports.createEmptyInvestigationState = createEmptyInvestigationState;
 exports.updateInvestigationState = updateInvestigationState;
 const stage1Contracts_1 = require("../types/stage1Contracts");
+const stage2ProblemUnits_1 = require("../types/stage2ProblemUnits");
 function uniqueStrings(values) {
     return Array.from(new Set(values.map((item) => item.trim()).filter(Boolean)));
 }
@@ -74,10 +75,101 @@ function collectOpenUncertainties(coverageReport, retrievalResults) {
     const limitationNotes = retrievalResults.flatMap((result) => result.limitations).slice(0, 6);
     return capStrings([...requirementNotes, ...limitationNotes], stage1Contracts_1.INVESTIGATION_MAX_UNCERTAINTIES);
 }
+function normalizeEntityBacklinks(values) {
+    const result = [];
+    const seen = new Set();
+    for (const item of values) {
+        const entity = String(item.entity ?? "").trim();
+        const id = String(item.id ?? "").trim();
+        if (!entity || !id) {
+            continue;
+        }
+        const key = `${entity}::${id}`;
+        if (seen.has(key)) {
+            continue;
+        }
+        seen.add(key);
+        result.push({
+            entity,
+            id
+        });
+    }
+    return result;
+}
+function collectProblemUnits(retrievalResults) {
+    return retrievalResults.flatMap((result) => result.problem_units ?? []);
+}
+function capProblemUnitState(state) {
+    return {
+        active_problem_units: capStrings(state.active_problem_units, stage2ProblemUnits_1.INVESTIGATION_MAX_ACTIVE_PROBLEM_UNITS),
+        resolved_problem_units: capStrings(state.resolved_problem_units, stage2ProblemUnits_1.INVESTIGATION_MAX_RESOLVED_PROBLEM_UNITS),
+        problem_unit_backlinks: state.problem_unit_backlinks
+            .map((item) => ({
+            problem_unit_id: String(item.problem_unit_id ?? "").trim(),
+            entity_backlinks: normalizeEntityBacklinks(item.entity_backlinks ?? [])
+        }))
+            .filter((item) => Boolean(item.problem_unit_id) && item.entity_backlinks.length > 0)
+            .slice(0, stage2ProblemUnits_1.INVESTIGATION_MAX_PROBLEM_UNIT_BACKLINKS),
+        focus_problem_types: capStrings(state.focus_problem_types.map((item) => String(item)), stage2ProblemUnits_1.INVESTIGATION_MAX_FOCUS_PROBLEM_TYPES)
+    };
+}
+function updateProblemUnitState(previous, retrievalResults) {
+    const previousState = previous.problem_unit_state;
+    const currentProblemUnits = collectProblemUnits(retrievalResults);
+    const currentIds = capStrings(currentProblemUnits.map((item) => String(item.problem_unit_id ?? "")), stage2ProblemUnits_1.INVESTIGATION_MAX_ACTIVE_PROBLEM_UNITS);
+    const currentTypes = capStrings(currentProblemUnits.map((item) => String(item.problem_unit_type ?? "")), stage2ProblemUnits_1.INVESTIGATION_MAX_FOCUS_PROBLEM_TYPES);
+    const currentBacklinksRaw = currentProblemUnits
+        .filter((item) => currentIds.includes(item.problem_unit_id))
+        .map((item) => ({
+        problem_unit_id: item.problem_unit_id,
+        entity_backlinks: normalizeEntityBacklinks(item.entity_backlinks ?? [])
+    }))
+        .filter((item) => item.entity_backlinks.length > 0);
+    const currentBacklinksById = new Map(currentBacklinksRaw.map((item) => [item.problem_unit_id, item.entity_backlinks]));
+    const previousBacklinksById = new Map((previousState?.problem_unit_backlinks ?? []).map((item) => [item.problem_unit_id, item.entity_backlinks]));
+    const active_problem_units = currentIds.length > 0
+        ? currentIds
+        : capStrings(previousState?.active_problem_units ?? [], stage2ProblemUnits_1.INVESTIGATION_MAX_ACTIVE_PROBLEM_UNITS);
+    const resolved_problem_units = currentIds.length > 0
+        ? capStrings([
+            ...(previousState?.active_problem_units ?? []).filter((item) => !currentIds.includes(item)),
+            ...(previousState?.resolved_problem_units ?? [])
+        ], stage2ProblemUnits_1.INVESTIGATION_MAX_RESOLVED_PROBLEM_UNITS)
+        : capStrings(previousState?.resolved_problem_units ?? [], stage2ProblemUnits_1.INVESTIGATION_MAX_RESOLVED_PROBLEM_UNITS);
+    const problem_unit_backlinks = active_problem_units
+        .map((problemUnitId) => {
+        const entity_backlinks = normalizeEntityBacklinks(currentBacklinksById.get(problemUnitId) ?? previousBacklinksById.get(problemUnitId) ?? []);
+        if (entity_backlinks.length === 0) {
+            return null;
+        }
+        return {
+            problem_unit_id: problemUnitId,
+            entity_backlinks
+        };
+    })
+        .filter((item) => item !== null)
+        .slice(0, stage2ProblemUnits_1.INVESTIGATION_MAX_PROBLEM_UNIT_BACKLINKS);
+    const focus_problem_types = currentTypes.length > 0
+        ? currentTypes
+        : capStrings((previousState?.focus_problem_types ?? []).map((item) => String(item)), stage2ProblemUnits_1.INVESTIGATION_MAX_FOCUS_PROBLEM_TYPES);
+    const nextState = capProblemUnitState({
+        active_problem_units,
+        resolved_problem_units,
+        problem_unit_backlinks,
+        focus_problem_types
+    });
+    if (nextState.active_problem_units.length === 0 &&
+        nextState.resolved_problem_units.length === 0 &&
+        nextState.problem_unit_backlinks.length === 0 &&
+        nextState.focus_problem_types.length === 0) {
+        return undefined;
+    }
+    return nextState;
+}
 function cloneInvestigationState(state) {
     if (!state)
         return null;
-    return {
+    const cloned = {
         ...state,
         focus: {
             ...state.focus,
@@ -92,6 +184,18 @@ function cloneInvestigationState(state) {
             }
             : null
     };
+    if (state.problem_unit_state) {
+        cloned.problem_unit_state = capProblemUnitState({
+            active_problem_units: [...state.problem_unit_state.active_problem_units],
+            resolved_problem_units: [...state.problem_unit_state.resolved_problem_units],
+            problem_unit_backlinks: state.problem_unit_state.problem_unit_backlinks.map((item) => ({
+                problem_unit_id: item.problem_unit_id,
+                entity_backlinks: [...item.entity_backlinks]
+            })),
+            focus_problem_types: [...state.problem_unit_state.focus_problem_types]
+        });
+    }
+    return cloned;
 }
 function createEmptyInvestigationState(sessionId, timestamp = new Date().toISOString()) {
     return {
@@ -120,6 +224,7 @@ function updateInvestigationState(input) {
     const focusFromMessage = capStrings(detectAccounts(input.userMessage), stage1Contracts_1.INVESTIGATION_MAX_PRIMARY_ACCOUNTS);
     const requirementIds = capStrings(input.requirements.map((item) => item.requirement_id), stage1Contracts_1.INVESTIGATION_MAX_REQUIREMENT_LINKS);
     const mainRequirement = input.requirements[0]?.requirement_text ?? input.userMessage;
+    const problemUnitState = updateProblemUnitState(previous, input.retrievalResults);
     return {
         schema_version: stage1Contracts_1.INVESTIGATION_STATE_SCHEMA_VERSION,
         session_id: previous.session_id,
@@ -142,6 +247,11 @@ function updateInvestigationState(input) {
             last_user_message: input.userMessage.slice(0, 240),
             referenced_requirement_ids: requirementIds
         },
-        query_mode_hint: deriveQueryModeHint(input.routeSummary)
+        query_mode_hint: deriveQueryModeHint(input.routeSummary),
+        ...(problemUnitState
+            ? {
+                problem_unit_state: problemUnitState
+            }
+            : {})
     };
 }

llm_normalizer/backend/dist/services/problemUnitAssembler.js

@@ -0,0 +1,467 @@
+"use strict";
+Object.defineProperty(exports, "__esModule", { value: true });
+exports.buildCandidateEvidence = buildCandidateEvidence;
+exports.clusterCandidateEvidence = clusterCandidateEvidence;
+exports.detectProblemUnitType = detectProblemUnitType;
+exports.scoreProblemSeverity = scoreProblemSeverity;
+exports.buildProblemUnit = buildProblemUnit;
+exports.collapseDuplicates = collapseDuplicates;
+exports.assembleProblemUnits = assembleProblemUnits;
+const stage2ProblemUnits_1 = require("../types/stage2ProblemUnits");
+function toObject(value) {
+    if (!value || typeof value !== "object" || Array.isArray(value)) {
+        return null;
+    }
+    return value;
+}
+function uniqueStrings(values) {
+    return Array.from(new Set(values.map((item) => item.trim()).filter(Boolean)));
+}
+function clampUnitScore(value) {
+    if (!Number.isFinite(value)) {
+        return 0;
+    }
+    if (value <= 0)
+        return 0;
+    if (value >= 1)
+        return 1;
+    return Number(value.toFixed(2));
+}
+function gradeForScore(score) {
+    if (score >= 0.7)
+        return "high";
+    if (score >= 0.4)
+        return "medium";
+    return "low";
+}
+function confidenceGradeForScore(score) {
+    if (score >= 0.75)
+        return "high";
+    if (score >= 0.45)
+        return "medium";
+    return "low";
+}
+function confidenceToScore(value) {
+    if (value === "high")
+        return 1;
+    if (value === "medium")
+        return 0.6;
+    return 0.3;
+}
+function valueFromPayload(item, key) {
+    const value = item.payload[key];
+    if (typeof value !== "string") {
+        return null;
+    }
+    const trimmed = value.trim();
+    return trimmed.length > 0 ? trimmed : null;
+}
+function stringArrayFromUnknown(value) {
+    if (!Array.isArray(value)) {
+        return [];
+    }
+    return uniqueStrings(value.map((entry) => String(entry)));
+}
+function stringArrayFromPayload(item, key) {
+    return stringArrayFromUnknown(item.payload[key]);
+}
+function extractSemanticProfile(summary) {
+    const semanticProfile = toObject(summary.semantic_profile);
+    return {
+        relation_patterns: stringArrayFromUnknown(semanticProfile?.relation_patterns),
+        anomaly_patterns: stringArrayFromUnknown(semanticProfile?.anomaly_patterns)
+    };
+}
+function resolveEntityOverlay(item, rawEntities) {
+    const sourceId = String(item.pointer.source.id ?? "").toLowerCase();
+    if (!sourceId) {
+        return null;
+    }
+    for (const entity of rawEntities) {
+        const candidates = [
+            String(entity.source_id ?? ""),
+            String(entity.entity_id ?? ""),
+            String(entity.id ?? "")
+        ]
+            .map((entry) => entry.toLowerCase())
+            .filter(Boolean);
+        if (candidates.includes(sourceId)) {
+            return entity;
+        }
+    }
+    return null;
+}
+function detectRelationPatternHits(item, overlay, context) {
+    const hits = [];
+    const failedEdge = valueFromPayload(item, "failed_expected_edge");
+    if (failedEdge) {
+        hits.push(`failed_edge:${failedEdge}`);
+    }
+    const expectedStep = valueFromPayload(item, "expected_next_step");
+    if (expectedStep) {
+        hits.push(`expected_step:${expectedStep}`);
+    }
+    const relationPattern = valueFromPayload(item, "relation_pattern");
+    if (relationPattern) {
+        hits.push(relationPattern);
+    }
+    hits.push(...stringArrayFromPayload(item, "relation_patterns"));
+    hits.push(...stringArrayFromPayload(item, "relation_pattern_hits"));
+    if (overlay) {
+        hits.push(...stringArrayFromUnknown(overlay.relation_pattern_hits));
+        hits.push(...stringArrayFromUnknown(overlay.relation_types));
+    }
+    hits.push(...context.summary_relation_patterns);
+    if (context.route === "hybrid_store_plus_live" && hits.length === 0) {
+        hits.push("chain_scope");
+    }
+    return uniqueStrings(hits);
+}
+function detectAnomalyPatterns(item, overlay, context) {
+    const patterns = [];
+    patterns.push(...stringArrayFromPayload(item, "anomaly_patterns"));
+    patterns.push(...stringArrayFromPayload(item, "risk_factors"));
+    patterns.push(...stringArrayFromPayload(item, "lifecycle_gaps"));
+    patterns.push(...stringArrayFromPayload(item, "lifecycle_markers"));
+    const explicit = valueFromPayload(item, "anomaly_pattern");
+    if (explicit) {
+        patterns.push(explicit);
+    }
+    if (overlay) {
+        patterns.push(...stringArrayFromUnknown(overlay.risk_factors));
+        patterns.push(...stringArrayFromUnknown(overlay.lifecycle_gaps));
+        patterns.push(...stringArrayFromUnknown(overlay.anomaly_patterns));
+    }
+    patterns.push(...context.summary_anomaly_patterns);
+    patterns.push(...context.risk_factors);
+    if (item.evidence_kind === "anomaly_signal") {
+        patterns.push("anomaly_signal");
+    }
+    if (item.limitation?.reason_code === "missing_mechanism") {
+        patterns.push("missing_mechanism");
+    }
+    if (item.limitation?.reason_code === "insufficient_detail") {
+        patterns.push("insufficient_detail");
+    }
+    const defectClass = valueFromPayload(item, "business_defect_class");
+    if (defectClass) {
+        patterns.push(defectClass);
+    }
+    if (context.route === "store_feature_risk") {
+        patterns.push("risk_route");
+    }
+    return uniqueStrings(patterns);
+}
+function buildEntityBacklinks(item, overlay) {
+    const backlinks = [];
+    const sourceEntity = String(item.pointer.source.entity ?? "").trim();
+    const sourceId = String(item.pointer.source.id ?? "").trim();
+    if (sourceEntity && sourceId) {
+        backlinks.push({
+            entity: sourceEntity,
+            id: sourceId
+        });
+    }
+    const payloadEntity = valueFromPayload(item, "source_entity");
+    const payloadId = valueFromPayload(item, "source_id");
+    if (payloadEntity && payloadId) {
+        backlinks.push({
+            entity: payloadEntity,
+            id: payloadId
+        });
+    }
+    const overlayEntity = overlay ? String(overlay.source_entity ?? "").trim() : "";
+    const overlayId = overlay ? String(overlay.source_id ?? "").trim() : "";
+    if (overlayEntity && overlayId) {
+        backlinks.push({
+            entity: overlayEntity,
+            id: overlayId
+        });
+    }
+    return Array.from(new Map(backlinks.map((entry) => [`${entry.entity.toLowerCase()}|${entry.id.toLowerCase()}`, entry])).values());
+}
+function inferExpectedState(item) {
+    return valueFromPayload(item, "expected_state") ?? valueFromPayload(item, "expected_next_step") ?? undefined;
+}
+function inferActualState(item) {
+    return valueFromPayload(item, "actual_state") ?? valueFromPayload(item, "mechanism_of_failure") ?? undefined;
+}
+function buildCandidateEvidence(items, route, contextInput) {
+    const context = {
+        route,
+        result_type: contextInput?.result_type,
+        raw_entities: contextInput?.raw_entities ?? [],
+        summary_relation_patterns: contextInput?.summary_relation_patterns ?? [],
+        summary_anomaly_patterns: contextInput?.summary_anomaly_patterns ?? [],
+        risk_factors: contextInput?.risk_factors ?? []
+    };
+    return items.map((item, index) => {
+        const overlay = resolveEntityOverlay(item, context.raw_entities);
+        return {
+            schema_version: stage2ProblemUnits_1.CANDIDATE_EVIDENCE_SCHEMA_VERSION,
+            candidate_id: `cand-${item.evidence_id || `${route}-${index + 1}`}`,
+            route,
+            source_ref: item.source_ref,
+            expected_state: inferExpectedState(item),
+            actual_state: inferActualState(item),
+            relation_pattern_hits: detectRelationPatternHits(item, overlay, context),
+            anomaly_patterns: detectAnomalyPatterns(item, overlay, context),
+            entity_backlinks: buildEntityBacklinks(item, overlay),
+            confidence_hint: item.confidence
+        };
+    });
+}
+function clusterSignature(candidate) {
+    const relation = candidate.relation_pattern_hits[0] ?? "none";
+    const anomaly = candidate.anomaly_patterns[0] ?? "none";
+    return [candidate.route, candidate.source_ref.canonical_ref, relation, anomaly].join("|");
+}
+function clusterCandidateEvidence(candidates) {
+    const byCluster = new Map();
+    for (const candidate of candidates) {
+        const signature = clusterSignature(candidate);
+        const current = byCluster.get(signature) ?? [];
+        current.push(candidate);
+        byCluster.set(signature, current);
+    }
+    return Array.from(byCluster.entries()).map(([cluster_id, clusterCandidates]) => ({
+        cluster_id,
+        candidates: clusterCandidates
+    }));
+}
+function hasAny(value, pattern) {
+    return pattern.test(value);
+}
+function detectProblemUnitType(cluster) {
+    const relationText = cluster.candidates.flatMap((item) => item.relation_pattern_hits).join(" ").toLowerCase();
+    const anomalyText = cluster.candidates.flatMap((item) => item.anomaly_patterns).join(" ").toLowerCase();
+    const sourceText = cluster.candidates
+        .map((item) => `${item.source_ref.entity} ${item.source_ref.id}`)
+        .join(" ")
+        .toLowerCase();
+    const routeText = cluster.candidates.map((item) => item.route).join(" ").toLowerCase();
+    if (hasAny(`${anomalyText} ${sourceText}`, /cross[_\s-]?branch|vat|nds|tax|ндс/)) {
+        return "cross_branch_inconsistency_cluster";
+    }
+    if (hasAny(`${anomalyText} ${relationText}`, /period|close[_\s-]?risk|reporting|закрыт|period_close/)) {
+        return "period_risk_cluster";
+    }
+    if (hasAny(anomalyText, /settlement|tail|unresolved|хвост|незакрыт/)) {
+        return "unresolved_settlement_cluster";
+    }
+    if (hasAny(anomalyText, /lifecycle|deferred|broken_lifecycle|списани|амортиз|рбп/)) {
+        return "lifecycle_anomaly_node";
+    }
+    if (hasAny(relationText, /failed_edge|statement_to_document|payment_to_settlement|chain|broken|цепоч|разрыв/)
+        || (routeText.includes("hybrid_store_plus_live") && relationText.length > 0)) {
+        return "broken_chain_segment";
+    }
+    return "document_conflict";
+}
+function scoreProblemSeverity(cluster) {
+    const candidates = cluster.candidates;
+    const averageConfidence = candidates.length > 0
+        ? candidates.reduce((acc, item) => acc + confidenceToScore(item.confidence_hint), 0) / candidates.length
+        : 0;
+    const hasEdgeBreak = candidates.some((item) => item.relation_pattern_hits.some((pattern) => /failed_edge|chain|statement_to_document|payment_to_settlement/i.test(pattern)));
+    const hasAnomaly = candidates.some((item) => item.anomaly_patterns.length > 0);
+    const hasPeriodRisk = candidates.some((item) => item.anomaly_patterns.some((pattern) => /period|close|reporting|закрыт/i.test(pattern)));
+    const candidateBoost = Math.min(candidates.length, 5) * 0.08;
+    let severityScore = 0.25 + candidateBoost;
+    if (hasEdgeBreak)
+        severityScore += 0.2;
+    if (hasAnomaly)
+        severityScore += 0.15;
+    if (hasPeriodRisk)
+        severityScore += 0.1;
+    const normalizedSeverity = clampUnitScore(severityScore);
+    let confidenceScore = averageConfidence;
+    if (hasEdgeBreak)
+        confidenceScore += 0.05;
+    const normalizedConfidence = clampUnitScore(confidenceScore);
+    return {
+        severity: {
+            score: normalizedSeverity,
+            grade: gradeForScore(normalizedSeverity)
+        },
+        confidence: {
+            score: normalizedConfidence,
+            grade: confidenceGradeForScore(normalizedConfidence)
+        }
+    };
+}
+function unitTitle(type) {
+    if (type === "document_conflict")
+        return "Document conflict detected";
+    if (type === "broken_chain_segment")
+        return "Broken chain segment detected";
+    if (type === "lifecycle_anomaly_node")
+        return "Lifecycle anomaly node detected";
+    if (type === "unresolved_settlement_cluster")
+        return "Unresolved settlement cluster detected";
+    if (type === "period_risk_cluster")
+        return "Period risk cluster detected";
+    return "Cross-branch inconsistency cluster detected";
+}
+function mechanismSummary(cluster, type) {
+    const relationHints = uniqueStrings(cluster.candidates.flatMap((item) => item.relation_pattern_hits));
+    const anomalyHints = uniqueStrings(cluster.candidates.flatMap((item) => item.anomaly_patterns));
+    const primaryRelation = relationHints[0];
+    const primaryAnomaly = anomalyHints[0];
+    if (primaryRelation) {
+        return `Mechanism candidate: ${primaryRelation}.`;
+    }
+    if (primaryAnomaly) {
+        return `Mechanism inferred from anomaly pattern: ${primaryAnomaly}.`;
+    }
+    return `Mechanism is currently inferred at baseline level for ${type}.`;
+}
+function businessDefectClass(cluster, type) {
+    const patterns = uniqueStrings([
+        ...cluster.candidates.flatMap((item) => item.relation_pattern_hits),
+        ...cluster.candidates.flatMap((item) => item.anomaly_patterns)
+    ]);
+    return patterns[0] ?? type;
+}
+function collectAffectedByEntity(backlinks, pattern) {
+    return uniqueStrings(backlinks.filter((entry) => pattern.test(entry.entity)).map((entry) => `${entry.entity}:${entry.id}`));
+}
+function parseFailedExpectedEdge(cluster) {
+    const withEdge = cluster.candidates.flatMap((item) => item.relation_pattern_hits).find((pattern) => pattern.startsWith("failed_edge:"));
+    if (!withEdge) {
+        return undefined;
+    }
+    return withEdge.replace(/^failed_edge:/, "").trim() || undefined;
+}
+function mergeBacklinks(candidates) {
+    return Array.from(new Map(candidates
+        .flatMap((item) => item.entity_backlinks)
+        .map((entry) => [`${entry.entity.toLowerCase()}|${entry.id.toLowerCase()}`, entry])).values());
+}
+function buildProblemUnit(cluster, index) {
+    const type = detectProblemUnitType(cluster);
+    const scored = scoreProblemSeverity(cluster);
+    const backlinks = mergeBacklinks(cluster.candidates);
+    const expectedState = cluster.candidates.find((item) => typeof item.expected_state === "string")?.expected_state;
+    const actualState = cluster.candidates.find((item) => typeof item.actual_state === "string")?.actual_state;
+    const failedExpectedEdge = parseFailedExpectedEdge(cluster);
+    const periodSensitive = cluster.candidates.some((item) => item.anomaly_patterns.some((pattern) => /period|close|reporting|закрыт/i.test(pattern)));
+    const hasLowConfidence = cluster.candidates.some((item) => item.confidence_hint === "low");
+    return {
+        schema_version: stage2ProblemUnits_1.PROBLEM_UNIT_SCHEMA_VERSION,
+        problem_unit_id: `pu-${type}-${index + 1}`,
+        problem_unit_type: type,
+        title: unitTitle(type),
+        mechanism_summary: mechanismSummary(cluster, type),
+        business_defect_class: businessDefectClass(cluster, type),
+        severity: scored.severity,
+        confidence: scored.confidence,
+        affected_entities: uniqueStrings(backlinks.map((entry) => `${entry.entity}:${entry.id}`)),
+        affected_documents: collectAffectedByEntity(backlinks, /doc|document|invoice|плат|реал|поступ/i),
+        affected_postings: collectAffectedByEntity(backlinks, /posting|journal|провод/i),
+        affected_accounts: collectAffectedByEntity(backlinks, /account|счет|сч/i),
+        affected_counterparties: collectAffectedByEntity(backlinks, /counterparty|supplier|buyer|контраг|постав|покуп/i),
+        affected_contracts: collectAffectedByEntity(backlinks, /contract|договор/i),
+        ...(expectedState ? { expected_state: expectedState } : {}),
+        ...(actualState ? { actual_state: actualState } : {}),
+        ...(failedExpectedEdge ? { failed_expected_edge: failedExpectedEdge } : {}),
+        ...(periodSensitive
+            ? {
+                period_impact: {
+                    is_period_sensitive: true,
+                    impact_class: "close_risk"
+                }
+            }
+            : {}),
+        evidence_pack: uniqueStrings(cluster.candidates.map((item) => item.candidate_id)),
+        entity_backlinks: backlinks,
+        snapshot_limitations: uniqueStrings(hasLowConfidence ? ["low_confidence_candidates_present"] : [])
+    };
+}
+function collapseSignature(unit) {
+    const backlink = unit.entity_backlinks[0] ? `${unit.entity_backlinks[0].entity}|${unit.entity_backlinks[0].id}` : "none";
+    return [unit.problem_unit_type, unit.business_defect_class, unit.failed_expected_edge ?? "none", backlink].join("|");
+}
+function collapseDuplicates(units) {
+    const bySignature = new Map();
+    let duplicateCollapses = 0;
+    for (const unit of units) {
+        const signature = collapseSignature(unit);
+        const existing = bySignature.get(signature);
+        if (!existing) {
+            bySignature.set(signature, unit);
+            continue;
+        }
+        duplicateCollapses += 1;
+        bySignature.set(signature, {
+            ...existing,
+            evidence_pack: uniqueStrings([...existing.evidence_pack, ...unit.evidence_pack]),
+            entity_backlinks: Array.from(new Map([...existing.entity_backlinks, ...unit.entity_backlinks].map((entry) => [
+                `${entry.entity.toLowerCase()}|${entry.id.toLowerCase()}`,
+                entry
+            ])).values()),
+            affected_entities: uniqueStrings([...existing.affected_entities, ...unit.affected_entities]),
+            affected_documents: uniqueStrings([...existing.affected_documents, ...unit.affected_documents]),
+            affected_postings: uniqueStrings([...existing.affected_postings, ...unit.affected_postings]),
+            affected_accounts: uniqueStrings([...existing.affected_accounts, ...unit.affected_accounts]),
+            affected_counterparties: uniqueStrings([...existing.affected_counterparties, ...unit.affected_counterparties]),
+            affected_contracts: uniqueStrings([...existing.affected_contracts, ...unit.affected_contracts]),
+            snapshot_limitations: uniqueStrings([...existing.snapshot_limitations, ...unit.snapshot_limitations]),
+            severity: unit.severity.score > existing.severity.score ? unit.severity : existing.severity,
+            confidence: unit.confidence.score > existing.confidence.score ? unit.confidence : existing.confidence
+        });
+    }
+    return {
+        problem_units: Array.from(bySignature.values()),
+        duplicate_collapses: duplicateCollapses
+    };
+}
+function buildSummary(units, duplicateCollapses) {
+    const unitTypes = uniqueStrings(units.map((item) => item.problem_unit_type));
+    const typeDistribution = {};
+    const severityDistribution = {
+        low: 0,
+        medium: 0,
+        high: 0
+    };
+    const confidenceDistribution = {
+        low: 0,
+        medium: 0,
+        high: 0
+    };
+    for (const unit of units) {
+        typeDistribution[unit.problem_unit_type] = (typeDistribution[unit.problem_unit_type] ?? 0) + 1;
+        severityDistribution[unit.severity.grade] += 1;
+        confidenceDistribution[unit.confidence.grade] += 1;
+    }
+    return {
+        schema_version: stage2ProblemUnits_1.PROBLEM_UNIT_SUMMARY_SCHEMA_VERSION,
+        units_total: units.length,
+        duplicate_collapses: duplicateCollapses,
+        unit_types: unitTypes,
+        type_distribution: typeDistribution,
+        severity_distribution: severityDistribution,
+        confidence_distribution: confidenceDistribution,
+        primary_unit_type: units[0]?.problem_unit_type ?? null
+    };
+}
+function assembleProblemUnits(input) {
+    const summary = input.summary ?? {};
+    const semanticProfile = extractSemanticProfile(summary);
+    const candidates = buildCandidateEvidence(input.evidence, input.route, {
+        route: input.route,
+        result_type: input.result_type,
+        raw_entities: input.raw_entities ?? [],
+        summary_relation_patterns: semanticProfile.relation_patterns,
+        summary_anomaly_patterns: semanticProfile.anomaly_patterns,
+        risk_factors: uniqueStrings(input.risk_factors ?? [])
+    });
+    const clusters = clusterCandidateEvidence(candidates);
+    const units = clusters.map((cluster, index) => buildProblemUnit(cluster, index));
+    const collapsed = collapseDuplicates(units);
+    return {
+        candidate_evidence: candidates,
+        problem_units: collapsed.problem_units,
+        problem_unit_summary: buildSummary(collapsed.problem_units, collapsed.duplicate_collapses)
+    };
+}

llm_normalizer/backend/dist/services/retrievalResultNormalizer.js

@@ -3,6 +3,7 @@ Object.defineProperty(exports, "__esModule", { value: true });
 exports.normalizeRetrievalResult = normalizeRetrievalResult;
 const config_1 = require("../config");
 const stage1Contracts_1 = require("../types/stage1Contracts");
+const problemUnitAssembler_1 = require("./problemUnitAssembler");
 function toObject(value) {
     if (!value || typeof value !== "object" || Array.isArray(value)) {
         return null;
@@ -59,6 +60,18 @@ function normalizeStringArray(value) {
     }
     return value.map((item) => String(item));
 }
+function mergeSummaryWithProblemUnitMeta(summary, input) {
+    return {
+        ...summary,
+        problem_units_enabled: true,
+        candidate_evidence_count: input.candidateEvidenceCount,
+        problem_units_count: input.problemUnitsCount,
+        problem_unit_types: input.unitTypes,
+        problem_unit_duplicate_collapses: input.duplicateCollapses,
+        problem_unit_severity_distribution: input.severityDistribution,
+        problem_unit_confidence_distribution: input.confidenceDistribution
+    };
+}
 function normalizeConfidence(value) {
     if (value === "high" || value === "medium" || value === "low") {
         return value;
@@ -359,15 +372,18 @@ function normalizeEvidenceItems(fragmentId, requirementIds, route, value) {
     });
 }
 function normalizeRetrievalResult(fragmentId, requirementIds, route, raw) {
-    return {
+    const items = normalizeObjectArray(raw.items);
+    const summary = normalizeSummary(raw.summary);
+    const evidence = normalizeEvidenceItems(fragmentId, requirementIds, route, raw.evidence);
+    const baseResult = {
         fragment_id: fragmentId,
         requirement_ids: requirementIds,
         route,
         status: normalizeStatus(raw.status),
         result_type: normalizeResultType(raw.result_type),
-        items: normalizeObjectArray(raw.items),
-        summary: normalizeSummary(raw.summary),
-        evidence: normalizeEvidenceItems(fragmentId, requirementIds, route, raw.evidence),
+        items,
+        summary,
+        evidence,
         why_included: normalizeStringArray(raw.why_included),
         selection_reason: normalizeStringArray(raw.selection_reason),
         risk_factors: normalizeStringArray(raw.risk_factors),
@@ -376,4 +392,33 @@ function normalizeRetrievalResult(fragmentId, requirementIds, route, raw) {
         limitations: normalizeStringArray(raw.limitations),
         errors: normalizeErrors(raw.errors)
     };
+    if (!config_1.FEATURE_ASSISTANT_PROBLEM_UNITS_V1) {
+        return baseResult;
+    }
+    const assembled = (0, problemUnitAssembler_1.assembleProblemUnits)({
+        route,
+        result_type: baseResult.result_type,
+        evidence,
+        raw_entities: items,
+        summary,
+        risk_factors: baseResult.risk_factors,
+        selection_reason: baseResult.selection_reason,
+        business_interpretation: baseResult.business_interpretation
+    });
+    const enrichedSummary = mergeSummaryWithProblemUnitMeta(summary, {
+        candidateEvidenceCount: assembled.candidate_evidence.length,
+        problemUnitsCount: assembled.problem_units.length,
+        unitTypes: assembled.problem_unit_summary.unit_types,
+        duplicateCollapses: assembled.problem_unit_summary.duplicate_collapses,
+        severityDistribution: assembled.problem_unit_summary.severity_distribution,
+        confidenceDistribution: assembled.problem_unit_summary.confidence_distribution
+    });
+    return {
+        ...baseResult,
+        summary: enrichedSummary,
+        raw_entities: items,
+        candidate_evidence: assembled.candidate_evidence,
+        problem_units: assembled.problem_units,
+        problem_unit_summary: assembled.problem_unit_summary
+    };
 }

llm_normalizer/backend/dist/types/stage2EvalContracts.js

@@ -0,0 +1,41 @@
+"use strict";
+Object.defineProperty(exports, "__esModule", { value: true });
+exports.ASSISTANT_STAGE2_SCORING_RUBRIC_V01 = exports.ASSISTANT_STAGE2_EVAL_RECORD_SCHEMA_VERSION = void 0;
+exports.ASSISTANT_STAGE2_EVAL_RECORD_SCHEMA_VERSION = "assistant_stage2_eval_record_v0_1";
+exports.ASSISTANT_STAGE2_SCORING_RUBRIC_V01 = {
+    problem_unit_precision: [
+        { score: 0, label: "Weak", description: "Problem unit typing is often mismatched against expected case profile." },
+        { score: 3, label: "Mixed", description: "Problem unit typing is partially aligned with expected case profile." },
+        { score: 5, label: "Strong", description: "Problem unit typing is consistently aligned with expected case profile." }
+    ],
+    problem_unit_recall_proxy: [
+        { score: 0, label: "Weak", description: "Expected problem categories are frequently missing from output." },
+        { score: 3, label: "Mixed", description: "Some expected problem categories are captured." },
+        { score: 5, label: "Strong", description: "Most expected problem categories are captured." }
+    ],
+    duplicate_collapse_rate: [
+        { score: 0, label: "Weak", description: "Duplicate collapse is rarely observed when candidate evidence is present." },
+        { score: 3, label: "Mixed", description: "Duplicate collapse works on part of candidate evidence." },
+        { score: 5, label: "Strong", description: "Duplicate collapse consistently reduces noisy candidate evidence." }
+    ],
+    mechanism_coherence_score: [
+        { score: 0, label: "Weak", description: "Mechanism narrative is missing or disconnected from problem units." },
+        { score: 3, label: "Mixed", description: "Mechanism narrative is partially connected to problem units." },
+        { score: 5, label: "Strong", description: "Mechanism narrative is explicit and coherent with problem units." }
+    ],
+    problem_clarity_score: [
+        { score: 0, label: "Weak", description: "Answer framing remains generic and unclear for accountant workflow." },
+        { score: 3, label: "Mixed", description: "Problem framing is partially explicit, but still uneven." },
+        { score: 5, label: "Strong", description: "Problem framing is explicit, scoped, and actionable." }
+    ],
+    problem_first_answer_rate: [
+        { score: 0, label: "Weak", description: "Problem-first rendering is rarely applied on applicable cases." },
+        { score: 3, label: "Mixed", description: "Problem-first rendering is applied inconsistently." },
+        { score: 5, label: "Strong", description: "Problem-first rendering is consistently applied on applicable cases." }
+    ],
+    entity_leakage_rate: [
+        { score: 0, label: "High Leakage", description: "User-facing answers frequently leak raw technical identifiers." },
+        { score: 3, label: "Moderate Leakage", description: "Technical identifier leakage appears in a minority of answers." },
+        { score: 5, label: "Low Leakage", description: "User-facing answers rarely leak raw technical identifiers." }
+    ]
+};

llm_normalizer/backend/dist/types/stage2ProblemUnits.js

@@ -0,0 +1,10 @@
+"use strict";
+Object.defineProperty(exports, "__esModule", { value: true });
+exports.INVESTIGATION_MAX_FOCUS_PROBLEM_TYPES = exports.INVESTIGATION_MAX_PROBLEM_UNIT_BACKLINKS = exports.INVESTIGATION_MAX_RESOLVED_PROBLEM_UNITS = exports.INVESTIGATION_MAX_ACTIVE_PROBLEM_UNITS = exports.PROBLEM_UNIT_SUMMARY_SCHEMA_VERSION = exports.PROBLEM_UNIT_SCHEMA_VERSION = exports.CANDIDATE_EVIDENCE_SCHEMA_VERSION = void 0;
+exports.CANDIDATE_EVIDENCE_SCHEMA_VERSION = "candidate_evidence_v0_1";
+exports.PROBLEM_UNIT_SCHEMA_VERSION = "problem_unit_v0_1";
+exports.PROBLEM_UNIT_SUMMARY_SCHEMA_VERSION = "problem_unit_summary_v0_1";
+exports.INVESTIGATION_MAX_ACTIVE_PROBLEM_UNITS = 8;
+exports.INVESTIGATION_MAX_RESOLVED_PROBLEM_UNITS = 16;
+exports.INVESTIGATION_MAX_PROBLEM_UNIT_BACKLINKS = 12;
+exports.INVESTIGATION_MAX_FOCUS_PROBLEM_TYPES = 6;

llm_normalizer/backend/src/config.ts

@@ -51,6 +51,22 @@ export const FEATURE_ASSISTANT_ACCOUNTANT_EVAL_V1 = toBooleanFlag(
   process.env.FEATURE_ASSISTANT_ACCOUNTANT_EVAL_V1,
   true
 );
+export const FEATURE_ASSISTANT_PROBLEM_UNITS_V1 = toBooleanFlag(
+  process.env.FEATURE_ASSISTANT_PROBLEM_UNITS_V1,
+  false
+);
+export const FEATURE_ASSISTANT_PROBLEM_CENTRIC_ANSWER_V1 = toBooleanFlag(
+  process.env.FEATURE_ASSISTANT_PROBLEM_CENTRIC_ANSWER_V1,
+  false
+);
+export const FEATURE_ASSISTANT_PROBLEM_UNIT_CONTINUITY_V1 = toBooleanFlag(
+  process.env.FEATURE_ASSISTANT_PROBLEM_UNIT_CONTINUITY_V1,
+  false
+);
+export const FEATURE_ASSISTANT_STAGE2_EVAL_V1 = toBooleanFlag(
+  process.env.FEATURE_ASSISTANT_STAGE2_EVAL_V1,
+  false
+);
 
 export const DATA_DIR = process.env.DATA_DIR ?? path.resolve(MODULE_ROOT, "data");
 export const TRACES_DIR = path.resolve(DATA_DIR, "traces");

llm_normalizer/backend/src/services/answerComposer.ts

@@ -8,6 +8,9 @@ import type {
 } from "../types/assistant";
 import type { RouteHintSummary } from "../types/normalizer";
 import type { AnswerStructureV11, EvidenceConfidence, EvidenceItem, EvidenceLimitationReasonCode } from "../types/stage1Contracts";
+import type { ProblemUnit, ProblemUnitSummary, ProblemUnitType } from "../types/stage2ProblemUnits";
+
+type ProblemAnswerMode = "stage1_policy_v11" | "stage2_problem_centric_v1";
 
 interface ComposeAnswerInput {
   userMessage: string;
@@ -17,6 +20,7 @@ interface ComposeAnswerInput {
   coverageReport: RequirementCoverageReport;
   groundingCheck: AnswerGroundingCheck;
   enableAnswerPolicyV11?: boolean;
+  enableProblemCentricAnswerV1?: boolean;
 }
 
 interface ComposeAnswerOutput {
@@ -24,6 +28,10 @@ interface ComposeAnswerOutput {
   fallback_type: AssistantFallbackType;
   reply_type: AssistantReplyType;
   answer_structure_v11?: AnswerStructureV11;
+  problem_centric_answer_applied?: boolean;
+  problem_units_used_count?: number;
+  problem_answer_mode?: ProblemAnswerMode;
+  problem_unit_ids_used?: string[];
 }
 
 function fallbackFromSummary(routeSummary: RouteHintSummary | null): AssistantFallbackType {
@@ -37,6 +45,84 @@ function uniqueStrings(values: string[], limit = 6): string[] {
   return Array.from(new Set(values.map((item) => item.trim()).filter(Boolean))).slice(0, limit);
 }
 
+const UUID_PATTERN = /\b[0-9a-f]{8}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{12}\b/gi;
+const LONG_HEX_PATTERN = /\b[0-9a-f]{24,}\b/gi;
+const RAW_REF_BLOB_PATTERN = /\bevidence_source_ref_v1\|[^\s,;]+/gi;
+const RAW_REF_TOKEN_PATTERN = /\b(?:source_ref|canonical_ref|entity_id|fragment_id|guid|uuid)\b/gi;
+
+function looksLikeMojibake(value: string): boolean {
+  const text = String(value ?? "");
+  if (!text.trim()) {
+    return false;
+  }
+  if (/(?:Р.|С.){5,}/u.test(text)) {
+    return true;
+  }
+  if (/[ЃѓЂђЌќЎў]/u.test(text)) {
+    return true;
+  }
+  return false;
+}
+
+function looksLikeTechnicalIdentifier(value: string): boolean {
+  const text = String(value ?? "").trim();
+  if (!text) {
+    return false;
+  }
+  if (UUID_PATTERN.test(text)) {
+    UUID_PATTERN.lastIndex = 0;
+    return true;
+  }
+  UUID_PATTERN.lastIndex = 0;
+  if (LONG_HEX_PATTERN.test(text)) {
+    LONG_HEX_PATTERN.lastIndex = 0;
+    return true;
+  }
+  LONG_HEX_PATTERN.lastIndex = 0;
+  return /(?:evidence_source_ref_v1\||cmp%3a|batch_refresh_then_store:|^cmp:)/i.test(text);
+}
+
+function scrubRawTechnicalRefs(value: string): string {
+  const raw = String(value ?? "").trim();
+  if (!raw) {
+    return "";
+  }
+  return raw
+    .replace(RAW_REF_BLOB_PATTERN, "linked source")
+    .replace(UUID_PATTERN, "[id]")
+    .replace(LONG_HEX_PATTERN, "[id]")
+    .replace(RAW_REF_TOKEN_PATTERN, "reference")
+    .replace(/\(\s*\[id\]\s*\)/g, "")
+    .replace(/\[\s*id\s*\](?:\s*,\s*\[\s*id\s*\])+/g, "[id]")
+    .replace(/\s{2,}/g, " ")
+    .trim();
+}
+
+function sanitizeUserFacingReply(value: string): string {
+  return scrubRawTechnicalRefs(value)
+    .replace(/[ \t]+\n/g, "\n")
+    .replace(/\n{3,}/g, "\n\n")
+    .trim();
+}
+
+function sanitizeUserText(value: string): string | null {
+  const normalized = scrubRawTechnicalRefs(String(value ?? "").replace(/\s+/g, " ").trim());
+  if (!normalized) {
+    return null;
+  }
+  if (looksLikeMojibake(normalized)) {
+    return null;
+  }
+  return normalized;
+}
+
+function sanitizeUserLines(values: string[], limit = 6): string[] {
+  const cleaned = values
+    .map((item) => sanitizeUserText(item))
+    .filter((item): item is string => Boolean(item));
+  return uniqueStrings(cleaned, limit);
+}
+
 function formatList(items: string[]): string {
   if (items.length === 0) {
     return "";
@@ -44,15 +130,28 @@ function formatList(items: string[]): string {
   return items.map((item) => `- ${item}`).join("\n");
 }
 
+function formatSafeItemLine(entity: unknown, sourceId: unknown, riskScore?: unknown): string {
+  const entityLabel = sanitizeUserText(String(entity ?? "")) ?? "Record";
+  const idRaw = String(sourceId ?? "").trim();
+  const exposeId = idRaw.length > 0 && !looksLikeTechnicalIdentifier(idRaw);
+  const subject = exposeId ? `${entityLabel} (${idRaw})` : entityLabel;
+  if (riskScore !== undefined) {
+    return `${subject} - risk ${String(riskScore)}.`;
+  }
+  return `${subject}.`;
+}
+
 function extractTopFacts(results: UnifiedRetrievalResult[]): string[] {
   const lines: string[] = [];
   for (const result of results.filter((item) => item.status === "ok").slice(0, 3)) {
     if (result.result_type === "chain") {
       const top = result.items.slice(0, 3).map((item) => {
-        const counterparty = String(item.counterparty_id ?? "не указан");
+        const counterparty = String(item.counterparty_id ?? "").trim();
         const operations = String(item.operations_count ?? "0");
         const docs = String(item.document_refs_count ?? "0");
-        return `Контрагент ${counterparty}: операций ${operations}, документов в связке ${docs}.`;
+        const counterpartyLabel =
+          counterparty.length > 0 && !looksLikeTechnicalIdentifier(counterparty) ? `Counterparty ${counterparty}` : "Counterparty";
+        return `${counterpartyLabel}: operations ${operations}, linked docs ${docs}.`;
       });
       lines.push(...top);
       continue;
@@ -60,46 +159,46 @@ function extractTopFacts(results: UnifiedRetrievalResult[]): string[] {
     if (result.result_type === "ranking") {
       const top = result.items
         .slice(0, 5)
-        .map((item) => `${item.rank ?? "•"}. ${String(item.entity ?? "Сущность")} — ${String(item.records_count ?? 0)}.`);
+        .map((item) => `${item.rank ?? "*"}. ${String(item.entity ?? "Entity")} - ${String(item.records_count ?? 0)}.`);
       lines.push(...top);
       continue;
     }
     if (result.result_type === "list") {
       const top = result.items.slice(0, 5).map((item) => {
         if (item.risk_score !== undefined) {
-          return `${String(item.source_entity ?? "Запись")} (${String(item.source_id ?? "")}) — риск ${String(item.risk_score)}.`;
+          return formatSafeItemLine(item.source_entity ?? "Record", item.source_id ?? "", item.risk_score);
         }
-        return `${String(item.source_entity ?? "Запись")} (${String(item.source_id ?? "")}).`;
+        return formatSafeItemLine(item.source_entity ?? "Record", item.source_id ?? "");
       });
       lines.push(...top);
       continue;
     }
     const top = result.items
       .slice(0, 3)
-      .map((item) => `${String(item.source_entity ?? "Запись")} (${String(item.source_id ?? "")}).`);
+      .map((item) => formatSafeItemLine(item.source_entity ?? "Record", item.source_id ?? ""));
     lines.push(...top);
   }
-  return lines;
+  return sanitizeUserLines(lines, 8);
 }
 
 function extractWhyIncluded(results: UnifiedRetrievalResult[]): string[] {
-  return uniqueStrings(results.flatMap((item) => item.why_included));
+  return sanitizeUserLines(results.flatMap((item) => item.why_included));
 }
 
 function extractSelectionReasons(results: UnifiedRetrievalResult[]): string[] {
-  return uniqueStrings(results.flatMap((item) => item.selection_reason));
+  return sanitizeUserLines(results.flatMap((item) => item.selection_reason));
 }
 
 function extractRiskFactors(results: UnifiedRetrievalResult[]): string[] {
-  return uniqueStrings(results.flatMap((item) => item.risk_factors));
+  return sanitizeUserLines(results.flatMap((item) => item.risk_factors));
 }
 
 function extractBusinessInterpretation(results: UnifiedRetrievalResult[]): string[] {
-  return uniqueStrings(results.flatMap((item) => item.business_interpretation));
+  return sanitizeUserLines(results.flatMap((item) => item.business_interpretation));
 }
 
 function extractLimitations(results: UnifiedRetrievalResult[]): string[] {
-  return uniqueStrings(results.flatMap((item) => item.limitations));
+  return sanitizeUserLines(results.flatMap((item) => item.limitations), 10);
 }
 
 function summaryValue(result: UnifiedRetrievalResult, key: string): unknown {
@@ -116,6 +215,63 @@ function summaryString(result: UnifiedRetrievalResult, key: string): string | nu
   return typeof value === "string" ? value : null;
 }
 
+function summaryNumber(result: UnifiedRetrievalResult, key: string): number | null {
+  const value = summaryValue(result, key);
+  return typeof value === "number" && Number.isFinite(value) ? value : null;
+}
+
+function summaryStringArray(result: UnifiedRetrievalResult, key: string): string[] {
+  const value = summaryValue(result, key);
+  if (!Array.isArray(value)) {
+    return [];
+  }
+  return sanitizeUserLines(value.map((item) => String(item)), 6);
+}
+
+function buildFallbackWhyIncluded(results: UnifiedRetrievalResult[]): string[] {
+  const lines: string[] = [];
+  for (const result of results.slice(0, 2)) {
+    const routeFocus = summaryString(result, "route_focus");
+    const sourceRecords = summaryNumber(result, "source_records");
+    const filteredRecords = summaryNumber(result, "filtered_records_after_narrowing");
+    const checkedRecords = summaryNumber(result, "checked_records");
+
+    if (routeFocus) {
+      lines.push(`Проверка выполнена по профилю ${routeFocus}.`);
+    }
+    if (sourceRecords !== null && filteredRecords !== null && filteredRecords < sourceRecords) {
+      lines.push(`Применено сужение выборки: ${filteredRecords} из ${sourceRecords} записей.`);
+    }
+    if (checkedRecords !== null) {
+      lines.push(`Проверено записей в текущем проходе: ${checkedRecords}.`);
+    }
+  }
+
+  return sanitizeUserLines(lines, 4);
+}
+
+function buildFallbackSelectionReasons(results: UnifiedRetrievalResult[]): string[] {
+  const lines: string[] = [];
+  for (const result of results.slice(0, 2)) {
+    if (summaryBoolean(result, "semantic_narrowing_applied")) {
+      lines.push("Отбор выполнен по семантическому сужению предметной области.");
+    }
+    const rankingBasis = summaryStringArray(result, "ranking_basis");
+    if (rankingBasis.length > 0) {
+      lines.push(`Ранжирование основано на: ${rankingBasis.join(", ")}.`);
+    }
+    if (summaryBoolean(result, "broad_guard_applied")) {
+      lines.push("Применен broad-query guard для контроля ложной точности.");
+    }
+  }
+
+  if (lines.length === 0) {
+    lines.push("Отбор выполнен по совпадению предметных сигналов и доступной evidence-опоры.");
+  }
+
+  return sanitizeUserLines(lines, 4);
+}
+
 function suggestNextStep(requirements: AssistantRequirement[], coverage: RequirementCoverageReport): string[] {
   const next: string[] = [];
   if (coverage.clarification_needed_for.length > 0) {
@@ -165,10 +321,155 @@ interface MissingAnchors {
   anomalyType: boolean;
 }
 
+const PROBLEM_HEAVY_TYPES = new Set<ProblemUnitType>([
+  "document_conflict",
+  "broken_chain_segment",
+  "lifecycle_anomaly_node",
+  "unresolved_settlement_cluster",
+  "period_risk_cluster",
+  "cross_branch_inconsistency_cluster"
+]);
+
 function flattenEvidence(results: UnifiedRetrievalResult[]): EvidenceItem[] {
   return results.flatMap((item) => item.evidence);
 }
 
+function flattenProblemUnits(results: UnifiedRetrievalResult[]): ProblemUnit[] {
+  const units: ProblemUnit[] = [];
+  for (const result of results) {
+    if (!Array.isArray(result.problem_units)) {
+      continue;
+    }
+    units.push(...result.problem_units);
+  }
+  const byId = new Map<string, ProblemUnit>();
+  for (const unit of units) {
+    byId.set(unit.problem_unit_id, unit);
+  }
+  return Array.from(byId.values());
+}
+
+function selectProblemUnitSummary(results: UnifiedRetrievalResult[]): ProblemUnitSummary | null {
+  let selected: ProblemUnitSummary | null = null;
+  for (const result of results) {
+    if (!result.problem_unit_summary) {
+      continue;
+    }
+    if (!selected || result.problem_unit_summary.units_total > selected.units_total) {
+      selected = result.problem_unit_summary;
+    }
+  }
+  return selected;
+}
+
+function formatAffectedScope(unit: ProblemUnit): string {
+  const scopeParts: string[] = [];
+  if (unit.affected_accounts.length > 0) {
+    scopeParts.push(`счета: ${unit.affected_accounts.slice(0, 2).join(", ")}`);
+  }
+  if (unit.affected_counterparties.length > 0) {
+    scopeParts.push(`контрагенты: ${unit.affected_counterparties.slice(0, 2).join(", ")}`);
+  }
+  if (unit.affected_documents.length > 0) {
+    scopeParts.push(`документы: ${unit.affected_documents.slice(0, 2).join(", ")}`);
+  }
+  if (scopeParts.length === 0 && unit.affected_entities.length > 0) {
+    scopeParts.push(`объекты: ${unit.affected_entities.slice(0, 2).join(", ")}`);
+  }
+  if (scopeParts.length === 0) {
+    return "затронутый контур требует уточнения";
+  }
+  return scopeParts.join("; ");
+}
+
+function buildProblemCentricActions(input: {
+  units: ProblemUnit[];
+  mode: PolicyMode;
+  missingAnchors: MissingAnchors;
+  coverageReport: RequirementCoverageReport;
+}): string[] {
+  const actions: string[] = [];
+  const unitTypes = new Set(input.units.map((item) => item.problem_unit_type));
+
+  if (unitTypes.has("broken_chain_segment")) {
+    actions.push("Проверьте связку выписка -> документ -> проводка по проблемным участкам цепочки.");
+  }
+  if (unitTypes.has("unresolved_settlement_cluster")) {
+    actions.push("Сверьте хвосты по расчетам: закрылся ли документ оплаты корректным закрывающим документом.");
+  }
+  if (unitTypes.has("period_risk_cluster")) {
+    actions.push("Оцените влияние дефекта на закрытие периода и корректность регламентных операций.");
+  }
+  if (unitTypes.has("cross_branch_inconsistency_cluster")) {
+    actions.push("Сверьте противоречия между документами, проводками и регистрами по НДС/межконтурным связям.");
+  }
+  if (unitTypes.has("lifecycle_anomaly_node")) {
+    actions.push("Проверьте lifecycle объекта: ожидаемый этап не должен оставаться в partially_linked состоянии.");
+  }
+
+  if (input.mode === "clarification_required") {
+    if (input.missingAnchors.period) {
+      actions.push("Уточните период проверки, чтобы зафиксировать границы проблемного контура.");
+    }
+    if (input.missingAnchors.account) {
+      actions.push("Уточните счет или группу счетов для предметной локализации дефекта.");
+    }
+    if (input.missingAnchors.documentOrObject) {
+      actions.push("Укажите конкретный документ или объект трассировки для проверки механизма отклонения.");
+    }
+    if (input.missingAnchors.counterparty) {
+      actions.push("Укажите контрагента/договор, чтобы проверить хвосты и разрывы на конкретной связке.");
+    }
+  }
+
+  if (input.coverageReport.requirements_uncovered.length > 0) {
+    actions.push(`Закройте непокрытые требования: ${input.coverageReport.requirements_uncovered.join(", ")}.`);
+  }
+
+  return uniqueStrings(actions, 6);
+}
+
+function buildProblemCentricClarifications(input: {
+  units: ProblemUnit[];
+  missingAnchors: MissingAnchors;
+  coverageReport: RequirementCoverageReport;
+  mode: PolicyMode;
+}): string[] {
+  if (input.mode !== "clarification_required") {
+    return [];
+  }
+
+  const questions: string[] = [];
+  const unitTypes = new Set(input.units.map((item) => item.problem_unit_type));
+
+  if (input.missingAnchors.period) {
+    questions.push("Уточните период (например, 2020-06), в котором нужно проверить проблемный кластер.");
+  }
+  if (input.missingAnchors.account) {
+    questions.push("Уточните счет или связку счетов (например, 51/60), где вы ожидаете дефект.");
+  }
+  if (input.missingAnchors.documentOrObject) {
+    questions.push("Укажите документ/объект, от которого нужно строить проверку цепочки.");
+  }
+  if (input.missingAnchors.counterparty) {
+    questions.push("Укажите контрагента или договор, по которому проверить незакрытую экспозицию.");
+  }
+  if (unitTypes.has("broken_chain_segment")) {
+    questions.push("Уточните участок цепочки: выписка, платежный документ или проводка.");
+  }
+  if (unitTypes.has("period_risk_cluster")) {
+    questions.push("Уточните, какой этап закрытия периода критичен: начисление, закрытие счетов или НДС-блок.");
+  }
+  if (unitTypes.has("unresolved_settlement_cluster")) {
+    questions.push("Уточните, интересуют хвосты поставщиков, покупателей или оба направления.");
+  }
+  if (input.coverageReport.clarification_needed_for.length > 0) {
+    questions.push(`Закройте уточнения для требований: ${input.coverageReport.clarification_needed_for.join(", ")}.`);
+  }
+
+  return uniqueStrings(questions, 6);
+}
+
 function buildClaimEvidenceLinks(results: UnifiedRetrievalResult[]): NonNullable<AnswerStructureV11["evidence_block"]["claim_evidence_links"]> {
   const byClaim = new Map<string, string[]>();
   for (const evidence of flattenEvidence(results)) {
@@ -333,7 +634,7 @@ function buildRecommendedActions(input: {
 }): string[] {
   const actions: string[] = [];
   if (input.mode === "focused_grounded") {
-    actions.push("Проверьте 1-2 ключевые записи по source_ref и зафиксируйте итог в рабочем файле проверки.");
+    actions.push("Проверьте 1-2 ключевые записи в учетной базе и зафиксируйте итог в рабочем файле проверки.");
   }
   if (input.mode === "broad_partial") {
     actions.push("Сузьте запрос до периода + счета или периода + документа и повторите проверку.");
@@ -357,7 +658,7 @@ function buildRecommendedActions(input: {
     actions.push("Проверьте source mapping для связей document/register по указанным ref.");
   }
   if (input.sourceRefs.length > 0) {
-    actions.push(`Начните проверку с source_ref: ${input.sourceRefs.slice(0, 2).join(", ")}.`);
+    actions.push(`Начните проверку с ${input.sourceRefs.length} подтвержденных записей и сверьте их с первичными документами.`);
   }
 
   return uniqueStrings(actions, 6);
@@ -520,6 +821,167 @@ function buildDirectAnswer(input: {
   return "Не удалось сформировать обоснованный ответ; нужно уточнение запроса.";
 }
 
+function buildProblemCentricAnswerSummary(input: {
+  mode: PolicyMode;
+  weakUnits: boolean;
+  summary: ProblemUnitSummary | null;
+}): string {
+  if (input.mode === "clarification_required") {
+    return "Выявлены проблемные кластеры, но для надежного вывода требуется предметное уточнение фокуса.";
+  }
+  if (input.weakUnits) {
+    return "Сформирован problem-centric срез с ограниченной опорой; вывод предварительный и требует до-проверки.";
+  }
+  if (input.summary?.units_total && input.summary.units_total > 1) {
+    return `Сформирован problem-centric срез: выделено ${input.summary.units_total} проблемных кластера с приоритетами.`;
+  }
+  return "Сформирован problem-centric срез: выделен ключевой проблемный кластер и затронутый контур.";
+}
+
+function buildProblemCentricDirectAnswer(input: {
+  mode: PolicyMode;
+  units: ProblemUnit[];
+  weakUnits: boolean;
+}): string {
+  const lead =
+    input.mode === "clarification_required"
+      ? "Обнаружены проблемные зоны, но без уточнения якорей сильный factual-вывод преждевременен."
+      : input.weakUnits
+        ? "Выделены проблемные зоны с ограниченной надежностью; вывод дан в ограниченном режиме."
+        : "Выделены ключевые проблемные зоны и их влияние на учетный контур.";
+
+  const unitLines = input.units.map((unit) => {
+    const scope = formatAffectedScope(unit);
+    return `- ${unit.title}: ${unit.business_defect_class}; ${scope}; severity=${unit.severity.grade}, confidence=${unit.confidence.grade}.`;
+  });
+
+  if (unitLines.length === 0) {
+    return `${lead}\nПроблемные кластеры не удалось детализировать в текущем срезе.`;
+  }
+
+  return [lead, "Проблемные кластеры:", ...unitLines].join("\n");
+}
+
+function buildProblemCentricAnswerStructure(input: {
+  mode: PolicyMode;
+  selectedUnits: ProblemUnit[];
+  problemSummary: ProblemUnitSummary | null;
+  evidenceItems: EvidenceItem[];
+  claimEvidenceLinks: NonNullable<AnswerStructureV11["evidence_block"]["claim_evidence_links"]>;
+  limitationReasonCodes: EvidenceLimitationReasonCode[];
+  groundingCheck: AnswerGroundingCheck;
+  retrievalResults: UnifiedRetrievalResult[];
+  missingAnchors: MissingAnchors;
+  coverageReport: RequirementCoverageReport;
+}): AnswerStructureV11 {
+  const weakUnits = input.selectedUnits.every((item) => item.confidence.grade === "low");
+  const unitMechanismNotes = uniqueStrings(
+    input.selectedUnits
+      .map((item) => item.mechanism_summary)
+      .filter((item) => typeof item === "string" && item.trim().length > 0),
+    6
+  );
+  const sourceRefs = uniqueStrings(
+    input.evidenceItems
+      .map((item) => item.source_ref?.canonical_ref)
+      .filter((item): item is string => typeof item === "string" && item.trim().length > 0),
+    6
+  );
+  const evidenceIds = uniqueStrings(input.evidenceItems.map((item) => item.evidence_id), 10);
+
+  const mechanismStatus: AnswerStructureV11["mechanism_block"]["status"] =
+    unitMechanismNotes.length === 0
+      ? "unresolved"
+      : weakUnits || input.limitationReasonCodes.includes("missing_mechanism")
+        ? "limited"
+        : "grounded";
+
+  const problemSpecificLimitations: string[] = [];
+  if (weakUnits) {
+    problemSpecificLimitations.push("Problem units remain weak-confidence; conclusions are intentionally limited.");
+  }
+  if (input.problemSummary?.duplicate_collapses && input.problemSummary.duplicate_collapses > 0) {
+    problemSpecificLimitations.push("Part of the problem signal was merged due to duplicate collapse.");
+  }
+
+  const limitations = uniqueStrings(
+    [
+      ...problemSpecificLimitations,
+      ...input.limitationReasonCodes.map((code) => limitationReasonToText(code)),
+      ...extractLimitations(input.retrievalResults),
+      ...input.groundingCheck.reasons
+    ],
+    10
+  );
+
+  const openUncertainties = uniqueStrings(
+    [
+      ...input.groundingCheck.missing_requirements,
+      ...(input.mode === "clarification_required" && input.missingAnchors.period ? ["missing_anchor:period"] : []),
+      ...(input.mode === "clarification_required" && input.missingAnchors.account ? ["missing_anchor:account"] : []),
+      ...(input.mode === "clarification_required" && input.missingAnchors.documentOrObject
+        ? ["missing_anchor:document_or_object"]
+        : []),
+      ...(input.mode === "clarification_required" && input.missingAnchors.counterparty ? ["missing_anchor:counterparty"] : [])
+    ],
+    8
+  );
+
+  return {
+    schema_version: "answer_structure_v1_1",
+    answer_summary: buildProblemCentricAnswerSummary({
+      mode: input.mode,
+      weakUnits,
+      summary: input.problemSummary
+    }),
+    direct_answer: buildProblemCentricDirectAnswer({
+      mode: input.mode,
+      units: input.selectedUnits,
+      weakUnits
+    }),
+    mechanism_block: {
+      status: mechanismStatus,
+      mechanism_notes: unitMechanismNotes,
+      limitation_reason_codes: input.limitationReasonCodes
+    },
+    evidence_block: {
+      evidence_ids: evidenceIds,
+      source_refs: sourceRefs,
+      mechanism_notes: unitMechanismNotes,
+      coverage_note:
+        input.coverageReport.requirements_total > 0 &&
+        input.coverageReport.requirements_total === input.coverageReport.requirements_covered &&
+        input.coverageReport.requirements_uncovered.length === 0 &&
+        input.coverageReport.requirements_partially_covered.length === 0
+          ? "coverage_full_or_near_full"
+          : "coverage_partial_or_limited",
+      ...(input.claimEvidenceLinks.length > 0
+        ? {
+            claim_evidence_links: input.claimEvidenceLinks
+          }
+        : {})
+    },
+    uncertainty_block: {
+      open_uncertainties: openUncertainties,
+      limitations
+    },
+    next_step_block: {
+      recommended_actions: buildProblemCentricActions({
+        units: input.selectedUnits,
+        mode: input.mode,
+        missingAnchors: input.missingAnchors,
+        coverageReport: input.coverageReport
+      }),
+      clarification_questions: buildProblemCentricClarifications({
+        units: input.selectedUnits,
+        missingAnchors: input.missingAnchors,
+        coverageReport: input.coverageReport,
+        mode: input.mode
+      })
+    }
+  };
+}
+
 function renderPolicyReply(structure: AnswerStructureV11): string {
   const mechanismLines: string[] = [`status=${structure.mechanism_block.status}`];
   if (structure.mechanism_block.mechanism_notes.length > 0) {
@@ -532,18 +994,18 @@ function renderPolicyReply(structure: AnswerStructureV11): string {
     mechanismLines.push("mechanism_note is intentionally omitted due to weak or missing mechanism evidence");
   }
 
+  const sourceRefCount = Array.isArray(structure.evidence_block.source_refs) ? structure.evidence_block.source_refs.length : 0;
+  const claimLinkCount = Array.isArray(structure.evidence_block.claim_evidence_links)
+    ? structure.evidence_block.claim_evidence_links.length
+    : 0;
   const evidenceLines: string[] = [
     `coverage=${structure.evidence_block.coverage_note}`,
-    `evidence_ids=${structure.evidence_block.evidence_ids.length > 0 ? structure.evidence_block.evidence_ids.join(", ") : "none"}`
+    `supporting_evidence_count=${structure.evidence_block.evidence_ids.length}`,
+    `supporting_source_count=${sourceRefCount}`,
+    `claim_support_links=${claimLinkCount}`
   ];
-  if (Array.isArray(structure.evidence_block.source_refs) && structure.evidence_block.source_refs.length > 0) {
-    evidenceLines.push(`source_refs=${structure.evidence_block.source_refs.join(", ")}`);
-  }
-  if (Array.isArray(structure.evidence_block.claim_evidence_links) && structure.evidence_block.claim_evidence_links.length > 0) {
-    const compactLinks = structure.evidence_block.claim_evidence_links
-      .slice(0, 4)
-      .map((item) => `${item.claim_ref}:${item.evidence_ids.join("|")}`);
-    evidenceLines.push(`claim_evidence_links=${compactLinks.join("; ")}`);
+  if (sourceRefCount > 0) {
+    evidenceLines.push("Detailed source references are available in debug payload.");
   }
 
   const uncertaintyLines = [
@@ -562,16 +1024,18 @@ function renderPolicyReply(structure: AnswerStructureV11): string {
     nextStepLines.push("No additional action is required for this scoped answer.");
   }
 
-  return [
-    `Answer summary: ${structure.answer_summary}`,
-    `Direct answer:\n${structure.direct_answer}`,
-    `Mechanism block:\n${formatList(mechanismLines)}`,
-    `Evidence block:\n${formatList(evidenceLines)}`,
-    `Uncertainty block:\n${formatList(uncertaintyLines)}`,
-    `Next step block:\n${formatList(nextStepLines)}`
-  ]
-    .filter(Boolean)
-    .join("\n\n");
+  return sanitizeUserFacingReply(
+    [
+      `Answer summary: ${structure.answer_summary}`,
+      `Direct answer:\n${structure.direct_answer}`,
+      `Mechanism block:\n${formatList(mechanismLines)}`,
+      `Evidence block:\n${formatList(evidenceLines)}`,
+      `Uncertainty block:\n${formatList(uncertaintyLines)}`,
+      `Next step block:\n${formatList(nextStepLines)}`
+    ]
+      .filter(Boolean)
+      .join("\n\n")
+  );
 }
 
 function composeAssistantAnswerV11(input: ComposeAnswerInput): ComposeAnswerOutput {
@@ -595,8 +1059,15 @@ function composeAssistantAnswerV11(input: ComposeAnswerInput): ComposeAnswerOutp
       .filter((item): item is string => typeof item === "string" && item.trim().length > 0),
     6
   );
+  const problemUnits = flattenProblemUnits(input.retrievalResults);
+  const problemUnitSummary = selectProblemUnitSummary(input.retrievalResults);
+  const problemHeavyUnits = problemUnits.filter((item) => PROBLEM_HEAVY_TYPES.has(item.problem_unit_type));
+  const selectedProblemUnits = problemHeavyUnits.slice(0, 4);
   const claimEvidenceLinks = buildClaimEvidenceLinks(input.retrievalResults);
   const aggregateEvidenceConfidence = aggregateConfidence(input.retrievalResults, evidenceItems);
+  const lowConfidenceSignals = evidenceItems.filter((item) => item.confidence === "low").length;
+  const lowConfidenceShare = evidenceItems.length > 0 ? lowConfidenceSignals / evidenceItems.length : 0;
+  const lowConfidenceConcentration = lowConfidenceShare >= 0.6;
   const hasSupport =
     okResults.length > 0 ||
     partialResults.length > 0 ||
@@ -637,6 +1108,51 @@ function composeAssistantAnswerV11(input: ComposeAnswerInput): ComposeAnswerOutp
   });
 
   const missingAnchors = detectMissingAnchors(input.userMessage);
+  const hasProblemWeakSignal =
+    policySignals.narrowing_strength !== "strong" ||
+    policySignals.minimum_evidence_failed ||
+    limitationReasonCodes.includes("missing_mechanism") ||
+    limitationReasonCodes.includes("weak_source_mapping") ||
+    aggregateEvidenceConfidence === "low" ||
+    lowConfidenceConcentration;
+  const hardBlockedMode = decision.mode === "out_of_scope" || decision.mode === "route_mismatch" || decision.mode === "backend_error";
+  const problemCentricModeEligible =
+    decision.mode === "broad_partial" ||
+    decision.mode === "clarification_required" ||
+    (decision.mode === "focused_grounded" && hasProblemWeakSignal);
+  const shouldUseProblemCentricAnswer =
+    Boolean(input.enableProblemCentricAnswerV1) &&
+    !hardBlockedMode &&
+    problemCentricModeEligible &&
+    (!focusedStrong || hasProblemWeakSignal) &&
+    selectedProblemUnits.length > 0;
+
+  if (shouldUseProblemCentricAnswer) {
+    const problemCentricStructure = buildProblemCentricAnswerStructure({
+      mode: decision.mode,
+      selectedUnits: selectedProblemUnits,
+      problemSummary: problemUnitSummary,
+      evidenceItems,
+      claimEvidenceLinks,
+      limitationReasonCodes,
+      groundingCheck: input.groundingCheck,
+      retrievalResults: input.retrievalResults,
+      missingAnchors,
+      coverageReport: input.coverageReport
+    });
+
+    return {
+      assistant_reply: renderPolicyReply(problemCentricStructure),
+      fallback_type: decision.fallback_type,
+      reply_type: decision.reply_type,
+      answer_structure_v11: problemCentricStructure,
+      problem_centric_answer_applied: true,
+      problem_units_used_count: selectedProblemUnits.length,
+      problem_answer_mode: "stage2_problem_centric_v1",
+      problem_unit_ids_used: selectedProblemUnits.map((item) => item.problem_unit_id)
+    };
+  }
+
   const clarificationQuestions = buildClarificationQuestions({
     mode: decision.mode,
     missingAnchors,
@@ -725,14 +1241,20 @@ function composeAssistantAnswerV11(input: ComposeAnswerInput): ComposeAnswerOutp
     assistant_reply: renderPolicyReply(answerStructure),
     fallback_type: decision.fallback_type,
     reply_type: decision.reply_type,
-    answer_structure_v11: answerStructure
+    answer_structure_v11: answerStructure,
+    problem_centric_answer_applied: false,
+    problem_units_used_count: 0,
+    problem_answer_mode: "stage1_policy_v11"
   };
 }
 
 function composeExplainableAnswer(input: ComposeAnswerInput, scopeLabel: "full" | "partial"): string {
   const facts = extractTopFacts(input.retrievalResults);
-  const whyIncluded = extractWhyIncluded(input.retrievalResults);
-  const selectionReasons = extractSelectionReasons(input.retrievalResults);
+  const whyIncludedRaw = extractWhyIncluded(input.retrievalResults);
+  const selectionReasonsRaw = extractSelectionReasons(input.retrievalResults);
+  const whyIncluded = whyIncludedRaw.length > 0 ? whyIncludedRaw : buildFallbackWhyIncluded(input.retrievalResults);
+  const selectionReasons =
+    selectionReasonsRaw.length > 0 ? selectionReasonsRaw : buildFallbackSelectionReasons(input.retrievalResults);
   const riskFactors = extractRiskFactors(input.retrievalResults);
   const interpretation = extractBusinessInterpretation(input.retrievalResults);
   const limitations = uniqueStrings([...extractLimitations(input.retrievalResults), ...input.groundingCheck.reasons]);
@@ -877,3 +1399,4 @@ export function composeAssistantAnswer(input: ComposeAnswerInput): ComposeAnswer
     reply_type: "backend_error"
   };
 }
+

llm_normalizer/backend/src/services/assistantService.ts

@@ -19,6 +19,8 @@ import {
   FEATURE_ASSISTANT_CONTRACTS_V11,
   FEATURE_ASSISTANT_EVIDENCE_ENRICHMENT_V1,
   FEATURE_ASSISTANT_INVESTIGATION_STATE_V1,
+  FEATURE_ASSISTANT_PROBLEM_CENTRIC_ANSWER_V1,
+  FEATURE_ASSISTANT_PROBLEM_UNIT_CONTINUITY_V1,
   FEATURE_ASSISTANT_STATE_FOLLOWUP_BINDING_V1
 } from "../config";
 import { logJson } from "../utils/log";
@@ -807,22 +809,28 @@ function hasAccountingSignal(text: string): boolean {
   if (/(?:^|[\s,;:])\d{2}(?:\.\d{2})?(?=$|[\s,.;:])/i.test(lower)) {
     return true;
   }
-  return /(проводк|документ|реализац|поступлен|взаиморасчет|сальдо|остатк|счет|ндс|амортиз|рбп|ос|контрагент|поставщик|покупател|оплат|банк|выписк|склад|товар|материал|counterparty|supplier|invoice|posting|ledger|account|anomaly|risk)/i.test(
+  return /(РїСЂРѕРІРѕРґРє|документ|реализац|поступлен|взаиморасчет|сальдо|остатк|счет|РЅРґСЃ|амортиз|СЂР±Рї|РѕСЃ|контрагент|поставщик|покупател|оплат|банк|выписк|склад|товар|материал|проводк|документ|реализац|поступлен|взаиморасчет|сальдо|остатк|счет|счёт|ндс|амортиз|рбп|контрагент|поставщик|покупател|оплат|банк|выписк|склад|товар|материал|закрыти|период|postavshchik|kontragent|schet|schetu|period|counterparty|supplier|invoice|posting|ledger|account|anomaly|risk)/i.test(
     lower
   );
 }
 
 function hasFollowupMarker(text: string): boolean {
   const compact = compactWhitespace(text.toLowerCase());
-  return /^(и|а еще|а ещё|еще|ещё|добав|уточн|продолж|также|plus|also|dobav|utochn|prodolzh)/i.test(compact);
+  return /^(Рё|Р° еще|Р° ещё|еще|ещё|добав|уточн|продолж|также|и|а если|а еще|а ещё|еще|ещё|добав|уточн|продолж|также|plus|also|dobav|utochn|prodolzh)/i.test(
+    compact
+  );
 }
 
 function hasReferentialPointer(text: string): boolean {
-  return /(по этому|по тому|это же|этой|этим|тому|same thing|that one|po etomu|po tomu)/i.test(text.toLowerCase());
+  return /(РїРѕ этому|РїРѕ тому|это же|этой|этим|тому|по этому|по тому|это же|этой|этим|этому|из этого|в этом|тот же|same thing|that one|po etomu|po tomu)/i.test(
+    text.toLowerCase()
+  );
 }
 
 function hasSmallTalkSignal(text: string): boolean {
-  return /(привет|как дела|спасибо|thanks|thank you|hello|hi)\b/i.test(text.toLowerCase());
+  return /(привет|как дела|спасибо|привет|как дела|спасибо|благодарю|thanks|thank you|hello|hi)\b/i.test(
+    text.toLowerCase()
+  );
 }
 
 function countTokens(text: string): number {
@@ -835,6 +843,44 @@ function hasPeriodLiteral(text: string): boolean {
   return /\b(20\d{2}(?:[-/.](?:0[1-9]|1[0-2]))?)\b/.test(text);
 }
 
+function extractNormalizedPeriodLiteral(text: string): string | null {
+  const monthly = text.match(/\b(20\d{2})[-/.](0[1-9]|1[0-2])\b/);
+  if (monthly) {
+    return `${monthly[1]}-${monthly[2]}`;
+  }
+  const yearly = text.match(/\b(20\d{2})\b/);
+  if (yearly) {
+    return yearly[1];
+  }
+  return null;
+}
+
+function hasStrongFollowupAnchors(
+  userMessage: string,
+  state: NonNullable<AssistantSessionState["investigation_state"]>
+): boolean {
+  const explicitPeriod = extractNormalizedPeriodLiteral(userMessage);
+  if (explicitPeriod && state.focus.period && explicitPeriod !== state.focus.period) {
+    const periodLooksLikeFollowupRefinement = hasFollowupMarker(userMessage) || hasReferentialPointer(userMessage);
+    if (!periodLooksLikeFollowupRefinement) {
+      return true;
+    }
+  }
+
+  const explicitAccounts = extractAccountTokens(userMessage);
+  if (explicitAccounts.length > 0) {
+    const knownAccounts = new Set(state.focus.primary_accounts.map((item) => item.trim()));
+    if (knownAccounts.size === 0) {
+      return true;
+    }
+    if (explicitAccounts.some((item) => !knownAccounts.has(item))) {
+      return true;
+    }
+  }
+
+  return false;
+}
+
 function routeFromInvestigationState(state: NonNullable<AssistantSessionState["investigation_state"]>): RouteHint | null {
   const rawDomain = compactWhitespace(state.focus.domain ?? "");
   if (!rawDomain) {
@@ -890,7 +936,17 @@ function buildFollowupStateBinding(input: {
   const referentialPointer = hasReferentialPointer(userMessage);
   const shortPrompt = countTokens(userMessage) <= 10;
   const smallTalkSignal = hasSmallTalkSignal(userMessage);
-  const shouldBind = !smallTalkSignal && (followupMarker || referentialPointer || (!strongSignal && shortPrompt));
+  const problemState = input.investigationState.problem_unit_state;
+  const problemContinuityAvailable =
+    FEATURE_ASSISTANT_PROBLEM_UNIT_CONTINUITY_V1 &&
+    Boolean(problemState) &&
+    ((problemState?.active_problem_units.length ?? 0) > 0 || (problemState?.focus_problem_types.length ?? 0) > 0);
+  const strongNewAnchorDetected = hasStrongFollowupAnchors(userMessage, input.investigationState);
+  const periodRefinementFollowup = hasPeriodLiteral(userMessage) && problemContinuityAvailable;
+  const shouldBind =
+    !smallTalkSignal &&
+    !strongNewAnchorDetected &&
+    (followupMarker || referentialPointer || periodRefinementFollowup || (!strongSignal && shortPrompt));
 
   if (!shouldBind) {
     return {
@@ -903,6 +959,7 @@ function buildFollowupStateBinding(input: {
   const context: NormalizeRequestPayload["context"] = {
     ...(input.payloadContext ?? {})
   };
+  const hasExplicitExpectedRoute = Boolean(input.payloadContext?.expected_route);
   const expectedRouteFromState = !context?.expected_route ? routeFromInvestigationState(input.investigationState) : null;
   const periodHintFromState = !context?.period_hint ? input.investigationState.focus.period : null;
 
@@ -915,6 +972,9 @@ function buildFollowupStateBinding(input: {
 
   const subject = withCappedLength(compactWhitespace(input.investigationState.focus.active_query_subject ?? ""), FOLLOWUP_SUBJECT_MAX);
   const businessContextPatch: string[] = ["followup_state_binding_v1"];
+  let problemContinuityApplied = false;
+  let problemContinuitySkippedReason: string | null = null;
+
   if (input.investigationState.focus.period) {
     businessContextPatch.push("active_period");
   }
@@ -924,6 +984,20 @@ function buildFollowupStateBinding(input: {
   if (input.investigationState.focus.primary_accounts.length > 0) {
     businessContextPatch.push(`focus_accounts:${input.investigationState.focus.primary_accounts.join(",")}`);
   }
+  if (problemContinuityAvailable) {
+    if (hasExplicitExpectedRoute) {
+      problemContinuitySkippedReason = "explicit_expected_route";
+    } else {
+      const focusTypes = (problemState?.focus_problem_types ?? []).slice(0, 3);
+      const activeCount = problemState?.active_problem_units.length ?? 0;
+      businessContextPatch.push("problem_unit_continuity_v1");
+      if (focusTypes.length > 0) {
+        businessContextPatch.push(`problem_focus_types:${focusTypes.join(",")}`);
+      }
+      businessContextPatch.push(`problem_active_count:${activeCount}`);
+      problemContinuityApplied = true;
+    }
+  }
 
   const mergedBusinessContext = mergeBusinessContext(context?.business_context, businessContextPatch);
   if (mergedBusinessContext) {
@@ -940,6 +1014,9 @@ function buildFollowupStateBinding(input: {
     if (periodHintFromState && !hasPeriodLiteral(userMessage)) {
       appendParts.push(`Период фокуса: ${periodHintFromState}`);
     }
+    if (problemContinuityApplied && (problemState?.focus_problem_types.length ?? 0) > 0) {
+      appendParts.push(`Problem focus types: ${(problemState?.focus_problem_types ?? []).slice(0, 3).join(", ")}`);
+    }
     const appendBlock = withCappedLength(compactWhitespace(appendParts.join("; ")), FOLLOWUP_QUESTION_APPEND_MAX);
     normalizedQuestion = `${userMessage}\n${appendBlock}`.trim();
   }
@@ -961,7 +1038,11 @@ function buildFollowupStateBinding(input: {
         period_hint_from_state: Boolean(periodHintFromState),
         expected_route_from_state: Boolean(expectedRouteFromState),
         business_context_from_state: Boolean(mergedBusinessContext),
-        question_augmented: shouldAugmentQuestion
+        question_augmented: shouldAugmentQuestion,
+        problem_continuity_available: problemContinuityAvailable,
+        problem_continuity_applied: problemContinuityApplied,
+        problem_continuity_skipped_reason: problemContinuityApplied ? null : problemContinuitySkippedReason,
+        strong_new_anchor_detected: strongNewAnchorDetected
       }
     }
   };
@@ -1121,7 +1202,8 @@ export class AssistantService {
       requirements: coverageEvaluation.requirements,
       coverageReport: coverageEvaluation.coverage,
       groundingCheck,
-      enableAnswerPolicyV11: FEATURE_ASSISTANT_ANSWER_POLICY_V11
+      enableAnswerPolicyV11: FEATURE_ASSISTANT_ANSWER_POLICY_V11,
+      enableProblemCentricAnswerV1: FEATURE_ASSISTANT_PROBLEM_CENTRIC_ANSWER_V1
     });
 
     const answerStructureV11 = FEATURE_ASSISTANT_CONTRACTS_V11
@@ -1175,6 +1257,14 @@ export class AssistantService {
       answer_grounding_check: groundingCheck,
       dropped_intent_segments: extractDiscardedIntentSegments(normalized.normalized),
       ...(followupBinding.usage ? { followup_state_usage: followupBinding.usage } : {}),
+      problem_centric_answer_applied: composition.problem_centric_answer_applied ?? false,
+      problem_units_used_count: composition.problem_units_used_count ?? 0,
+      problem_answer_mode: composition.problem_answer_mode ?? "stage1_policy_v11",
+      ...(Array.isArray(composition.problem_unit_ids_used) && composition.problem_unit_ids_used.length > 0
+        ? {
+            problem_unit_ids_used: composition.problem_unit_ids_used
+          }
+        : {}),
       answer_structure_v11: answerStructureV11,
       investigation_state_snapshot: investigationStateSnapshot,
       normalized: normalized.normalized
@@ -1234,6 +1324,14 @@ export class AssistantService {
         clarification_target: coverageEvaluation.coverage.clarification_needed_for,
         dropped_intent_segments: extractDiscardedIntentSegments(normalized.normalized),
         ...(followupBinding.usage ? { followup_state_usage: followupBinding.usage } : {}),
+        problem_centric_answer_applied: composition.problem_centric_answer_applied ?? false,
+        problem_units_used_count: composition.problem_units_used_count ?? 0,
+        problem_answer_mode: composition.problem_answer_mode ?? "stage1_policy_v11",
+        ...(Array.isArray(composition.problem_unit_ids_used) && composition.problem_unit_ids_used.length > 0
+          ? {
+              problem_unit_ids_used: composition.problem_unit_ids_used
+            }
+          : {}),
         answer_structure_v11: answerStructureV11,
         investigation_state_snapshot: investigationStateSnapshot,
         fallback_type: composition.fallback_type,

llm_normalizer/backend/src/services/investigationState.ts

@@ -16,9 +16,21 @@ import {
   INVESTIGATION_MAX_UNCERTAINTIES,
   INVESTIGATION_STATE_SCHEMA_VERSION
 } from "../types/stage1Contracts";
+import type {
+  InvestigationProblemUnitState,
+  InvestigationStateWithProblemUnits,
+  ProblemUnit,
+  ProblemUnitEntityBacklink
+} from "../types/stage2ProblemUnits";
+import {
+  INVESTIGATION_MAX_ACTIVE_PROBLEM_UNITS,
+  INVESTIGATION_MAX_FOCUS_PROBLEM_TYPES,
+  INVESTIGATION_MAX_PROBLEM_UNIT_BACKLINKS,
+  INVESTIGATION_MAX_RESOLVED_PROBLEM_UNITS
+} from "../types/stage2ProblemUnits";
 
 interface UpdateInvestigationStateInput {
-  previous: InvestigationState;
+  previous: InvestigationStateWithProblemUnits;
   timestamp: string;
   questionId: string;
   userMessage: string;
@@ -115,9 +127,142 @@ function collectOpenUncertainties(
   return capStrings([...requirementNotes, ...limitationNotes], INVESTIGATION_MAX_UNCERTAINTIES);
 }
 
-export function cloneInvestigationState(state: InvestigationState | null): InvestigationState | null {
-  if (!state) return null;
+function normalizeEntityBacklinks(values: ProblemUnitEntityBacklink[]): ProblemUnitEntityBacklink[] {
+  const result: ProblemUnitEntityBacklink[] = [];
+  const seen = new Set<string>();
+  for (const item of values) {
+    const entity = String(item.entity ?? "").trim();
+    const id = String(item.id ?? "").trim();
+    if (!entity || !id) {
+      continue;
+    }
+    const key = `${entity}::${id}`;
+    if (seen.has(key)) {
+      continue;
+    }
+    seen.add(key);
+    result.push({
+      entity,
+      id
+    });
+  }
+  return result;
+}
+
+function collectProblemUnits(retrievalResults: UnifiedRetrievalResult[]): ProblemUnit[] {
+  return retrievalResults.flatMap((result) => result.problem_units ?? []);
+}
+
+function capProblemUnitState(state: InvestigationProblemUnitState): InvestigationProblemUnitState {
   return {
+    active_problem_units: capStrings(state.active_problem_units, INVESTIGATION_MAX_ACTIVE_PROBLEM_UNITS),
+    resolved_problem_units: capStrings(state.resolved_problem_units, INVESTIGATION_MAX_RESOLVED_PROBLEM_UNITS),
+    problem_unit_backlinks: state.problem_unit_backlinks
+      .map((item) => ({
+        problem_unit_id: String(item.problem_unit_id ?? "").trim(),
+        entity_backlinks: normalizeEntityBacklinks(item.entity_backlinks ?? [])
+      }))
+      .filter((item) => Boolean(item.problem_unit_id) && item.entity_backlinks.length > 0)
+      .slice(0, INVESTIGATION_MAX_PROBLEM_UNIT_BACKLINKS),
+    focus_problem_types: capStrings(
+      state.focus_problem_types.map((item) => String(item)),
+      INVESTIGATION_MAX_FOCUS_PROBLEM_TYPES
+    ) as InvestigationProblemUnitState["focus_problem_types"]
+  };
+}
+
+function updateProblemUnitState(
+  previous: InvestigationStateWithProblemUnits,
+  retrievalResults: UnifiedRetrievalResult[]
+): InvestigationProblemUnitState | undefined {
+  const previousState = previous.problem_unit_state;
+  const currentProblemUnits = collectProblemUnits(retrievalResults);
+  const currentIds = capStrings(
+    currentProblemUnits.map((item) => String(item.problem_unit_id ?? "")),
+    INVESTIGATION_MAX_ACTIVE_PROBLEM_UNITS
+  );
+  const currentTypes = capStrings(
+    currentProblemUnits.map((item) => String(item.problem_unit_type ?? "")),
+    INVESTIGATION_MAX_FOCUS_PROBLEM_TYPES
+  ) as InvestigationProblemUnitState["focus_problem_types"];
+
+  const currentBacklinksRaw = currentProblemUnits
+    .filter((item) => currentIds.includes(item.problem_unit_id))
+    .map((item) => ({
+      problem_unit_id: item.problem_unit_id,
+      entity_backlinks: normalizeEntityBacklinks(item.entity_backlinks ?? [])
+    }))
+    .filter((item) => item.entity_backlinks.length > 0);
+
+  const currentBacklinksById = new Map(
+    currentBacklinksRaw.map((item) => [item.problem_unit_id, item.entity_backlinks] as const)
+  );
+  const previousBacklinksById = new Map(
+    (previousState?.problem_unit_backlinks ?? []).map((item) => [item.problem_unit_id, item.entity_backlinks] as const)
+  );
+
+  const active_problem_units =
+    currentIds.length > 0
+      ? currentIds
+      : capStrings(previousState?.active_problem_units ?? [], INVESTIGATION_MAX_ACTIVE_PROBLEM_UNITS);
+
+  const resolved_problem_units =
+    currentIds.length > 0
+      ? capStrings(
+          [
+            ...(previousState?.active_problem_units ?? []).filter((item) => !currentIds.includes(item)),
+            ...(previousState?.resolved_problem_units ?? [])
+          ],
+          INVESTIGATION_MAX_RESOLVED_PROBLEM_UNITS
+        )
+      : capStrings(previousState?.resolved_problem_units ?? [], INVESTIGATION_MAX_RESOLVED_PROBLEM_UNITS);
+
+  const problem_unit_backlinks = active_problem_units
+    .map((problemUnitId) => {
+      const entity_backlinks = normalizeEntityBacklinks(
+        currentBacklinksById.get(problemUnitId) ?? previousBacklinksById.get(problemUnitId) ?? []
+      );
+      if (entity_backlinks.length === 0) {
+        return null;
+      }
+      return {
+        problem_unit_id: problemUnitId,
+        entity_backlinks
+      };
+    })
+    .filter((item): item is NonNullable<typeof item> => item !== null)
+    .slice(0, INVESTIGATION_MAX_PROBLEM_UNIT_BACKLINKS);
+
+  const focus_problem_types =
+    currentTypes.length > 0
+      ? currentTypes
+      : capStrings(
+          (previousState?.focus_problem_types ?? []).map((item) => String(item)),
+          INVESTIGATION_MAX_FOCUS_PROBLEM_TYPES
+        ) as InvestigationProblemUnitState["focus_problem_types"];
+
+  const nextState = capProblemUnitState({
+    active_problem_units,
+    resolved_problem_units,
+    problem_unit_backlinks,
+    focus_problem_types
+  });
+
+  if (
+    nextState.active_problem_units.length === 0 &&
+    nextState.resolved_problem_units.length === 0 &&
+    nextState.problem_unit_backlinks.length === 0 &&
+    nextState.focus_problem_types.length === 0
+  ) {
+    return undefined;
+  }
+
+  return nextState;
+}
+
+export function cloneInvestigationState(state: InvestigationStateWithProblemUnits | null): InvestigationStateWithProblemUnits | null {
+  if (!state) return null;
+  const cloned: InvestigationStateWithProblemUnits = {
     ...state,
     focus: {
       ...state.focus,
@@ -132,9 +277,24 @@ export function cloneInvestigationState(state: InvestigationState | null): Inves
         }
       : null
   };
+  if (state.problem_unit_state) {
+    cloned.problem_unit_state = capProblemUnitState({
+      active_problem_units: [...state.problem_unit_state.active_problem_units],
+      resolved_problem_units: [...state.problem_unit_state.resolved_problem_units],
+      problem_unit_backlinks: state.problem_unit_state.problem_unit_backlinks.map((item) => ({
+        problem_unit_id: item.problem_unit_id,
+        entity_backlinks: [...item.entity_backlinks]
+      })),
+      focus_problem_types: [...state.problem_unit_state.focus_problem_types]
+    });
+  }
+  return cloned;
 }
 
-export function createEmptyInvestigationState(sessionId: string, timestamp = new Date().toISOString()): InvestigationState {
+export function createEmptyInvestigationState(
+  sessionId: string,
+  timestamp = new Date().toISOString()
+): InvestigationStateWithProblemUnits {
   return {
     schema_version: INVESTIGATION_STATE_SCHEMA_VERSION,
     session_id: sessionId,
@@ -157,7 +317,7 @@ export function createEmptyInvestigationState(sessionId: string, timestamp = new
   };
 }
 
-export function updateInvestigationState(input: UpdateInvestigationStateInput): InvestigationState {
+export function updateInvestigationState(input: UpdateInvestigationStateInput): InvestigationStateWithProblemUnits {
   const previous = input.previous;
   const focusFromMessage = capStrings(detectAccounts(input.userMessage), INVESTIGATION_MAX_PRIMARY_ACCOUNTS);
   const requirementIds = capStrings(
@@ -165,6 +325,7 @@ export function updateInvestigationState(input: UpdateInvestigationStateInput):
     INVESTIGATION_MAX_REQUIREMENT_LINKS
   );
   const mainRequirement = input.requirements[0]?.requirement_text ?? input.userMessage;
+  const problemUnitState = updateProblemUnitState(previous, input.retrievalResults);
 
   return {
     schema_version: INVESTIGATION_STATE_SCHEMA_VERSION,
@@ -194,6 +355,11 @@ export function updateInvestigationState(input: UpdateInvestigationStateInput):
       last_user_message: input.userMessage.slice(0, 240),
       referenced_requirement_ids: requirementIds
     },
-    query_mode_hint: deriveQueryModeHint(input.routeSummary)
+    query_mode_hint: deriveQueryModeHint(input.routeSummary),
+    ...(problemUnitState
+      ? {
+          problem_unit_state: problemUnitState
+        }
+      : {})
   };
 }

llm_normalizer/backend/src/services/problemUnitAssembler.ts

@@ -0,0 +1,592 @@
+import type { EvidenceItem } from "../types/stage1Contracts";
+import type {
+  CandidateEvidenceItem,
+  ProblemConfidence,
+  ProblemScore,
+  ProblemUnit,
+  ProblemUnitEntityBacklink,
+  ProblemUnitSummary,
+  ProblemUnitType
+} from "../types/stage2ProblemUnits";
+import {
+  CANDIDATE_EVIDENCE_SCHEMA_VERSION,
+  PROBLEM_UNIT_SCHEMA_VERSION,
+  PROBLEM_UNIT_SUMMARY_SCHEMA_VERSION
+} from "../types/stage2ProblemUnits";
+
+type RetrievalResultType = "list" | "summary" | "object" | "chain" | "ranking";
+
+interface AssembleProblemUnitsInput {
+  route: string;
+  result_type?: RetrievalResultType;
+  evidence: EvidenceItem[];
+  raw_entities?: Array<Record<string, unknown>>;
+  summary?: Record<string, unknown>;
+  risk_factors?: string[];
+  selection_reason?: string[];
+  business_interpretation?: string[];
+}
+
+interface CandidateCluster {
+  cluster_id: string;
+  candidates: CandidateEvidenceItem[];
+}
+
+interface SeverityResult {
+  severity: ProblemScore;
+  confidence: ProblemConfidence;
+}
+
+interface CandidateBuildContext {
+  route: string;
+  result_type?: RetrievalResultType;
+  raw_entities: Array<Record<string, unknown>>;
+  summary_relation_patterns: string[];
+  summary_anomaly_patterns: string[];
+  risk_factors: string[];
+}
+
+function toObject(value: unknown): Record<string, unknown> | null {
+  if (!value || typeof value !== "object" || Array.isArray(value)) {
+    return null;
+  }
+  return value as Record<string, unknown>;
+}
+
+function uniqueStrings(values: string[]): string[] {
+  return Array.from(new Set(values.map((item) => item.trim()).filter(Boolean)));
+}
+
+function clampUnitScore(value: number): number {
+  if (!Number.isFinite(value)) {
+    return 0;
+  }
+  if (value <= 0) return 0;
+  if (value >= 1) return 1;
+  return Number(value.toFixed(2));
+}
+
+function gradeForScore(score: number): "low" | "medium" | "high" {
+  if (score >= 0.7) return "high";
+  if (score >= 0.4) return "medium";
+  return "low";
+}
+
+function confidenceGradeForScore(score: number): "low" | "medium" | "high" {
+  if (score >= 0.75) return "high";
+  if (score >= 0.45) return "medium";
+  return "low";
+}
+
+function confidenceToScore(value: CandidateEvidenceItem["confidence_hint"]): number {
+  if (value === "high") return 1;
+  if (value === "medium") return 0.6;
+  return 0.3;
+}
+
+function valueFromPayload(item: EvidenceItem, key: string): string | null {
+  const value = item.payload[key];
+  if (typeof value !== "string") {
+    return null;
+  }
+  const trimmed = value.trim();
+  return trimmed.length > 0 ? trimmed : null;
+}
+
+function stringArrayFromUnknown(value: unknown): string[] {
+  if (!Array.isArray(value)) {
+    return [];
+  }
+  return uniqueStrings(value.map((entry) => String(entry)));
+}
+
+function stringArrayFromPayload(item: EvidenceItem, key: string): string[] {
+  return stringArrayFromUnknown(item.payload[key]);
+}
+
+function extractSemanticProfile(summary: Record<string, unknown>): {
+  relation_patterns: string[];
+  anomaly_patterns: string[];
+} {
+  const semanticProfile = toObject(summary.semantic_profile);
+  return {
+    relation_patterns: stringArrayFromUnknown(semanticProfile?.relation_patterns),
+    anomaly_patterns: stringArrayFromUnknown(semanticProfile?.anomaly_patterns)
+  };
+}
+
+function resolveEntityOverlay(item: EvidenceItem, rawEntities: Array<Record<string, unknown>>): Record<string, unknown> | null {
+  const sourceId = String(item.pointer.source.id ?? "").toLowerCase();
+  if (!sourceId) {
+    return null;
+  }
+  for (const entity of rawEntities) {
+    const candidates = [
+      String(entity.source_id ?? ""),
+      String(entity.entity_id ?? ""),
+      String(entity.id ?? "")
+    ]
+      .map((entry) => entry.toLowerCase())
+      .filter(Boolean);
+    if (candidates.includes(sourceId)) {
+      return entity;
+    }
+  }
+  return null;
+}
+
+function detectRelationPatternHits(
+  item: EvidenceItem,
+  overlay: Record<string, unknown> | null,
+  context: CandidateBuildContext
+): string[] {
+  const hits: string[] = [];
+  const failedEdge = valueFromPayload(item, "failed_expected_edge");
+  if (failedEdge) {
+    hits.push(`failed_edge:${failedEdge}`);
+  }
+  const expectedStep = valueFromPayload(item, "expected_next_step");
+  if (expectedStep) {
+    hits.push(`expected_step:${expectedStep}`);
+  }
+  const relationPattern = valueFromPayload(item, "relation_pattern");
+  if (relationPattern) {
+    hits.push(relationPattern);
+  }
+
+  hits.push(...stringArrayFromPayload(item, "relation_patterns"));
+  hits.push(...stringArrayFromPayload(item, "relation_pattern_hits"));
+
+  if (overlay) {
+    hits.push(...stringArrayFromUnknown(overlay.relation_pattern_hits));
+    hits.push(...stringArrayFromUnknown(overlay.relation_types));
+  }
+
+  hits.push(...context.summary_relation_patterns);
+  if (context.route === "hybrid_store_plus_live" && hits.length === 0) {
+    hits.push("chain_scope");
+  }
+
+  return uniqueStrings(hits);
+}
+
+function detectAnomalyPatterns(
+  item: EvidenceItem,
+  overlay: Record<string, unknown> | null,
+  context: CandidateBuildContext
+): string[] {
+  const patterns: string[] = [];
+  patterns.push(...stringArrayFromPayload(item, "anomaly_patterns"));
+  patterns.push(...stringArrayFromPayload(item, "risk_factors"));
+  patterns.push(...stringArrayFromPayload(item, "lifecycle_gaps"));
+  patterns.push(...stringArrayFromPayload(item, "lifecycle_markers"));
+
+  const explicit = valueFromPayload(item, "anomaly_pattern");
+  if (explicit) {
+    patterns.push(explicit);
+  }
+
+  if (overlay) {
+    patterns.push(...stringArrayFromUnknown(overlay.risk_factors));
+    patterns.push(...stringArrayFromUnknown(overlay.lifecycle_gaps));
+    patterns.push(...stringArrayFromUnknown(overlay.anomaly_patterns));
+  }
+
+  patterns.push(...context.summary_anomaly_patterns);
+  patterns.push(...context.risk_factors);
+
+  if (item.evidence_kind === "anomaly_signal") {
+    patterns.push("anomaly_signal");
+  }
+  if (item.limitation?.reason_code === "missing_mechanism") {
+    patterns.push("missing_mechanism");
+  }
+  if (item.limitation?.reason_code === "insufficient_detail") {
+    patterns.push("insufficient_detail");
+  }
+
+  const defectClass = valueFromPayload(item, "business_defect_class");
+  if (defectClass) {
+    patterns.push(defectClass);
+  }
+
+  if (context.route === "store_feature_risk") {
+    patterns.push("risk_route");
+  }
+
+  return uniqueStrings(patterns);
+}
+
+function buildEntityBacklinks(item: EvidenceItem, overlay: Record<string, unknown> | null): ProblemUnitEntityBacklink[] {
+  const backlinks: ProblemUnitEntityBacklink[] = [];
+  const sourceEntity = String(item.pointer.source.entity ?? "").trim();
+  const sourceId = String(item.pointer.source.id ?? "").trim();
+  if (sourceEntity && sourceId) {
+    backlinks.push({
+      entity: sourceEntity,
+      id: sourceId
+    });
+  }
+
+  const payloadEntity = valueFromPayload(item, "source_entity");
+  const payloadId = valueFromPayload(item, "source_id");
+  if (payloadEntity && payloadId) {
+    backlinks.push({
+      entity: payloadEntity,
+      id: payloadId
+    });
+  }
+
+  const overlayEntity = overlay ? String(overlay.source_entity ?? "").trim() : "";
+  const overlayId = overlay ? String(overlay.source_id ?? "").trim() : "";
+  if (overlayEntity && overlayId) {
+    backlinks.push({
+      entity: overlayEntity,
+      id: overlayId
+    });
+  }
+
+  return Array.from(
+    new Map(backlinks.map((entry) => [`${entry.entity.toLowerCase()}|${entry.id.toLowerCase()}`, entry])).values()
+  );
+}
+
+function inferExpectedState(item: EvidenceItem): string | undefined {
+  return valueFromPayload(item, "expected_state") ?? valueFromPayload(item, "expected_next_step") ?? undefined;
+}
+
+function inferActualState(item: EvidenceItem): string | undefined {
+  return valueFromPayload(item, "actual_state") ?? valueFromPayload(item, "mechanism_of_failure") ?? undefined;
+}
+
+export function buildCandidateEvidence(
+  items: EvidenceItem[],
+  route: string,
+  contextInput?: Partial<CandidateBuildContext>
+): CandidateEvidenceItem[] {
+  const context: CandidateBuildContext = {
+    route,
+    result_type: contextInput?.result_type,
+    raw_entities: contextInput?.raw_entities ?? [],
+    summary_relation_patterns: contextInput?.summary_relation_patterns ?? [],
+    summary_anomaly_patterns: contextInput?.summary_anomaly_patterns ?? [],
+    risk_factors: contextInput?.risk_factors ?? []
+  };
+
+  return items.map((item, index) => {
+    const overlay = resolveEntityOverlay(item, context.raw_entities);
+    return {
+      schema_version: CANDIDATE_EVIDENCE_SCHEMA_VERSION,
+      candidate_id: `cand-${item.evidence_id || `${route}-${index + 1}`}`,
+      route,
+      source_ref: item.source_ref,
+      expected_state: inferExpectedState(item),
+      actual_state: inferActualState(item),
+      relation_pattern_hits: detectRelationPatternHits(item, overlay, context),
+      anomaly_patterns: detectAnomalyPatterns(item, overlay, context),
+      entity_backlinks: buildEntityBacklinks(item, overlay),
+      confidence_hint: item.confidence
+    };
+  });
+}
+
+function clusterSignature(candidate: CandidateEvidenceItem): string {
+  const relation = candidate.relation_pattern_hits[0] ?? "none";
+  const anomaly = candidate.anomaly_patterns[0] ?? "none";
+  return [candidate.route, candidate.source_ref.canonical_ref, relation, anomaly].join("|");
+}
+
+export function clusterCandidateEvidence(candidates: CandidateEvidenceItem[]): CandidateCluster[] {
+  const byCluster = new Map<string, CandidateEvidenceItem[]>();
+  for (const candidate of candidates) {
+    const signature = clusterSignature(candidate);
+    const current = byCluster.get(signature) ?? [];
+    current.push(candidate);
+    byCluster.set(signature, current);
+  }
+  return Array.from(byCluster.entries()).map(([cluster_id, clusterCandidates]) => ({
+    cluster_id,
+    candidates: clusterCandidates
+  }));
+}
+
+function hasAny(value: string, pattern: RegExp): boolean {
+  return pattern.test(value);
+}
+
+export function detectProblemUnitType(cluster: CandidateCluster): ProblemUnitType {
+  const relationText = cluster.candidates.flatMap((item) => item.relation_pattern_hits).join(" ").toLowerCase();
+  const anomalyText = cluster.candidates.flatMap((item) => item.anomaly_patterns).join(" ").toLowerCase();
+  const sourceText = cluster.candidates
+    .map((item) => `${item.source_ref.entity} ${item.source_ref.id}`)
+    .join(" ")
+    .toLowerCase();
+  const routeText = cluster.candidates.map((item) => item.route).join(" ").toLowerCase();
+
+  if (hasAny(`${anomalyText} ${sourceText}`, /cross[_\s-]?branch|vat|nds|tax|ндс/)) {
+    return "cross_branch_inconsistency_cluster";
+  }
+  if (hasAny(`${anomalyText} ${relationText}`, /period|close[_\s-]?risk|reporting|закрыт|period_close/)) {
+    return "period_risk_cluster";
+  }
+  if (hasAny(anomalyText, /settlement|tail|unresolved|хвост|незакрыт/)) {
+    return "unresolved_settlement_cluster";
+  }
+  if (hasAny(anomalyText, /lifecycle|deferred|broken_lifecycle|списани|амортиз|рбп/)) {
+    return "lifecycle_anomaly_node";
+  }
+  if (
+    hasAny(relationText, /failed_edge|statement_to_document|payment_to_settlement|chain|broken|цепоч|разрыв/)
+    || (routeText.includes("hybrid_store_plus_live") && relationText.length > 0)
+  ) {
+    return "broken_chain_segment";
+  }
+  return "document_conflict";
+}
+
+export function scoreProblemSeverity(cluster: CandidateCluster): SeverityResult {
+  const candidates = cluster.candidates;
+  const averageConfidence =
+    candidates.length > 0
+      ? candidates.reduce((acc, item) => acc + confidenceToScore(item.confidence_hint), 0) / candidates.length
+      : 0;
+  const hasEdgeBreak = candidates.some((item) =>
+    item.relation_pattern_hits.some((pattern) => /failed_edge|chain|statement_to_document|payment_to_settlement/i.test(pattern))
+  );
+  const hasAnomaly = candidates.some((item) => item.anomaly_patterns.length > 0);
+  const hasPeriodRisk = candidates.some((item) => item.anomaly_patterns.some((pattern) => /period|close|reporting|закрыт/i.test(pattern)));
+  const candidateBoost = Math.min(candidates.length, 5) * 0.08;
+
+  let severityScore = 0.25 + candidateBoost;
+  if (hasEdgeBreak) severityScore += 0.2;
+  if (hasAnomaly) severityScore += 0.15;
+  if (hasPeriodRisk) severityScore += 0.1;
+  const normalizedSeverity = clampUnitScore(severityScore);
+
+  let confidenceScore = averageConfidence;
+  if (hasEdgeBreak) confidenceScore += 0.05;
+  const normalizedConfidence = clampUnitScore(confidenceScore);
+
+  return {
+    severity: {
+      score: normalizedSeverity,
+      grade: gradeForScore(normalizedSeverity)
+    },
+    confidence: {
+      score: normalizedConfidence,
+      grade: confidenceGradeForScore(normalizedConfidence)
+    }
+  };
+}
+
+function unitTitle(type: ProblemUnitType): string {
+  if (type === "document_conflict") return "Document conflict detected";
+  if (type === "broken_chain_segment") return "Broken chain segment detected";
+  if (type === "lifecycle_anomaly_node") return "Lifecycle anomaly node detected";
+  if (type === "unresolved_settlement_cluster") return "Unresolved settlement cluster detected";
+  if (type === "period_risk_cluster") return "Period risk cluster detected";
+  return "Cross-branch inconsistency cluster detected";
+}
+
+function mechanismSummary(cluster: CandidateCluster, type: ProblemUnitType): string {
+  const relationHints = uniqueStrings(cluster.candidates.flatMap((item) => item.relation_pattern_hits));
+  const anomalyHints = uniqueStrings(cluster.candidates.flatMap((item) => item.anomaly_patterns));
+  const primaryRelation = relationHints[0];
+  const primaryAnomaly = anomalyHints[0];
+
+  if (primaryRelation) {
+    return `Mechanism candidate: ${primaryRelation}.`;
+  }
+  if (primaryAnomaly) {
+    return `Mechanism inferred from anomaly pattern: ${primaryAnomaly}.`;
+  }
+  return `Mechanism is currently inferred at baseline level for ${type}.`;
+}
+
+function businessDefectClass(cluster: CandidateCluster, type: ProblemUnitType): string {
+  const patterns = uniqueStrings([
+    ...cluster.candidates.flatMap((item) => item.relation_pattern_hits),
+    ...cluster.candidates.flatMap((item) => item.anomaly_patterns)
+  ]);
+  return patterns[0] ?? type;
+}
+
+function collectAffectedByEntity(backlinks: ProblemUnitEntityBacklink[], pattern: RegExp): string[] {
+  return uniqueStrings(
+    backlinks.filter((entry) => pattern.test(entry.entity)).map((entry) => `${entry.entity}:${entry.id}`)
+  );
+}
+
+function parseFailedExpectedEdge(cluster: CandidateCluster): string | undefined {
+  const withEdge = cluster.candidates.flatMap((item) => item.relation_pattern_hits).find((pattern) => pattern.startsWith("failed_edge:"));
+  if (!withEdge) {
+    return undefined;
+  }
+  return withEdge.replace(/^failed_edge:/, "").trim() || undefined;
+}
+
+function mergeBacklinks(candidates: CandidateEvidenceItem[]): ProblemUnitEntityBacklink[] {
+  return Array.from(
+    new Map(
+      candidates
+        .flatMap((item) => item.entity_backlinks)
+        .map((entry) => [`${entry.entity.toLowerCase()}|${entry.id.toLowerCase()}`, entry] as const)
+    ).values()
+  );
+}
+
+export function buildProblemUnit(cluster: CandidateCluster, index: number): ProblemUnit {
+  const type = detectProblemUnitType(cluster);
+  const scored = scoreProblemSeverity(cluster);
+  const backlinks = mergeBacklinks(cluster.candidates);
+  const expectedState = cluster.candidates.find((item) => typeof item.expected_state === "string")?.expected_state;
+  const actualState = cluster.candidates.find((item) => typeof item.actual_state === "string")?.actual_state;
+  const failedExpectedEdge = parseFailedExpectedEdge(cluster);
+  const periodSensitive = cluster.candidates.some((item) =>
+    item.anomaly_patterns.some((pattern) => /period|close|reporting|закрыт/i.test(pattern))
+  );
+  const hasLowConfidence = cluster.candidates.some((item) => item.confidence_hint === "low");
+
+  return {
+    schema_version: PROBLEM_UNIT_SCHEMA_VERSION,
+    problem_unit_id: `pu-${type}-${index + 1}`,
+    problem_unit_type: type,
+    title: unitTitle(type),
+    mechanism_summary: mechanismSummary(cluster, type),
+    business_defect_class: businessDefectClass(cluster, type),
+    severity: scored.severity,
+    confidence: scored.confidence,
+    affected_entities: uniqueStrings(backlinks.map((entry) => `${entry.entity}:${entry.id}`)),
+    affected_documents: collectAffectedByEntity(backlinks, /doc|document|invoice|плат|реал|поступ/i),
+    affected_postings: collectAffectedByEntity(backlinks, /posting|journal|провод/i),
+    affected_accounts: collectAffectedByEntity(backlinks, /account|счет|сч/i),
+    affected_counterparties: collectAffectedByEntity(backlinks, /counterparty|supplier|buyer|контраг|постав|покуп/i),
+    affected_contracts: collectAffectedByEntity(backlinks, /contract|договор/i),
+    ...(expectedState ? { expected_state: expectedState } : {}),
+    ...(actualState ? { actual_state: actualState } : {}),
+    ...(failedExpectedEdge ? { failed_expected_edge: failedExpectedEdge } : {}),
+    ...(periodSensitive
+      ? {
+          period_impact: {
+            is_period_sensitive: true,
+            impact_class: "close_risk" as const
+          }
+        }
+      : {}),
+    evidence_pack: uniqueStrings(cluster.candidates.map((item) => item.candidate_id)),
+    entity_backlinks: backlinks,
+    snapshot_limitations: uniqueStrings(
+      hasLowConfidence ? ["low_confidence_candidates_present"] : []
+    )
+  };
+}
+
+function collapseSignature(unit: ProblemUnit): string {
+  const backlink = unit.entity_backlinks[0] ? `${unit.entity_backlinks[0].entity}|${unit.entity_backlinks[0].id}` : "none";
+  return [unit.problem_unit_type, unit.business_defect_class, unit.failed_expected_edge ?? "none", backlink].join("|");
+}
+
+export function collapseDuplicates(units: ProblemUnit[]): {
+  problem_units: ProblemUnit[];
+  duplicate_collapses: number;
+} {
+  const bySignature = new Map<string, ProblemUnit>();
+  let duplicateCollapses = 0;
+
+  for (const unit of units) {
+    const signature = collapseSignature(unit);
+    const existing = bySignature.get(signature);
+    if (!existing) {
+      bySignature.set(signature, unit);
+      continue;
+    }
+
+    duplicateCollapses += 1;
+    bySignature.set(signature, {
+      ...existing,
+      evidence_pack: uniqueStrings([...existing.evidence_pack, ...unit.evidence_pack]),
+      entity_backlinks: Array.from(
+        new Map(
+          [...existing.entity_backlinks, ...unit.entity_backlinks].map((entry) => [
+            `${entry.entity.toLowerCase()}|${entry.id.toLowerCase()}`,
+            entry
+          ])
+        ).values()
+      ),
+      affected_entities: uniqueStrings([...existing.affected_entities, ...unit.affected_entities]),
+      affected_documents: uniqueStrings([...existing.affected_documents, ...unit.affected_documents]),
+      affected_postings: uniqueStrings([...existing.affected_postings, ...unit.affected_postings]),
+      affected_accounts: uniqueStrings([...existing.affected_accounts, ...unit.affected_accounts]),
+      affected_counterparties: uniqueStrings([...existing.affected_counterparties, ...unit.affected_counterparties]),
+      affected_contracts: uniqueStrings([...existing.affected_contracts, ...unit.affected_contracts]),
+      snapshot_limitations: uniqueStrings([...existing.snapshot_limitations, ...unit.snapshot_limitations]),
+      severity: unit.severity.score > existing.severity.score ? unit.severity : existing.severity,
+      confidence: unit.confidence.score > existing.confidence.score ? unit.confidence : existing.confidence
+    });
+  }
+
+  return {
+    problem_units: Array.from(bySignature.values()),
+    duplicate_collapses: duplicateCollapses
+  };
+}
+
+function buildSummary(units: ProblemUnit[], duplicateCollapses: number): ProblemUnitSummary {
+  const unitTypes = uniqueStrings(units.map((item) => item.problem_unit_type)) as ProblemUnitType[];
+  const typeDistribution: Partial<Record<ProblemUnitType, number>> = {};
+  const severityDistribution: Record<"low" | "medium" | "high", number> = {
+    low: 0,
+    medium: 0,
+    high: 0
+  };
+  const confidenceDistribution: Record<"low" | "medium" | "high", number> = {
+    low: 0,
+    medium: 0,
+    high: 0
+  };
+
+  for (const unit of units) {
+    typeDistribution[unit.problem_unit_type] = (typeDistribution[unit.problem_unit_type] ?? 0) + 1;
+    severityDistribution[unit.severity.grade] += 1;
+    confidenceDistribution[unit.confidence.grade] += 1;
+  }
+
+  return {
+    schema_version: PROBLEM_UNIT_SUMMARY_SCHEMA_VERSION,
+    units_total: units.length,
+    duplicate_collapses: duplicateCollapses,
+    unit_types: unitTypes,
+    type_distribution: typeDistribution,
+    severity_distribution: severityDistribution,
+    confidence_distribution: confidenceDistribution,
+    primary_unit_type: units[0]?.problem_unit_type ?? null
+  };
+}
+
+export function assembleProblemUnits(input: AssembleProblemUnitsInput): {
+  candidate_evidence: CandidateEvidenceItem[];
+  problem_units: ProblemUnit[];
+  problem_unit_summary: ProblemUnitSummary;
+} {
+  const summary = input.summary ?? {};
+  const semanticProfile = extractSemanticProfile(summary);
+
+  const candidates = buildCandidateEvidence(input.evidence, input.route, {
+    route: input.route,
+    result_type: input.result_type,
+    raw_entities: input.raw_entities ?? [],
+    summary_relation_patterns: semanticProfile.relation_patterns,
+    summary_anomaly_patterns: semanticProfile.anomaly_patterns,
+    risk_factors: uniqueStrings(input.risk_factors ?? [])
+  });
+  const clusters = clusterCandidateEvidence(candidates);
+  const units = clusters.map((cluster, index) => buildProblemUnit(cluster, index));
+  const collapsed = collapseDuplicates(units);
+
+  return {
+    candidate_evidence: candidates,
+    problem_units: collapsed.problem_units,
+    problem_unit_summary: buildSummary(collapsed.problem_units, collapsed.duplicate_collapses)
+  };
+}
+

llm_normalizer/backend/src/services/retrievalResultNormalizer.ts

@@ -4,7 +4,7 @@ import type {
   RetrievalResultType,
   UnifiedRetrievalResult
 } from "../types/assistant";
-import { FEATURE_ASSISTANT_EVIDENCE_ENRICHMENT_V1 } from "../config";
+import { FEATURE_ASSISTANT_EVIDENCE_ENRICHMENT_V1, FEATURE_ASSISTANT_PROBLEM_UNITS_V1 } from "../config";
 import { EVIDENCE_SOURCE_REF_SCHEMA_VERSION } from "../types/stage1Contracts";
 import type {
   EvidenceConfidence,
@@ -14,6 +14,7 @@ import type {
   EvidencePointer,
   EvidenceSourceRef
 } from "../types/stage1Contracts";
+import { assembleProblemUnits } from "./problemUnitAssembler";
 
 interface RawRetrievalResult {
   status?: string;
@@ -94,6 +95,29 @@ function normalizeStringArray(value: unknown): string[] {
   return value.map((item) => String(item));
 }
 
+function mergeSummaryWithProblemUnitMeta(
+  summary: Record<string, unknown>,
+  input: {
+    candidateEvidenceCount: number;
+    problemUnitsCount: number;
+    unitTypes: string[];
+    duplicateCollapses: number;
+    severityDistribution: Record<string, number>;
+    confidenceDistribution: Record<string, number>;
+  }
+): Record<string, unknown> {
+  return {
+    ...summary,
+    problem_units_enabled: true,
+    candidate_evidence_count: input.candidateEvidenceCount,
+    problem_units_count: input.problemUnitsCount,
+    problem_unit_types: input.unitTypes,
+    problem_unit_duplicate_collapses: input.duplicateCollapses,
+    problem_unit_severity_distribution: input.severityDistribution,
+    problem_unit_confidence_distribution: input.confidenceDistribution
+  };
+}
+
 function normalizeConfidence(value: unknown): RetrievalConfidence {
   if (value === "high" || value === "medium" || value === "low") {
     return value;
@@ -459,15 +483,19 @@ export function normalizeRetrievalResult(
   route: string,
   raw: RawRetrievalResult
 ): UnifiedRetrievalResult {
-  return {
+  const items = normalizeObjectArray(raw.items);
+  const summary = normalizeSummary(raw.summary);
+  const evidence = normalizeEvidenceItems(fragmentId, requirementIds, route, raw.evidence);
+
+  const baseResult: UnifiedRetrievalResult = {
     fragment_id: fragmentId,
     requirement_ids: requirementIds,
     route,
     status: normalizeStatus(raw.status),
     result_type: normalizeResultType(raw.result_type),
-    items: normalizeObjectArray(raw.items),
-    summary: normalizeSummary(raw.summary),
-    evidence: normalizeEvidenceItems(fragmentId, requirementIds, route, raw.evidence),
+    items,
+    summary,
+    evidence,
     why_included: normalizeStringArray(raw.why_included),
     selection_reason: normalizeStringArray(raw.selection_reason),
     risk_factors: normalizeStringArray(raw.risk_factors),
@@ -476,4 +504,37 @@ export function normalizeRetrievalResult(
     limitations: normalizeStringArray(raw.limitations),
     errors: normalizeErrors(raw.errors)
   };
+
+  if (!FEATURE_ASSISTANT_PROBLEM_UNITS_V1) {
+    return baseResult;
+  }
+
+  const assembled = assembleProblemUnits({
+    route,
+    result_type: baseResult.result_type,
+    evidence,
+    raw_entities: items,
+    summary,
+    risk_factors: baseResult.risk_factors,
+    selection_reason: baseResult.selection_reason,
+    business_interpretation: baseResult.business_interpretation
+  });
+
+  const enrichedSummary = mergeSummaryWithProblemUnitMeta(summary, {
+    candidateEvidenceCount: assembled.candidate_evidence.length,
+    problemUnitsCount: assembled.problem_units.length,
+    unitTypes: assembled.problem_unit_summary.unit_types,
+    duplicateCollapses: assembled.problem_unit_summary.duplicate_collapses,
+    severityDistribution: assembled.problem_unit_summary.severity_distribution,
+    confidenceDistribution: assembled.problem_unit_summary.confidence_distribution
+  });
+
+  return {
+    ...baseResult,
+    summary: enrichedSummary,
+    raw_entities: items,
+    candidate_evidence: assembled.candidate_evidence,
+    problem_units: assembled.problem_units,
+    problem_unit_summary: assembled.problem_unit_summary
+  };
 }

llm_normalizer/backend/src/types/assistant.ts

@@ -1,5 +1,11 @@
 import type { NormalizeRequestPayload, NormalizeResponsePayload, RouteHintSummary } from "./normalizer";
-import type { AnswerStructureV11, EvidenceItem, InvestigationState } from "./stage1Contracts";
+import type { AnswerStructureV11, EvidenceItem } from "./stage1Contracts";
+import type {
+  CandidateEvidenceItem,
+  InvestigationStateWithProblemUnits,
+  ProblemUnit,
+  ProblemUnitSummary
+} from "./stage2ProblemUnits";
 
 export type AssistantFallbackType = "none" | "out_of_scope" | "clarification" | "partial" | "unknown";
 export type AssistantReplyType =
@@ -15,6 +21,7 @@ export type AssistantReplyType =
 export type RetrievalResultStatus = "ok" | "empty" | "partial" | "error";
 export type RetrievalResultType = "list" | "summary" | "object" | "chain" | "ranking";
 export type RetrievalConfidence = "high" | "medium" | "low";
+export type AssistantProblemAnswerMode = "stage1_policy_v11" | "stage2_problem_centric_v1";
 
 export interface AssistantRequirement {
   requirement_id: string;
@@ -52,6 +59,10 @@ export interface FollowupStateUsageDebug {
     expected_route_from_state: boolean;
     business_context_from_state: boolean;
     question_augmented: boolean;
+    problem_continuity_available?: boolean;
+    problem_continuity_applied?: boolean;
+    problem_continuity_skipped_reason?: string | null;
+    strong_new_anchor_detected?: boolean;
   };
 }
 
@@ -81,6 +92,10 @@ export interface UnifiedRetrievalResult {
   status: RetrievalResultStatus;
   result_type: RetrievalResultType;
   items: Array<Record<string, unknown>>;
+  raw_entities?: Array<Record<string, unknown>>;
+  candidate_evidence?: CandidateEvidenceItem[];
+  problem_units?: ProblemUnit[];
+  problem_unit_summary?: ProblemUnitSummary | null;
   summary: Record<string, unknown>;
   evidence: EvidenceItem[];
   why_included: string[];
@@ -113,8 +128,12 @@ export interface AssistantDebugPayload {
   answer_grounding_check: AnswerGroundingCheck;
   dropped_intent_segments: string[];
   followup_state_usage?: FollowupStateUsageDebug;
+  problem_centric_answer_applied?: boolean;
+  problem_units_used_count?: number;
+  problem_answer_mode?: AssistantProblemAnswerMode;
+  problem_unit_ids_used?: string[];
   answer_structure_v11: AnswerStructureV11 | null;
-  investigation_state_snapshot: InvestigationState | null;
+  investigation_state_snapshot: InvestigationStateWithProblemUnits | null;
   normalized: NormalizeResponsePayload["normalized"];
 }
 
@@ -133,7 +152,7 @@ export interface AssistantSessionState {
   session_id: string;
   updated_at: string;
   items: AssistantConversationItem[];
-  investigation_state: InvestigationState | null;
+  investigation_state: InvestigationStateWithProblemUnits | null;
 }
 
 export interface AssistantMessageResponsePayload {

llm_normalizer/backend/src/types/assistantEval.ts

@@ -1,6 +1,7 @@
 import type { AssistantEvalBroadnessLevel, AssistantEvalQuestionType } from "./stage1Contracts";
+import type { ProblemUnitType } from "./stage2ProblemUnits";
 
-export type EvalTarget = "normalizer" | "assistant_stage1";
+export type EvalTarget = "normalizer" | "assistant_stage1" | "assistant_stage2";
 
 export interface AssistantStage1SuiteCaseTurn {
   user_message: string;
@@ -29,3 +30,21 @@ export interface AssistantStage1SuiteFile {
   case_ids: string[];
   cases: AssistantStage1SuiteCase[];
 }
+
+export interface AssistantStage2ExpectedHints extends AssistantStage1ExpectedHints {
+  expected_problem_first?: boolean;
+  expected_problem_unit_types?: ProblemUnitType[];
+}
+
+export interface AssistantStage2SuiteCase extends Omit<AssistantStage1SuiteCase, "expected_hints"> {
+  expected_hints?: AssistantStage2ExpectedHints;
+}
+
+export interface AssistantStage2SuiteFile {
+  suite_id: string;
+  suite_version: string;
+  schema_version?: string;
+  scenario_count: number;
+  case_ids: string[];
+  cases: AssistantStage2SuiteCase[];
+}

llm_normalizer/backend/src/types/stage2EvalContracts.ts

@@ -0,0 +1,81 @@
+import type { AssistantEvalBroadnessLevel, AssistantEvalQuestionType } from "./stage1Contracts";
+import type { ProblemUnitType } from "./stage2ProblemUnits";
+
+export const ASSISTANT_STAGE2_EVAL_RECORD_SCHEMA_VERSION = "assistant_stage2_eval_record_v0_1" as const;
+
+export interface AssistantStage2MetricVector {
+  problem_unit_precision: number | null;
+  problem_unit_recall_proxy: number | null;
+  duplicate_collapse_rate: number | null;
+  mechanism_coherence_score: number | null;
+  problem_clarity_score: number | null;
+  problem_first_answer_rate: number | null;
+  entity_leakage_rate: number | null;
+}
+
+export type AssistantStage2MetricName = keyof AssistantStage2MetricVector;
+export type AssistantStage2RubricScore = 0 | 3 | 5;
+
+export interface AssistantStage2RubricBand {
+  score: AssistantStage2RubricScore;
+  label: string;
+  description: string;
+}
+
+export interface AssistantStage2EvalRecord {
+  schema_version: typeof ASSISTANT_STAGE2_EVAL_RECORD_SCHEMA_VERSION;
+  created_at: string;
+  case_id: string;
+  scenario_tag: string;
+  session_id: string | null;
+  trace_id: string | null;
+  question_type: AssistantEvalQuestionType;
+  broadness_level: AssistantEvalBroadnessLevel;
+  expected_problem_unit_types: ProblemUnitType[];
+  expected_problem_first: boolean;
+  problem_units_detected: number;
+  candidate_evidence_detected: number;
+  duplicate_collapses_detected: number;
+  metric_subscores: AssistantStage2MetricVector;
+  raw_signals: Record<string, unknown>;
+  limitations: string[];
+  notes: string[];
+}
+
+export const ASSISTANT_STAGE2_SCORING_RUBRIC_V01: Record<AssistantStage2MetricName, AssistantStage2RubricBand[]> = {
+  problem_unit_precision: [
+    { score: 0, label: "Weak", description: "Problem unit typing is often mismatched against expected case profile." },
+    { score: 3, label: "Mixed", description: "Problem unit typing is partially aligned with expected case profile." },
+    { score: 5, label: "Strong", description: "Problem unit typing is consistently aligned with expected case profile." }
+  ],
+  problem_unit_recall_proxy: [
+    { score: 0, label: "Weak", description: "Expected problem categories are frequently missing from output." },
+    { score: 3, label: "Mixed", description: "Some expected problem categories are captured." },
+    { score: 5, label: "Strong", description: "Most expected problem categories are captured." }
+  ],
+  duplicate_collapse_rate: [
+    { score: 0, label: "Weak", description: "Duplicate collapse is rarely observed when candidate evidence is present." },
+    { score: 3, label: "Mixed", description: "Duplicate collapse works on part of candidate evidence." },
+    { score: 5, label: "Strong", description: "Duplicate collapse consistently reduces noisy candidate evidence." }
+  ],
+  mechanism_coherence_score: [
+    { score: 0, label: "Weak", description: "Mechanism narrative is missing or disconnected from problem units." },
+    { score: 3, label: "Mixed", description: "Mechanism narrative is partially connected to problem units." },
+    { score: 5, label: "Strong", description: "Mechanism narrative is explicit and coherent with problem units." }
+  ],
+  problem_clarity_score: [
+    { score: 0, label: "Weak", description: "Answer framing remains generic and unclear for accountant workflow." },
+    { score: 3, label: "Mixed", description: "Problem framing is partially explicit, but still uneven." },
+    { score: 5, label: "Strong", description: "Problem framing is explicit, scoped, and actionable." }
+  ],
+  problem_first_answer_rate: [
+    { score: 0, label: "Weak", description: "Problem-first rendering is rarely applied on applicable cases." },
+    { score: 3, label: "Mixed", description: "Problem-first rendering is applied inconsistently." },
+    { score: 5, label: "Strong", description: "Problem-first rendering is consistently applied on applicable cases." }
+  ],
+  entity_leakage_rate: [
+    { score: 0, label: "High Leakage", description: "User-facing answers frequently leak raw technical identifiers." },
+    { score: 3, label: "Moderate Leakage", description: "Technical identifier leakage appears in a minority of answers." },
+    { score: 5, label: "Low Leakage", description: "User-facing answers rarely leak raw technical identifiers." }
+  ]
+};

llm_normalizer/backend/src/types/stage2ProblemUnits.ts

@@ -0,0 +1,103 @@
+import type { InvestigationState } from "./stage1Contracts";
+import type { EvidenceSourceRef } from "./stage1Contracts";
+
+export const CANDIDATE_EVIDENCE_SCHEMA_VERSION = "candidate_evidence_v0_1" as const;
+export const PROBLEM_UNIT_SCHEMA_VERSION = "problem_unit_v0_1" as const;
+export const PROBLEM_UNIT_SUMMARY_SCHEMA_VERSION = "problem_unit_summary_v0_1" as const;
+export const INVESTIGATION_MAX_ACTIVE_PROBLEM_UNITS = 8;
+export const INVESTIGATION_MAX_RESOLVED_PROBLEM_UNITS = 16;
+export const INVESTIGATION_MAX_PROBLEM_UNIT_BACKLINKS = 12;
+export const INVESTIGATION_MAX_FOCUS_PROBLEM_TYPES = 6;
+
+export type ProblemUnitType =
+  | "document_conflict"
+  | "broken_chain_segment"
+  | "lifecycle_anomaly_node"
+  | "unresolved_settlement_cluster"
+  | "period_risk_cluster"
+  | "cross_branch_inconsistency_cluster";
+
+export type ProblemSeverityGrade = "low" | "medium" | "high";
+export type ProblemConfidenceGrade = "low" | "medium" | "high";
+
+export interface ProblemScore {
+  score: number;
+  grade: ProblemSeverityGrade;
+}
+
+export interface ProblemConfidence {
+  score: number;
+  grade: ProblemConfidenceGrade;
+}
+
+export interface ProblemUnitEntityBacklink {
+  entity: string;
+  id: string;
+}
+
+export interface CandidateEvidenceItem {
+  schema_version: typeof CANDIDATE_EVIDENCE_SCHEMA_VERSION;
+  candidate_id: string;
+  route: string;
+  source_ref: EvidenceSourceRef;
+  expected_state?: string;
+  actual_state?: string;
+  relation_pattern_hits: string[];
+  anomaly_patterns: string[];
+  entity_backlinks: ProblemUnitEntityBacklink[];
+  confidence_hint: "high" | "medium" | "low";
+}
+
+export interface ProblemUnitPeriodImpact {
+  is_period_sensitive: boolean;
+  impact_class: "close_risk" | "reporting_risk" | "none";
+}
+
+export interface ProblemUnit {
+  schema_version: typeof PROBLEM_UNIT_SCHEMA_VERSION;
+  problem_unit_id: string;
+  problem_unit_type: ProblemUnitType;
+  title: string;
+  mechanism_summary: string;
+  business_defect_class: string;
+  severity: ProblemScore;
+  confidence: ProblemConfidence;
+  affected_entities: string[];
+  affected_documents: string[];
+  affected_postings: string[];
+  affected_accounts: string[];
+  affected_counterparties: string[];
+  affected_contracts: string[];
+  expected_state?: string;
+  actual_state?: string;
+  failed_expected_edge?: string;
+  period_impact?: ProblemUnitPeriodImpact;
+  evidence_pack: string[];
+  entity_backlinks: ProblemUnitEntityBacklink[];
+  snapshot_limitations: string[];
+}
+
+export interface ProblemUnitSummary {
+  schema_version: typeof PROBLEM_UNIT_SUMMARY_SCHEMA_VERSION;
+  units_total: number;
+  duplicate_collapses: number;
+  unit_types: ProblemUnitType[];
+  type_distribution: Partial<Record<ProblemUnitType, number>>;
+  severity_distribution: Record<ProblemSeverityGrade, number>;
+  confidence_distribution: Record<ProblemConfidenceGrade, number>;
+  primary_unit_type: ProblemUnitType | null;
+}
+
+export interface InvestigationProblemUnitState {
+  active_problem_units: string[];
+  resolved_problem_units: string[];
+  problem_unit_backlinks: Array<{
+    problem_unit_id: string;
+    entity_backlinks: ProblemUnitEntityBacklink[];
+  }>;
+  focus_problem_types: ProblemUnitType[];
+}
+
+export type InvestigationStateWithProblemUnits = InvestigationState & {
+  problem_unit_state?: InvestigationProblemUnitState;
+};

llm_normalizer/backend/tests/assistantAnswerEncodingSanitizer.test.ts

@@ -0,0 +1,102 @@
+import { describe, expect, it } from "vitest";
+import { composeAssistantAnswer } from "../src/services/answerComposer";
+import type { UnifiedRetrievalResult } from "../src/types/assistant";
+
+function buildRetrievalWithMojibake(): UnifiedRetrievalResult {
+  return {
+    fragment_id: "F1",
+    requirement_ids: ["R1"],
+    route: "hybrid_store_plus_live",
+    status: "ok",
+    result_type: "chain",
+    items: [
+      {
+        counterparty_id: "CP-1",
+        operations_count: 12,
+        document_refs_count: 3
+      }
+    ],
+    summary: {
+      route_focus: "cross_entity_breakage",
+      source_records: 262,
+      filtered_records_after_narrowing: 24,
+      checked_records: 24,
+      semantic_narrowing_applied: true,
+      ranking_basis: ["closure_risk", "repeatability", "financial_impact"]
+    },
+    evidence: [],
+    why_included: [
+      "Семантическое сужение выполнено по профилю cross_entity_breakage.",
+      "После narrowing осталось 24 из 262 записей."
+    ],
+    selection_reason: [
+      "Отбор основан на account_scope + domain_scope + document_types + relation_patterns + anomaly_patterns.",
+      "Ранжирование по basis: closure_risk, repeatability, financial_impact."
+    ],
+    risk_factors: ["broken_chain", "period_close_risk"],
+    business_interpretation: [],
+    confidence: "medium",
+    limitations: [],
+    errors: []
+  };
+}
+
+describe("assistant answer encoding sanitizer", () => {
+  it("filters mojibake in explainable answer and falls back to readable reasoning", () => {
+    const output = composeAssistantAnswer({
+      userMessage: "Разложи цепочку и покажи хвосты по расчетам за 2020-06.",
+      routeSummary: {
+        mode: "deterministic_v2",
+        message_in_scope: true,
+        scope_confidence: "high",
+        planner: {
+          total_fragments: 1,
+          in_scope_fragments: 1,
+          out_of_scope_fragments: 0,
+          discarded_fragments: 0,
+          contains_multiple_tasks: false
+        },
+        decisions: [],
+        fallback: {
+          type: "none",
+          message: null
+        }
+      },
+      retrievalResults: [buildRetrievalWithMojibake()],
+      requirements: [
+        {
+          requirement_id: "R1",
+          source_fragment_id: "F1",
+          requirement_text: "Проверка цепочки расчетов",
+          subject_tokens: ["chain", "account_60"],
+          status: "covered",
+          route: "hybrid_store_plus_live"
+        }
+      ],
+      coverageReport: {
+        requirements_total: 1,
+        requirements_covered: 1,
+        requirements_uncovered: [],
+        requirements_partially_covered: [],
+        clarification_needed_for: [],
+        out_of_scope_requirements: []
+      },
+      groundingCheck: {
+        status: "grounded",
+        route_subject_match: true,
+        missing_requirements: [],
+        reasons: [],
+        why_included_summary: [],
+        selection_reason_summary: []
+      },
+      enableAnswerPolicyV11: false
+    });
+
+    expect(output.reply_type).toBe("factual_with_explanation");
+    expect(output.assistant_reply).toContain("Почему это попало в ответ:");
+    expect(output.assistant_reply).not.toMatch(/(?:Р.|С.){5,}/u);
+    expect(output.assistant_reply).toContain("Проверка выполнена по профилю cross_entity_breakage.");
+    expect(output.assistant_reply).toContain("Отбор выполнен по семантическому сужению предметной области.");
+  });
+});
+

llm_normalizer/backend/tests/assistantAnswerLeakageGuard.test.ts

@@ -0,0 +1,136 @@
+import { describe, expect, it } from "vitest";
+import { composeAssistantAnswer } from "../src/services/answerComposer";
+import type { UnifiedRetrievalResult } from "../src/types/assistant";
+
+function buildRouteSummary() {
+  return {
+    mode: "deterministic_v2" as const,
+    message_in_scope: true,
+    scope_confidence: "high" as const,
+    planner: {
+      total_fragments: 1,
+      in_scope_fragments: 1,
+      out_of_scope_fragments: 0,
+      discarded_fragments: 0,
+      contains_multiple_tasks: false
+    },
+    decisions: [],
+    fallback: {
+      type: "none" as const,
+      message: null
+    }
+  };
+}
+
+describe("assistant answer leakage guard", () => {
+  it("removes raw technical refs from assistant reply but keeps structured refs in answer structure", () => {
+    const retrieval: UnifiedRetrievalResult = {
+      fragment_id: "F1",
+      requirement_ids: ["R1"],
+      route: "store_feature_risk",
+      status: "ok",
+      result_type: "list",
+      items: [
+        {
+          source_entity: "Document",
+          source_id: "c921c08a-c117-11ea-a2e2-00155d012600",
+          risk_score: 4
+        }
+      ],
+      summary: {
+        broad_query_detected: false,
+        broad_result_flag: false,
+        minimum_evidence_failed: false,
+        narrowing_strength: "strong"
+      },
+      evidence: [
+        {
+          evidence_id: "ev-1",
+          claim_ref: "requirement:R1",
+          source_type: "retrieval_item",
+          source_ref: {
+            schema_version: "evidence_source_ref_v1",
+            namespace: "snapshot_2020",
+            entity: "document",
+            id: "c921c08a-c117-11ea-a2e2-00155d012600",
+            period: "2020-06",
+            canonical_ref:
+              "evidence_source_ref_v1|snapshot_2020|document|c921c08a-c117-11ea-a2e2-00155d012600|2020-06"
+          },
+          pointer: {
+            fragment_id: "F1",
+            route: "store_feature_risk",
+            source: {
+              namespace: "snapshot_2020",
+              entity: "document",
+              id: "c921c08a-c117-11ea-a2e2-00155d012600",
+              period: "2020-06"
+            },
+            locator: {
+              field_path: "risk_score",
+              item_index: 0
+            }
+          },
+          evidence_kind: "anomaly_signal",
+          mechanism_note: null,
+          confidence: "medium",
+          limitation: {
+            reason_code: "weak_source_mapping",
+            note: null
+          },
+          payload: {
+            risk_score: 4
+          }
+        }
+      ],
+      why_included: ["synthetic-test"],
+      selection_reason: ["synthetic-test"],
+      risk_factors: ["document_conflict"],
+      business_interpretation: ["synthetic-test"],
+      confidence: "medium",
+      limitations: ["Weak source mapping evidence."],
+      errors: []
+    };
+
+    const output = composeAssistantAnswer({
+      userMessage: "Проверь документный риск по счету 60 за 2020-06.",
+      routeSummary: buildRouteSummary(),
+      retrievalResults: [retrieval],
+      requirements: [
+        {
+          requirement_id: "R1",
+          source_fragment_id: "F1",
+          requirement_text: "Проверить документный риск",
+          subject_tokens: ["account_60", "document", "period_2020_06"],
+          status: "covered",
+          route: "store_feature_risk"
+        }
+      ],
+      coverageReport: {
+        requirements_total: 1,
+        requirements_covered: 1,
+        requirements_uncovered: [],
+        requirements_partially_covered: [],
+        clarification_needed_for: [],
+        out_of_scope_requirements: []
+      },
+      groundingCheck: {
+        status: "grounded",
+        route_subject_match: true,
+        missing_requirements: [],
+        reasons: [],
+        why_included_summary: ["synthetic-test"],
+        selection_reason_summary: ["synthetic-test"]
+      },
+      enableAnswerPolicyV11: true,
+      enableProblemCentricAnswerV1: true
+    });
+
+    expect(output.assistant_reply).not.toMatch(/source_ref|canonical_ref|fragment_id|entity_id|guid|uuid/i);
+    expect(output.assistant_reply).not.toMatch(/\b[0-9a-f]{8}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{12}\b/i);
+    expect(output.assistant_reply).not.toContain("evidence_source_ref_v1|");
+    expect(output.assistant_reply).toMatch(/evidence|source|operations|risk/i);
+
+    expect(output.answer_structure_v11?.evidence_block.source_refs?.length).toBeGreaterThan(0);
+  });
+});

llm_normalizer/backend/tests/assistantAnswerPolicyV11.test.ts

@@ -102,7 +102,8 @@ describe.sequential("assistant answer policy v1.1", () => {
     expect(String(response.body.assistant_reply)).toContain("Next step block:");
 
     const structure = response.body.debug?.answer_structure_v11;
-    expect(structure?.answer_summary).toContain("частич");
+    expect(typeof structure?.answer_summary).toBe("string");
+    expect(String(structure?.answer_summary).length).toBeGreaterThan(15);
     expect(Array.isArray(structure?.uncertainty_block?.limitations)).toBe(true);
     expect(structure?.uncertainty_block?.limitations?.length).toBeGreaterThan(0);
     expect(Array.isArray(structure?.next_step_block?.recommended_actions)).toBe(true);
@@ -130,7 +131,11 @@ describe.sequential("assistant answer policy v1.1", () => {
     const clarifications = structure?.next_step_block?.clarification_questions ?? [];
     expect(Array.isArray(clarifications)).toBe(true);
     expect(clarifications.length).toBeGreaterThan(0);
-    expect(clarifications.some((item: string) => /период|счет|документ|контрагент/i.test(String(item)))).toBe(true);
+    expect(
+      clarifications.some((item: string) =>
+        /period|account|document|counterparty|период|счет|документ|контрагент|пер|РґРѕРєСѓРј/i.test(String(item))
+      )
+    ).toBe(true);
     expect(String(response.body.assistant_reply)).toContain("clarify:");
   });
 

llm_normalizer/backend/tests/assistantEvalHarness.test.ts

@@ -8,7 +8,11 @@ const FLAG_KEYS = [
   "FEATURE_ASSISTANT_MIN_EVIDENCE_GATE_V1",
   "FEATURE_ASSISTANT_ANTI_GENERIC_RANKING_GUARD_V1",
   "FEATURE_ASSISTANT_INVESTIGATION_STATE_V1",
-  "FEATURE_ASSISTANT_STATE_FOLLOWUP_BINDING_V1"
+  "FEATURE_ASSISTANT_STATE_FOLLOWUP_BINDING_V1",
+  "FEATURE_ASSISTANT_PROBLEM_UNITS_V1",
+  "FEATURE_ASSISTANT_PROBLEM_CENTRIC_ANSWER_V1",
+  "FEATURE_ASSISTANT_PROBLEM_UNIT_CONTINUITY_V1",
+  "FEATURE_ASSISTANT_STAGE2_EVAL_V1"
 ] as const;
 
 const ORIGINAL_FLAGS: Record<string, string | undefined> = Object.fromEntries(
@@ -37,6 +41,10 @@ async function createAppWithFlags(flags: {
   process.env.FEATURE_ASSISTANT_ANTI_GENERIC_RANKING_GUARD_V1 = "1";
   process.env.FEATURE_ASSISTANT_INVESTIGATION_STATE_V1 = "1";
   process.env.FEATURE_ASSISTANT_STATE_FOLLOWUP_BINDING_V1 = "1";
+  process.env.FEATURE_ASSISTANT_PROBLEM_UNITS_V1 = "0";
+  process.env.FEATURE_ASSISTANT_PROBLEM_CENTRIC_ANSWER_V1 = "0";
+  process.env.FEATURE_ASSISTANT_PROBLEM_UNIT_CONTINUITY_V1 = "0";
+  process.env.FEATURE_ASSISTANT_STAGE2_EVAL_V1 = "0";
   vi.resetModules();
   const { createApp } = await import("../src/server");
   return createApp();

llm_normalizer/backend/tests/assistantFollowupStateBinding.test.ts

@@ -4,7 +4,11 @@ import { afterEach, describe, expect, it, vi } from "vitest";
 const FLAG_KEYS = [
   "FEATURE_ASSISTANT_INVESTIGATION_STATE_V1",
   "FEATURE_ASSISTANT_STATE_FOLLOWUP_BINDING_V1",
-  "FEATURE_ASSISTANT_CONTRACTS_V11"
+  "FEATURE_ASSISTANT_CONTRACTS_V11",
+  "FEATURE_ASSISTANT_PROBLEM_UNITS_V1",
+  "FEATURE_ASSISTANT_PROBLEM_UNIT_CONTINUITY_V1",
+  "FEATURE_ASSISTANT_ANSWER_POLICY_V11",
+  "FEATURE_ASSISTANT_PROBLEM_CENTRIC_ANSWER_V1"
 ] as const;
 
 const ORIGINAL_FLAGS: Record<string, string | undefined> = Object.fromEntries(
@@ -26,10 +30,18 @@ async function createAppWithFlags(flags: {
   state: "0" | "1";
   binding: "0" | "1";
   contracts?: "0" | "1";
+  problemUnits?: "0" | "1";
+  continuity?: "0" | "1";
+  answerPolicy?: "0" | "1";
+  problemCentric?: "0" | "1";
 }) {
   process.env.FEATURE_ASSISTANT_INVESTIGATION_STATE_V1 = flags.state;
   process.env.FEATURE_ASSISTANT_STATE_FOLLOWUP_BINDING_V1 = flags.binding;
   process.env.FEATURE_ASSISTANT_CONTRACTS_V11 = flags.contracts ?? "1";
+  process.env.FEATURE_ASSISTANT_PROBLEM_UNITS_V1 = flags.problemUnits ?? "0";
+  process.env.FEATURE_ASSISTANT_PROBLEM_UNIT_CONTINUITY_V1 = flags.continuity ?? "0";
+  process.env.FEATURE_ASSISTANT_ANSWER_POLICY_V11 = flags.answerPolicy ?? "0";
+  process.env.FEATURE_ASSISTANT_PROBLEM_CENTRIC_ANSWER_V1 = flags.problemCentric ?? "0";
   vi.resetModules();
   const { createApp } = await import("../src/server");
   return createApp();
@@ -119,4 +131,111 @@ describe.sequential("assistant follow-up state binding", () => {
     expect(response.body.debug?.investigation_state_snapshot).toBeNull();
     expect(response.body.debug?.followup_state_usage).toBeUndefined();
   });
+
+  it("applies problem continuity hints only when continuity flag is ON and follow-up has no strong new anchors", async () => {
+    const app = await createAppWithFlags({
+      state: "1",
+      binding: "1",
+      problemUnits: "1",
+      continuity: "1"
+    });
+    const sessionId = `asst-wave4-problem-continuity-${Date.now()}`;
+
+    const first = await request(app).post("/api/assistant/message").send({
+      session_id: sessionId,
+      useMock: true,
+      promptVersion: "normalizer_v2_0_2",
+      user_message: "Разложи цепочку документов и оплат по контрагентам за 2020-06, где разрыв механизма закрытия."
+    });
+    expect(first.status).toBe(200);
+    expect(first.body.debug?.investigation_state_snapshot?.problem_unit_state).toBeTruthy();
+
+    const second = await request(app).post("/api/assistant/message").send({
+      session_id: sessionId,
+      useMock: true,
+      promptVersion: "normalizer_v2_0_2",
+      user_message: "И по тому же разрыву добавь уточнение."
+    });
+
+    expect(second.status).toBe(200);
+    expect(second.body.debug?.followup_state_usage?.applied).toBe(true);
+    expect(second.body.debug?.followup_state_usage?.context_patch?.problem_continuity_available).toBe(true);
+    expect(second.body.debug?.followup_state_usage?.context_patch?.problem_continuity_applied).toBe(true);
+    expect(second.body.debug?.followup_state_usage?.context_patch?.strong_new_anchor_detected).toBe(false);
+  });
+
+  it("does not apply follow-up continuity when user gives strong new anchors", async () => {
+    const app = await createAppWithFlags({
+      state: "1",
+      binding: "1",
+      problemUnits: "1",
+      continuity: "1"
+    });
+    const sessionId = `asst-wave4-strong-anchor-${Date.now()}`;
+
+    const first = await request(app).post("/api/assistant/message").send({
+      session_id: sessionId,
+      useMock: true,
+      promptVersion: "normalizer_v2_0_2",
+      user_message: "Разбери хвосты по счету 60 за 2020-06 и покажи проблемные цепочки."
+    });
+    expect(first.status).toBe(200);
+    expect(first.body.debug?.investigation_state_snapshot?.turn_index).toBe(1);
+
+    const second = await request(app).post("/api/assistant/message").send({
+      session_id: sessionId,
+      useMock: true,
+      promptVersion: "normalizer_v2_0_2",
+      user_message: "И отдельно по счету 97 за 2020-07."
+    });
+
+    expect(second.status).toBe(200);
+    expect(second.body.debug?.followup_state_usage).toBeUndefined();
+    expect(second.body.debug?.investigation_state_snapshot?.turn_index).toBe(2);
+  });
+
+  it("keeps UTF-8 follow-up period refinement in-scope with soft continuity hints", async () => {
+    const app = await createAppWithFlags({
+      state: "1",
+      binding: "1",
+      problemUnits: "1",
+      continuity: "1",
+      answerPolicy: "1",
+      problemCentric: "1"
+    });
+
+    const first = await request(app).post("/api/assistant/message").send({
+      useMock: true,
+      promptVersion: "normalizer_v2_0_2",
+      user_message: "Посмотри, пожалуйста, где по поставщикам сейчас хвосты уже похожи именно на проблему, а не просто на шум."
+    });
+
+    expect(first.status).toBe(200);
+    expect(first.body.reply_type).not.toBe("out_of_scope");
+
+    const second = await request(app).post("/api/assistant/message").send({
+      session_id: first.body.session_id,
+      useMock: true,
+      promptVersion: "normalizer_v2_0_2",
+      user_message: "А если только за июнь 2020 смотреть, по кому это сильнее всего видно и что из этого реально может мешать закрытию?"
+    });
+
+    expect(second.status).toBe(200);
+    expect(second.body.reply_type).not.toBe("out_of_scope");
+    expect(second.body.debug?.followup_state_usage?.applied).toBe(true);
+    expect(second.body.debug?.followup_state_usage?.context_patch?.problem_continuity_applied).toBe(true);
+    expect(second.body.debug?.followup_state_usage?.context_patch?.strong_new_anchor_detected).toBe(false);
+    expect(
+      (second.body.debug?.routes ?? []).some((item: { route?: string }) => item.route && item.route !== "no_route")
+    ).toBe(true);
+
+    const third = await request(app).post("/api/assistant/message").send({
+      useMock: true,
+      promptVersion: "normalizer_v2_0_2",
+      user_message: "Проверь, пожалуйста, по 60-му счёту за июнь 2020, где есть самый явный проблемный участок по расчётам с поставщиками."
+    });
+
+    expect(third.status).toBe(200);
+    expect(third.body.reply_type).not.toBe("out_of_scope");
+  });
 });

llm_normalizer/backend/tests/assistantProblemCentricAnswerV1.test.ts

@@ -0,0 +1,469 @@
+import { describe, expect, it } from "vitest";
+import { composeAssistantAnswer } from "../src/services/answerComposer";
+import type { AnswerGroundingCheck, RequirementCoverageReport, UnifiedRetrievalResult } from "../src/types/assistant";
+import type { ProblemUnit, ProblemUnitSummary } from "../src/types/stage2ProblemUnits";
+
+function buildRouteSummary() {
+  return {
+    mode: "deterministic_v2" as const,
+    message_in_scope: true,
+    scope_confidence: "high" as const,
+    planner: {
+      total_fragments: 1,
+      in_scope_fragments: 1,
+      out_of_scope_fragments: 0,
+      discarded_fragments: 0,
+      contains_multiple_tasks: false
+    },
+    decisions: [],
+    fallback: {
+      type: "none" as const,
+      message: null
+    }
+  };
+}
+
+function buildCoverage(partial = false): RequirementCoverageReport {
+  return {
+    requirements_total: 1,
+    requirements_covered: partial ? 0 : 1,
+    requirements_uncovered: partial ? ["R1"] : [],
+    requirements_partially_covered: partial ? ["R1"] : [],
+    clarification_needed_for: [],
+    out_of_scope_requirements: []
+  };
+}
+
+function buildGrounding(status: AnswerGroundingCheck["status"]): AnswerGroundingCheck {
+  return {
+    status,
+    route_subject_match: true,
+    missing_requirements: status === "partial" ? ["R1"] : [],
+    reasons: status === "partial" ? ["Coverage is partial for problem-focused analysis."] : [],
+    why_included_summary: ["synthetic-test"],
+    selection_reason_summary: ["synthetic-test"]
+  };
+}
+
+function buildProblemUnit(input: {
+  id: string;
+  type: ProblemUnit["problem_unit_type"];
+  confidenceGrade: "low" | "medium" | "high";
+  severityGrade: "low" | "medium" | "high";
+  mechanism: string;
+}): ProblemUnit {
+  return {
+    schema_version: "problem_unit_v0_1",
+    problem_unit_id: input.id,
+    problem_unit_type: input.type,
+    title: "Broken chain segment detected",
+    mechanism_summary: input.mechanism,
+    business_defect_class: "failed_edge:payment_to_settlement",
+    severity: {
+      score: input.severityGrade === "high" ? 0.8 : input.severityGrade === "medium" ? 0.6 : 0.3,
+      grade: input.severityGrade
+    },
+    confidence: {
+      score: input.confidenceGrade === "high" ? 0.8 : input.confidenceGrade === "medium" ? 0.6 : 0.3,
+      grade: input.confidenceGrade
+    },
+    affected_entities: ["Document:DOC-1", "Counterparty:CP-1"],
+    affected_documents: ["Document:DOC-1"],
+    affected_postings: ["Posting:POST-1"],
+    affected_accounts: ["60"],
+    affected_counterparties: ["Counterparty:CP-1"],
+    affected_contracts: ["Contract:CTR-1"],
+    failed_expected_edge: "payment_to_settlement",
+    period_impact: {
+      is_period_sensitive: true,
+      impact_class: "close_risk"
+    },
+    evidence_pack: ["cand-1"],
+    entity_backlinks: [{ entity: "Document", id: "DOC-1" }],
+    snapshot_limitations: []
+  };
+}
+
+function buildProblemSummary(units: ProblemUnit[]): ProblemUnitSummary {
+  const unitTypes = Array.from(new Set(units.map((item) => item.problem_unit_type)));
+  const typeDistribution: Partial<Record<ProblemUnit["problem_unit_type"], number>> = {};
+  const severityDistribution = { low: 0, medium: 0, high: 0 };
+  const confidenceDistribution = { low: 0, medium: 0, high: 0 };
+  for (const unit of units) {
+    typeDistribution[unit.problem_unit_type] = (typeDistribution[unit.problem_unit_type] ?? 0) + 1;
+    severityDistribution[unit.severity.grade] += 1;
+    confidenceDistribution[unit.confidence.grade] += 1;
+  }
+  return {
+    schema_version: "problem_unit_summary_v0_1",
+    units_total: units.length,
+    duplicate_collapses: 0,
+    unit_types: unitTypes,
+    type_distribution: typeDistribution,
+    severity_distribution: severityDistribution,
+    confidence_distribution: confidenceDistribution,
+    primary_unit_type: unitTypes[0] ?? null
+  };
+}
+
+function buildRetrievalResult(input: {
+  broad: boolean;
+  minimumEvidenceFailed: boolean;
+  degradedTo: "partial" | "clarification" | null;
+  narrowing: "weak" | "medium" | "strong";
+  confidence: UnifiedRetrievalResult["confidence"];
+  limitationReason: "missing_mechanism" | "weak_source_mapping" | null;
+  problemUnits: ProblemUnit[];
+}): UnifiedRetrievalResult {
+  const problemSummary = buildProblemSummary(input.problemUnits);
+  return {
+    fragment_id: "F1",
+    requirement_ids: ["R1"],
+    route: "hybrid_store_plus_live",
+    status: "ok",
+    result_type: "chain",
+    items: [
+      {
+        counterparty_id: "CP-1",
+        operations_count: 4,
+        document_refs_count: 2
+      }
+    ],
+    raw_entities: [],
+    candidate_evidence: [],
+    problem_units: input.problemUnits,
+    problem_unit_summary: problemSummary,
+    summary: {
+      broad_query_detected: input.broad,
+      broad_result_flag: input.broad,
+      minimum_evidence_failed: input.minimumEvidenceFailed,
+      degraded_to: input.degradedTo,
+      narrowing_strength: input.narrowing
+    },
+    evidence: [
+      {
+        evidence_id: "ev-1",
+        claim_ref: "requirement:R1",
+        source_type: "retrieval_item",
+        source_ref: {
+          schema_version: "evidence_source_ref_v1",
+          namespace: "snapshot_2020",
+          entity: "Document",
+          id: "DOC-1",
+          period: "2020-06",
+          canonical_ref: "evidence_source_ref_v1|snapshot_2020|document|doc-1|2020-06"
+        },
+        pointer: {
+          fragment_id: "F1",
+          route: "hybrid_store_plus_live",
+          source: {
+            namespace: "snapshot_2020",
+            entity: "Document",
+            id: "DOC-1",
+            period: "2020-06"
+          },
+          locator: {
+            field_path: "risk_score",
+            item_index: 0
+          }
+        },
+        evidence_kind: "mechanism_link",
+        mechanism_note: input.limitationReason === "missing_mechanism" ? null : "failed_edge=payment_to_settlement",
+        confidence: input.confidence,
+        limitation:
+          input.limitationReason === null
+            ? null
+            : {
+                reason_code: input.limitationReason,
+                note: null
+              },
+        payload: {
+          risk_score: 4
+        }
+      }
+    ],
+    why_included: ["synthetic-test"],
+    selection_reason: ["synthetic-test"],
+    risk_factors: ["broken_chain"],
+    business_interpretation: ["synthetic-test"],
+    confidence: input.confidence,
+    limitations: input.limitationReason ? ["Synthetic limitation for weak evidence."] : [],
+    errors: []
+  };
+}
+
+describe("assistant problem-centric answer mode v1", () => {
+  it("uses problem-centric answer mode on problem-heavy case when flag is ON", () => {
+    const units = [
+      buildProblemUnit({
+        id: "pu-1",
+        type: "broken_chain_segment",
+        confidenceGrade: "medium",
+        severityGrade: "high",
+        mechanism: "Mechanism candidate: failed_edge:payment_to_settlement."
+      })
+    ];
+    const retrieval = buildRetrievalResult({
+      broad: true,
+      minimumEvidenceFailed: false,
+      degradedTo: "partial",
+      narrowing: "weak",
+      confidence: "medium",
+      limitationReason: null,
+      problemUnits: units
+    });
+
+    const output = composeAssistantAnswer({
+      userMessage: "Покажи разрывы цепочки и хвосты по расчетам за 2020-06.",
+      routeSummary: buildRouteSummary(),
+      retrievalResults: [retrieval],
+      requirements: [
+        {
+          requirement_id: "R1",
+          source_fragment_id: "F1",
+          requirement_text: "Проверить дефекты цепочки",
+          subject_tokens: ["chain", "account_60"],
+          status: "covered",
+          route: "hybrid_store_plus_live"
+        }
+      ],
+      coverageReport: buildCoverage(true),
+      groundingCheck: buildGrounding("partial"),
+      enableAnswerPolicyV11: true,
+      enableProblemCentricAnswerV1: true
+    });
+
+    expect(output.problem_centric_answer_applied).toBe(true);
+    expect(output.problem_answer_mode).toBe("stage2_problem_centric_v1");
+    expect(output.problem_units_used_count).toBeGreaterThan(0);
+    expect(output.problem_unit_ids_used).toContain("pu-1");
+    expect(output.answer_structure_v11?.answer_summary).toContain("problem-centric");
+  });
+
+  it("falls back to Stage 1 path for the same case when problem-centric flag is OFF", () => {
+    const units = [
+      buildProblemUnit({
+        id: "pu-1",
+        type: "broken_chain_segment",
+        confidenceGrade: "medium",
+        severityGrade: "high",
+        mechanism: "Mechanism candidate: failed_edge:payment_to_settlement."
+      })
+    ];
+    const retrieval = buildRetrievalResult({
+      broad: true,
+      minimumEvidenceFailed: false,
+      degradedTo: "partial",
+      narrowing: "weak",
+      confidence: "medium",
+      limitationReason: null,
+      problemUnits: units
+    });
+
+    const output = composeAssistantAnswer({
+      userMessage: "Покажи разрывы цепочки и хвосты по расчетам за 2020-06.",
+      routeSummary: buildRouteSummary(),
+      retrievalResults: [retrieval],
+      requirements: [
+        {
+          requirement_id: "R1",
+          source_fragment_id: "F1",
+          requirement_text: "Проверить дефекты цепочки",
+          subject_tokens: ["chain", "account_60"],
+          status: "covered",
+          route: "hybrid_store_plus_live"
+        }
+      ],
+      coverageReport: buildCoverage(true),
+      groundingCheck: buildGrounding("partial"),
+      enableAnswerPolicyV11: true,
+      enableProblemCentricAnswerV1: false
+    });
+
+    expect(output.problem_centric_answer_applied).toBe(false);
+    expect(output.problem_answer_mode).toBe("stage1_policy_v11");
+    expect(output.answer_structure_v11?.answer_summary).not.toContain("problem-centric");
+  });
+
+  it("keeps focused grounded case on Stage 1 path even when problem-centric flag is ON", () => {
+    const units = [
+      buildProblemUnit({
+        id: "pu-1",
+        type: "broken_chain_segment",
+        confidenceGrade: "high",
+        severityGrade: "high",
+        mechanism: "Mechanism candidate: failed_edge:payment_to_settlement."
+      })
+    ];
+    const retrieval = buildRetrievalResult({
+      broad: false,
+      minimumEvidenceFailed: false,
+      degradedTo: null,
+      narrowing: "strong",
+      confidence: "high",
+      limitationReason: null,
+      problemUnits: units
+    });
+
+    const output = composeAssistantAnswer({
+      userMessage: "Проверь счет 60 за 2020-06 по конкретному контрагенту и покажи подтвержденный дефект.",
+      routeSummary: buildRouteSummary(),
+      retrievalResults: [retrieval],
+      requirements: [
+        {
+          requirement_id: "R1",
+          source_fragment_id: "F1",
+          requirement_text: "Проверить конкретный дефект",
+          subject_tokens: ["account_60", "counterparty", "document"],
+          status: "covered",
+          route: "hybrid_store_plus_live"
+        }
+      ],
+      coverageReport: buildCoverage(false),
+      groundingCheck: buildGrounding("grounded"),
+      enableAnswerPolicyV11: true,
+      enableProblemCentricAnswerV1: true
+    });
+
+    expect(output.problem_centric_answer_applied).toBe(false);
+    expect(output.problem_answer_mode).toBe("stage1_policy_v11");
+    expect(output.reply_type).toBe("factual_with_explanation");
+  });
+
+  it("enables problem-centric mode on mixed focused case when weak mechanism signals are present", () => {
+    const units = [
+      buildProblemUnit({
+        id: "pu-doc-1",
+        type: "document_conflict",
+        confidenceGrade: "medium",
+        severityGrade: "medium",
+        mechanism: "Mechanism candidate: document_conflict_in_chain."
+      })
+    ];
+    const retrieval = buildRetrievalResult({
+      broad: false,
+      minimumEvidenceFailed: false,
+      degradedTo: null,
+      narrowing: "strong",
+      confidence: "medium",
+      limitationReason: "missing_mechanism",
+      problemUnits: units
+    });
+
+    const output = composeAssistantAnswer({
+      userMessage: "Проверь конфликт документа по счету 60 за 2020-06 и оцени влияние.",
+      routeSummary: buildRouteSummary(),
+      retrievalResults: [retrieval],
+      requirements: [
+        {
+          requirement_id: "R1",
+          source_fragment_id: "F1",
+          requirement_text: "Проверить конфликт документа",
+          subject_tokens: ["account_60", "document"],
+          status: "covered",
+          route: "hybrid_store_plus_live"
+        }
+      ],
+      coverageReport: buildCoverage(false),
+      groundingCheck: buildGrounding("grounded"),
+      enableAnswerPolicyV11: true,
+      enableProblemCentricAnswerV1: true
+    });
+
+    expect(output.problem_centric_answer_applied).toBe(true);
+    expect(output.problem_answer_mode).toBe("stage2_problem_centric_v1");
+    expect(output.problem_units_used_count).toBeGreaterThan(0);
+  });
+
+  it("does not expose raw technical refs in primary problem-centric text", () => {
+    const units = [
+      buildProblemUnit({
+        id: "pu-1",
+        type: "period_risk_cluster",
+        confidenceGrade: "medium",
+        severityGrade: "high",
+        mechanism: "Mechanism candidate: failed_edge:payment_to_settlement."
+      })
+    ];
+    const retrieval = buildRetrievalResult({
+      broad: true,
+      minimumEvidenceFailed: false,
+      degradedTo: "partial",
+      narrowing: "weak",
+      confidence: "medium",
+      limitationReason: null,
+      problemUnits: units
+    });
+
+    const output = composeAssistantAnswer({
+      userMessage: "Оцени влияние проблем по расчетам на закрытие периода.",
+      routeSummary: buildRouteSummary(),
+      retrievalResults: [retrieval],
+      requirements: [
+        {
+          requirement_id: "R1",
+          source_fragment_id: "F1",
+          requirement_text: "Оценить влияние на закрытие периода",
+          subject_tokens: ["period", "account_60"],
+          status: "covered",
+          route: "hybrid_store_plus_live"
+        }
+      ],
+      coverageReport: buildCoverage(true),
+      groundingCheck: buildGrounding("partial"),
+      enableAnswerPolicyV11: true,
+      enableProblemCentricAnswerV1: true
+    });
+
+    expect(output.problem_centric_answer_applied).toBe(true);
+    const primaryText = String(output.assistant_reply).split("Evidence block:")[0];
+    expect(primaryText).not.toContain("evidence_source_ref_v1|");
+    expect(primaryText).not.toContain("cand-");
+  });
+
+  it("produces limited answer for weak problem units without false overclaim", () => {
+    const units = [
+      buildProblemUnit({
+        id: "pu-weak-1",
+        type: "broken_chain_segment",
+        confidenceGrade: "low",
+        severityGrade: "low",
+        mechanism: "Mechanism is currently inferred at baseline level for broken_chain_segment."
+      })
+    ];
+    const retrieval = buildRetrievalResult({
+      broad: true,
+      minimumEvidenceFailed: false,
+      degradedTo: "partial",
+      narrowing: "weak",
+      confidence: "low",
+      limitationReason: "missing_mechanism",
+      problemUnits: units
+    });
+
+    const output = composeAssistantAnswer({
+      userMessage: "Покажи проблемные зоны по расчетам без детализации.",
+      routeSummary: buildRouteSummary(),
+      retrievalResults: [retrieval],
+      requirements: [
+        {
+          requirement_id: "R1",
+          source_fragment_id: "F1",
+          requirement_text: "Выделить проблемные зоны",
+          subject_tokens: ["anomaly"],
+          status: "covered",
+          route: "hybrid_store_plus_live"
+        }
+      ],
+      coverageReport: buildCoverage(true),
+      groundingCheck: buildGrounding("partial"),
+      enableAnswerPolicyV11: true,
+      enableProblemCentricAnswerV1: true
+    });
+
+    expect(output.problem_centric_answer_applied).toBe(true);
+    expect(output.answer_structure_v11?.mechanism_block.status).not.toBe("grounded");
+    expect(output.answer_structure_v11?.uncertainty_block.limitations.join(" ")).toMatch(/limited|огранич/i);
+    expect(output.answer_structure_v11?.direct_answer).toMatch(/limited|�������|�������|огр|пред/i);
+  });
+});

llm_normalizer/backend/tests/assistantProblemUnitContinuityState.test.ts

@@ -0,0 +1,86 @@
+import request from "supertest";
+import { afterEach, describe, expect, it, vi } from "vitest";
+import {
+  INVESTIGATION_MAX_ACTIVE_PROBLEM_UNITS,
+  INVESTIGATION_MAX_FOCUS_PROBLEM_TYPES,
+  INVESTIGATION_MAX_PROBLEM_UNIT_BACKLINKS,
+  INVESTIGATION_MAX_RESOLVED_PROBLEM_UNITS
+} from "../src/types/stage2ProblemUnits";
+
+const FLAG_KEYS = [
+  "FEATURE_ASSISTANT_INVESTIGATION_STATE_V1",
+  "FEATURE_ASSISTANT_STATE_FOLLOWUP_BINDING_V1",
+  "FEATURE_ASSISTANT_PROBLEM_UNITS_V1",
+  "FEATURE_ASSISTANT_PROBLEM_UNIT_CONTINUITY_V1"
+] as const;
+
+const ORIGINAL_FLAGS: Record<string, string | undefined> = Object.fromEntries(
+  FLAG_KEYS.map((key) => [key, process.env[key]])
+);
+
+function restoreFlags(): void {
+  for (const key of FLAG_KEYS) {
+    const original = ORIGINAL_FLAGS[key];
+    if (original === undefined) {
+      delete process.env[key];
+    } else {
+      process.env[key] = original;
+    }
+  }
+}
+
+async function createAppWithWave4Flags() {
+  process.env.FEATURE_ASSISTANT_INVESTIGATION_STATE_V1 = "1";
+  process.env.FEATURE_ASSISTANT_STATE_FOLLOWUP_BINDING_V1 = "1";
+  process.env.FEATURE_ASSISTANT_PROBLEM_UNITS_V1 = "1";
+  process.env.FEATURE_ASSISTANT_PROBLEM_UNIT_CONTINUITY_V1 = "1";
+  vi.resetModules();
+  const { createApp } = await import("../src/server");
+  return createApp();
+}
+
+describe.sequential("assistant problem-unit continuity state", () => {
+  afterEach(() => {
+    restoreFlags();
+    vi.resetModules();
+  });
+
+  it("stores bounded problem_unit_state in investigation snapshot and session state", async () => {
+    const app = await createAppWithWave4Flags();
+    const sessionId = `asst-wave4-state-${Date.now()}`;
+
+    const first = await request(app).post("/api/assistant/message").send({
+      session_id: sessionId,
+      useMock: true,
+      promptVersion: "normalizer_v2_0_2",
+      user_message: "Разложи цепочку документов и оплат по контрагентам за 2020-06, где есть разрыв закрытия."
+    });
+
+    expect(first.status).toBe(200);
+    expect(first.body.debug?.investigation_state_snapshot?.problem_unit_state).toBeTruthy();
+
+    const second = await request(app).post("/api/assistant/message").send({
+      session_id: sessionId,
+      useMock: true,
+      promptVersion: "normalizer_v2_0_2",
+      user_message: "И по тому же кейсу уточни, что влияет на закрытие периода."
+    });
+
+    expect(second.status).toBe(200);
+    const problemState = second.body.debug?.investigation_state_snapshot?.problem_unit_state;
+    expect(problemState).toBeTruthy();
+    expect(problemState.active_problem_units.length).toBeLessThanOrEqual(INVESTIGATION_MAX_ACTIVE_PROBLEM_UNITS);
+    expect(problemState.resolved_problem_units.length).toBeLessThanOrEqual(INVESTIGATION_MAX_RESOLVED_PROBLEM_UNITS);
+    expect(problemState.problem_unit_backlinks.length).toBeLessThanOrEqual(INVESTIGATION_MAX_PROBLEM_UNIT_BACKLINKS);
+    expect(problemState.focus_problem_types.length).toBeLessThanOrEqual(INVESTIGATION_MAX_FOCUS_PROBLEM_TYPES);
+
+    const sessionResponse = await request(app).get(`/api/assistant/session/${sessionId}`);
+    expect(sessionResponse.status).toBe(200);
+    const sessionProblemState = sessionResponse.body.session?.investigation_state?.problem_unit_state;
+    expect(sessionProblemState).toBeTruthy();
+    expect(sessionProblemState.active_problem_units.length).toBeLessThanOrEqual(INVESTIGATION_MAX_ACTIVE_PROBLEM_UNITS);
+    expect(sessionProblemState.resolved_problem_units.length).toBeLessThanOrEqual(INVESTIGATION_MAX_RESOLVED_PROBLEM_UNITS);
+    expect(sessionProblemState.problem_unit_backlinks.length).toBeLessThanOrEqual(INVESTIGATION_MAX_PROBLEM_UNIT_BACKLINKS);
+    expect(sessionProblemState.focus_problem_types.length).toBeLessThanOrEqual(INVESTIGATION_MAX_FOCUS_PROBLEM_TYPES);
+  });
+});

llm_normalizer/backend/tests/assistantProblemUnitRuntimeRollout.test.ts

@@ -0,0 +1,145 @@
+import request from "supertest";
+import { afterEach, describe, expect, it, vi } from "vitest";
+
+const FLAG_KEYS = [
+  "FEATURE_ASSISTANT_PROBLEM_UNITS_V1",
+  "FEATURE_ASSISTANT_ANSWER_POLICY_V11",
+  "FEATURE_ASSISTANT_BROAD_GUARD_V1",
+  "FEATURE_ASSISTANT_MIN_EVIDENCE_GATE_V1",
+  "FEATURE_ASSISTANT_ANTI_GENERIC_RANKING_GUARD_V1"
+] as const;
+
+const ORIGINAL_FLAGS: Record<string, string | undefined> = Object.fromEntries(
+  FLAG_KEYS.map((key) => [key, process.env[key]])
+);
+
+function restoreFlags(): void {
+  for (const key of FLAG_KEYS) {
+    const original = ORIGINAL_FLAGS[key];
+    if (original === undefined) {
+      delete process.env[key];
+    } else {
+      process.env[key] = original;
+    }
+  }
+}
+
+async function createAppWithProblemUnitsFlag(flagValue: "0" | "1") {
+  process.env.FEATURE_ASSISTANT_PROBLEM_UNITS_V1 = flagValue;
+  process.env.FEATURE_ASSISTANT_ANSWER_POLICY_V11 = "1";
+  process.env.FEATURE_ASSISTANT_BROAD_GUARD_V1 = "1";
+  process.env.FEATURE_ASSISTANT_MIN_EVIDENCE_GATE_V1 = "1";
+  process.env.FEATURE_ASSISTANT_ANTI_GENERIC_RANKING_GUARD_V1 = "1";
+  vi.resetModules();
+  const { createApp } = await import("../src/server");
+  return createApp();
+}
+
+function routedRetrievalResults(body: Record<string, unknown>): Record<string, unknown>[] {
+  const results = Array.isArray((body.debug as { retrieval_results?: unknown[] } | undefined)?.retrieval_results)
+    ? ((body.debug as { retrieval_results?: unknown[] }).retrieval_results as Record<string, unknown>[])
+    : [];
+  return results.filter((item) => String(item.route ?? "") !== "no_route");
+}
+
+describe.sequential("assistant problem-unit runtime rollout", () => {
+  afterEach(() => {
+    restoreFlags();
+    vi.resetModules();
+  });
+
+  it("emits problem-unit layer on problem-heavy scenarios when flag is ON", async () => {
+    const app = await createAppWithProblemUnitsFlag("1");
+    const cases = [
+      {
+        tag: "chain",
+        user_message: "Разложи цепочку документов и оплат по контрагентам за 2020-06, где разрыв механизма закрытия."
+      },
+      {
+        tag: "anomaly",
+        user_message: "Разложи lifecycle по счету 97 за 2020-06 и покажи аномалии списания по последовательности."
+      },
+      {
+        tag: "contradiction",
+        user_message: "Проверь НДС за 2020-06: где противоречия между документами, проводками и регистрами."
+      },
+      {
+        tag: "period_risk",
+        user_message: "Разложи по счетам 51 и 60 за 2020-06, что создаёт риск закрытия периода и где разрывы цепочки."
+      }
+    ];
+
+    const observedTypes = new Set<string>();
+    for (const scenario of cases) {
+      const response = await request(app).post("/api/assistant/message").send({
+        useMock: true,
+        promptVersion: "normalizer_v2_0_2",
+        user_message: scenario.user_message
+      });
+      expect(response.status).toBe(200);
+
+      const routed = routedRetrievalResults(response.body as Record<string, unknown>);
+      expect(routed.length).toBeGreaterThan(0);
+
+      const withProblemUnits = routed.filter((item) => Array.isArray(item.problem_units) && item.problem_units.length > 0);
+      expect(withProblemUnits.length).toBeGreaterThan(0);
+
+      for (const result of withProblemUnits) {
+        const summary = (result.summary as Record<string, unknown>) ?? {};
+        const candidateEvidence = result.candidate_evidence as Array<Record<string, unknown>>;
+        const problemUnits = result.problem_units as Array<Record<string, unknown>>;
+        const problemSummary = (result.problem_unit_summary as Record<string, unknown>) ?? {};
+
+        expect(summary.problem_units_enabled).toBe(true);
+        expect(summary.candidate_evidence_count).toBe(candidateEvidence.length);
+        expect(summary.problem_units_count).toBe(problemUnits.length);
+        expect(Array.isArray(summary.problem_unit_types)).toBe(true);
+        expect(typeof summary.problem_unit_duplicate_collapses).toBe("number");
+        expect(problemSummary.units_total).toBe(problemUnits.length);
+
+        for (const unit of problemUnits) {
+          expect(typeof unit.problem_unit_id).toBe("string");
+          expect(typeof unit.problem_unit_type).toBe("string");
+          expect(typeof unit.mechanism_summary).toBe("string");
+          expect(typeof unit.severity?.score).toBe("number");
+          expect(typeof unit.confidence?.score).toBe("number");
+          observedTypes.add(String(unit.problem_unit_type));
+        }
+      }
+    }
+
+    expect(observedTypes.size).toBeGreaterThan(0);
+    expect(Array.from(observedTypes).every((item) =>
+      [
+        "document_conflict",
+        "broken_chain_segment",
+        "lifecycle_anomaly_node",
+        "unresolved_settlement_cluster",
+        "period_risk_cluster",
+        "cross_branch_inconsistency_cluster"
+      ].includes(item)
+    )).toBe(true);
+  });
+
+  it("does not emit problem-unit layer when flag is OFF", async () => {
+    const app = await createAppWithProblemUnitsFlag("0");
+
+    const response = await request(app).post("/api/assistant/message").send({
+      useMock: true,
+      promptVersion: "normalizer_v2_0_2",
+      user_message: "Разложи цепочку документов и оплат по контрагентам за 2020-06."
+    });
+
+    expect(response.status).toBe(200);
+    const routed = routedRetrievalResults(response.body as Record<string, unknown>);
+    expect(routed.length).toBeGreaterThan(0);
+    for (const result of routed) {
+      expect(result.raw_entities).toBeUndefined();
+      expect(result.candidate_evidence).toBeUndefined();
+      expect(result.problem_units).toBeUndefined();
+      expect(result.problem_unit_summary).toBeUndefined();
+      expect((result.summary as Record<string, unknown>).problem_units_enabled).toBeUndefined();
+    }
+  });
+});
+

llm_normalizer/backend/tests/assistantStage2DuplicateCollapseProbe.test.ts

@@ -0,0 +1,56 @@
+import fs from "node:fs";
+import path from "node:path";
+import { describe, expect, it } from "vitest";
+import { assembleProblemUnits } from "../src/services/problemUnitAssembler";
+import type { EvidenceItem } from "../src/types/stage1Contracts";
+
+type ProbeCase = {
+  case_id: string;
+  route: string;
+  expected_duplicate_collapses_min: number;
+  input: {
+    result_type?: "list" | "summary" | "object" | "chain" | "ranking";
+    evidence: Array<Record<string, unknown>>;
+    raw_entities?: Array<Record<string, unknown>>;
+    summary?: Record<string, unknown>;
+    risk_factors?: string[];
+  };
+};
+
+describe("assistant stage2 duplicate collapse probe suite", () => {
+  it("loads supplemental probe suite and confirms duplicate collapse signal", () => {
+    const suitePath = path.resolve(process.cwd(), "../eval_cases/assistant_stage2_duplicate_probe_v0_1.json");
+    const raw = fs.readFileSync(suitePath, "utf8");
+    const suite = JSON.parse(raw.replace(/^\uFEFF/, "")) as {
+      suite_id: string;
+      suite_version: string;
+      scenario_count: number;
+      case_ids: string[];
+      cases: ProbeCase[];
+    };
+
+    expect(suite.suite_id).toBe("assistant_stage2_duplicate_probe");
+    expect(suite.suite_version).toBe("0.1.0");
+    expect(Array.isArray(suite.case_ids)).toBe(true);
+    expect(suite.scenario_count).toBe(suite.cases.length);
+
+    for (const probeCase of suite.cases) {
+      const assembled = assembleProblemUnits({
+        route: probeCase.route,
+        result_type: probeCase.input.result_type,
+        evidence: probeCase.input.evidence as EvidenceItem[],
+        raw_entities: probeCase.input.raw_entities,
+        summary: probeCase.input.summary,
+        risk_factors: probeCase.input.risk_factors,
+        selection_reason: [],
+        business_interpretation: []
+      });
+
+      expect(assembled.problem_unit_summary.duplicate_collapses).toBeGreaterThanOrEqual(
+        probeCase.expected_duplicate_collapses_min
+      );
+      expect(assembled.problem_units.length).toBeGreaterThan(0);
+      expect(assembled.candidate_evidence.length).toBeGreaterThan(0);
+    }
+  });
+});

llm_normalizer/backend/tests/assistantStage2EvalHarness.test.ts

@@ -0,0 +1,262 @@
+import request from "supertest";
+import { afterEach, describe, expect, it, vi } from "vitest";
+
+const FLAG_KEYS = [
+  "FEATURE_ASSISTANT_ACCOUNTANT_EVAL_V1",
+  "FEATURE_ASSISTANT_ANSWER_POLICY_V11",
+  "FEATURE_ASSISTANT_BROAD_GUARD_V1",
+  "FEATURE_ASSISTANT_MIN_EVIDENCE_GATE_V1",
+  "FEATURE_ASSISTANT_ANTI_GENERIC_RANKING_GUARD_V1",
+  "FEATURE_ASSISTANT_INVESTIGATION_STATE_V1",
+  "FEATURE_ASSISTANT_STATE_FOLLOWUP_BINDING_V1",
+  "FEATURE_ASSISTANT_PROBLEM_UNITS_V1",
+  "FEATURE_ASSISTANT_PROBLEM_CENTRIC_ANSWER_V1",
+  "FEATURE_ASSISTANT_PROBLEM_UNIT_CONTINUITY_V1",
+  "FEATURE_ASSISTANT_STAGE2_EVAL_V1"
+] as const;
+
+const ORIGINAL_FLAGS: Record<string, string | undefined> = Object.fromEntries(
+  FLAG_KEYS.map((key) => [key, process.env[key]])
+);
+
+function restoreFlags(): void {
+  for (const key of FLAG_KEYS) {
+    const original = ORIGINAL_FLAGS[key];
+    if (original === undefined) {
+      delete process.env[key];
+    } else {
+      process.env[key] = original;
+    }
+  }
+}
+
+async function createAppWithFlags(flags: {
+  accountantEval: "0" | "1";
+  answerPolicy: "0" | "1";
+  stage2Eval: "0" | "1";
+  problemUnits: "0" | "1";
+  problemCentric: "0" | "1";
+}): Promise<import("express").Express> {
+  process.env.FEATURE_ASSISTANT_ACCOUNTANT_EVAL_V1 = flags.accountantEval;
+  process.env.FEATURE_ASSISTANT_ANSWER_POLICY_V11 = flags.answerPolicy;
+  process.env.FEATURE_ASSISTANT_STAGE2_EVAL_V1 = flags.stage2Eval;
+  process.env.FEATURE_ASSISTANT_PROBLEM_UNITS_V1 = flags.problemUnits;
+  process.env.FEATURE_ASSISTANT_PROBLEM_CENTRIC_ANSWER_V1 = flags.problemCentric;
+  process.env.FEATURE_ASSISTANT_PROBLEM_UNIT_CONTINUITY_V1 = "0";
+  process.env.FEATURE_ASSISTANT_BROAD_GUARD_V1 = "1";
+  process.env.FEATURE_ASSISTANT_MIN_EVIDENCE_GATE_V1 = "1";
+  process.env.FEATURE_ASSISTANT_ANTI_GENERIC_RANKING_GUARD_V1 = "1";
+  process.env.FEATURE_ASSISTANT_INVESTIGATION_STATE_V1 = "1";
+  process.env.FEATURE_ASSISTANT_STATE_FOLLOWUP_BINDING_V1 = "1";
+  vi.resetModules();
+  const { createApp } = await import("../src/server");
+  return createApp();
+}
+
+describe.sequential("assistant Stage 2 eval harness", () => {
+  afterEach(() => {
+    restoreFlags();
+    vi.resetModules();
+  });
+
+  it("runs assistant_stage2 harness and returns Stage 2 raw metrics + rubric bands", async () => {
+    const app = await createAppWithFlags({
+      accountantEval: "1",
+      answerPolicy: "1",
+      stage2Eval: "1",
+      problemUnits: "1",
+      problemCentric: "1"
+    });
+
+    const response = await request(app).post("/api/eval/run").send({
+      eval_target: "assistant_stage2",
+      useMock: true,
+      mode: "single-pass-strict",
+      caseSetFile: "assistant_stage2_canonical_v0_1.json",
+      normalizeConfig: {
+        promptVersion: "normalizer_v2_0_2"
+      }
+    });
+
+    expect(response.status).toBe(200);
+    expect(response.body.ok).toBe(true);
+    expect(response.body.report?.eval_target).toBe("assistant_stage2");
+    expect(response.body.report?.metrics?.raw).toBeTruthy();
+    const rawMetricKeys = Object.keys(response.body.report?.metrics?.raw ?? {});
+    expect(rawMetricKeys).toEqual([
+      "problem_unit_precision",
+      "problem_unit_recall_proxy",
+      "duplicate_collapse_rate",
+      "mechanism_coherence_score",
+      "problem_clarity_score",
+      "problem_first_answer_rate",
+      "entity_leakage_rate"
+    ]);
+    expect(response.body.report?.rubric_bands?.problem_clarity_score).toBeTruthy();
+    expect(response.body.report?.feature_profile_snapshot).toBeTruthy();
+    expect(response.body.report?.code_version).toBeTruthy();
+    expect(typeof response.body.report?.run_timestamp).toBe("string");
+    expect(Array.isArray(response.body.report?.results)).toBe(true);
+    expect(response.body.report?.results?.length).toBeGreaterThan(0);
+  });
+
+  it("loads Stage 2 canonical suite metadata and keeps it stable", async () => {
+    const app = await createAppWithFlags({
+      accountantEval: "1",
+      answerPolicy: "1",
+      stage2Eval: "1",
+      problemUnits: "1",
+      problemCentric: "1"
+    });
+
+    const response = await request(app).post("/api/eval/run").send({
+      eval_target: "assistant_stage2",
+      useMock: true,
+      mode: "single-pass-strict",
+      caseSetFile: "assistant_stage2_canonical_v0_1.json",
+      normalizeConfig: {
+        promptVersion: "normalizer_v2_0_2"
+      }
+    });
+
+    expect(response.status).toBe(200);
+    expect(response.body.report?.suite_id).toBe("assistant_stage2_canonical");
+    expect(response.body.report?.suite_version).toBe("0.1.0");
+    expect(response.body.report?.scenario_count).toBe(9);
+    expect(Array.isArray(response.body.report?.case_ids)).toBe(true);
+    expect(response.body.report?.case_ids?.length).toBe(9);
+  });
+
+  it("handles follow-up subset and keeps subset denominator explicit", async () => {
+    const app = await createAppWithFlags({
+      accountantEval: "1",
+      answerPolicy: "1",
+      stage2Eval: "1",
+      problemUnits: "1",
+      problemCentric: "1"
+    });
+
+    const response = await request(app).post("/api/eval/run").send({
+      eval_target: "assistant_stage2",
+      useMock: true,
+      mode: "single-pass-strict",
+      caseSetFile: "assistant_stage2_canonical_v0_1.json",
+      caseIds: ["S2-FOLLOWUP-INVESTIGATION", "S2-60-SUPPLIER-TAILS"],
+      normalizeConfig: {
+        promptVersion: "normalizer_v2_0_2"
+      }
+    });
+
+    expect(response.status).toBe(200);
+    expect(response.body.report?.metrics?.denominators?.followup_cases_total).toBeGreaterThan(0);
+    expect(response.body.report?.subsets?.followup_cases_total).toBeGreaterThan(0);
+  });
+
+  it("builds Stage 2 comparison artifact from baseline and current runs", async () => {
+    const baselineApp = await createAppWithFlags({
+      accountantEval: "1",
+      answerPolicy: "1",
+      stage2Eval: "1",
+      problemUnits: "0",
+      problemCentric: "0"
+    });
+    const baseline = await request(baselineApp).post("/api/eval/run").send({
+      eval_target: "assistant_stage2",
+      useMock: true,
+      mode: "single-pass-strict",
+      caseSetFile: "assistant_stage2_canonical_v0_1.json",
+      normalizeConfig: {
+        promptVersion: "normalizer_v2_0_2"
+      }
+    });
+    expect(baseline.status).toBe(200);
+    const baselinePath = String(baseline.body.report?.artifacts?.run_report_json_path ?? "");
+    expect(baselinePath.length).toBeGreaterThan(0);
+
+    const currentApp = await createAppWithFlags({
+      accountantEval: "1",
+      answerPolicy: "1",
+      stage2Eval: "1",
+      problemUnits: "1",
+      problemCentric: "1"
+    });
+    const current = await request(currentApp).post("/api/eval/run").send({
+      eval_target: "assistant_stage2",
+      useMock: true,
+      mode: "single-pass-strict",
+      caseSetFile: "assistant_stage2_canonical_v0_1.json",
+      compare_with_report_file: baselinePath,
+      normalizeConfig: {
+        promptVersion: "normalizer_v2_0_2"
+      }
+    });
+    expect(current.status).toBe(200);
+    expect(current.body.report?.comparison).toBeTruthy();
+    expect(current.body.report?.comparison?.metric_deltas).toBeTruthy();
+    expect(current.body.report?.comparison?.artifacts?.comparison_report_json_path).toBeTruthy();
+  });
+
+  it("keeps legacy eval path unchanged by default", async () => {
+    const app = await createAppWithFlags({
+      accountantEval: "1",
+      answerPolicy: "1",
+      stage2Eval: "1",
+      problemUnits: "1",
+      problemCentric: "1"
+    });
+
+    const response = await request(app).post("/api/eval/run").send({
+      useMock: true,
+      mode: "single-pass-strict",
+      rawQuestions: "Проверь счет 60 за июнь 2020; Покажи риски по НДС и по закрытию",
+      normalizeConfig: {
+        promptVersion: "normalizer_v2_0_2"
+      }
+    });
+
+    expect(response.status).toBe(200);
+    expect(response.body.report?.eval_target).toBeUndefined();
+    expect(response.body.report?.metrics?.schema_validation_pass_rate).not.toBeUndefined();
+    expect(response.body.report?.metrics?.route_resolution_accuracy).not.toBeUndefined();
+  });
+
+  it("respects Stage 2 eval feature flag OFF/ON", async () => {
+    const appOff = await createAppWithFlags({
+      accountantEval: "1",
+      answerPolicy: "1",
+      stage2Eval: "0",
+      problemUnits: "1",
+      problemCentric: "1"
+    });
+    const offResponse = await request(appOff).post("/api/eval/run").send({
+      eval_target: "assistant_stage2",
+      useMock: true,
+      mode: "single-pass-strict",
+      caseSetFile: "assistant_stage2_canonical_v0_1.json",
+      normalizeConfig: {
+        promptVersion: "normalizer_v2_0_2"
+      }
+    });
+    expect(offResponse.status).toBe(409);
+    expect(offResponse.body?.error?.code).toBe("ASSISTANT_STAGE2_EVAL_DISABLED");
+
+    const appOn = await createAppWithFlags({
+      accountantEval: "1",
+      answerPolicy: "1",
+      stage2Eval: "1",
+      problemUnits: "1",
+      problemCentric: "1"
+    });
+    const onResponse = await request(appOn).post("/api/eval/run").send({
+      eval_target: "assistant_stage2",
+      useMock: true,
+      mode: "single-pass-strict",
+      caseSetFile: "assistant_stage2_canonical_v0_1.json",
+      normalizeConfig: {
+        promptVersion: "normalizer_v2_0_2"
+      }
+    });
+    expect(onResponse.status).toBe(200);
+    expect(onResponse.body.report?.eval_target).toBe("assistant_stage2");
+  });
+});

llm_normalizer/backend/tests/problemUnitAssembler.test.ts

@@ -0,0 +1,190 @@
+import { describe, expect, it } from "vitest";
+import type { EvidenceItem } from "../src/types/stage1Contracts";
+import type { ProblemUnit } from "../src/types/stage2ProblemUnits";
+import {
+  assembleProblemUnits,
+  buildCandidateEvidence,
+  clusterCandidateEvidence,
+  collapseDuplicates,
+  detectProblemUnitType
+} from "../src/services/problemUnitAssembler";
+
+function buildEvidence(input: {
+  evidenceId: string;
+  sourceId: string;
+  payload?: Record<string, unknown>;
+  confidence?: "high" | "medium" | "low";
+}): EvidenceItem {
+  const payload = input.payload ?? {};
+  return {
+    evidence_id: input.evidenceId,
+    claim_ref: "requirement:R1",
+    source_type: "retrieval_item",
+    source_ref: {
+      schema_version: "evidence_source_ref_v1",
+      namespace: "snapshot_2020",
+      entity: "Document",
+      id: input.sourceId,
+      period: "2020-06",
+      canonical_ref: `evidence_source_ref_v1|snapshot_2020|document|${input.sourceId.toLowerCase()}|2020-06`
+    },
+    pointer: {
+      fragment_id: "F1",
+      route: "hybrid_store_plus_live",
+      source: {
+        namespace: "snapshot_2020",
+        entity: "Document",
+        id: input.sourceId,
+        period: "2020-06"
+      },
+      locator: {
+        field_path: "risk_score",
+        item_index: 0
+      }
+    },
+    evidence_kind: "mechanism_link",
+    mechanism_note: null,
+    confidence: input.confidence ?? "medium",
+    limitation: null,
+    payload
+  };
+}
+
+describe("problemUnitAssembler scaffold", () => {
+  it("groups candidate evidence by route/source/pattern signature", () => {
+    const evidence = [
+      buildEvidence({
+        evidenceId: "ev-1",
+        sourceId: "DOC-1",
+        payload: {
+          failed_expected_edge: "statement_to_document"
+        }
+      }),
+      buildEvidence({
+        evidenceId: "ev-2",
+        sourceId: "DOC-1",
+        payload: {
+          failed_expected_edge: "statement_to_document"
+        }
+      }),
+      buildEvidence({
+        evidenceId: "ev-3",
+        sourceId: "DOC-2",
+        payload: {
+          anomaly_patterns: ["lifecycle_gap"]
+        }
+      })
+    ];
+
+    const candidates = buildCandidateEvidence(evidence, "hybrid_store_plus_live");
+    expect(candidates[0].candidate_id).toBe("cand-ev-1");
+    expect(candidates[0].relation_pattern_hits).toContain("failed_edge:statement_to_document");
+    expect(candidates[0].entity_backlinks.length).toBeGreaterThan(0);
+
+    const clusters = clusterCandidateEvidence(candidates);
+    expect(clusters.length).toBe(2);
+    expect(clusters.some((item) => item.candidates.length === 2)).toBe(true);
+  });
+
+  it("detects baseline problem unit type from anomaly hints", () => {
+    const candidates = buildCandidateEvidence(
+      [
+        buildEvidence({
+          evidenceId: "ev-lifecycle",
+          sourceId: "DOC-LC",
+          payload: {
+            anomaly_patterns: ["lifecycle_gap"]
+          }
+        })
+      ],
+      "store_feature_risk"
+    );
+    const cluster = clusterCandidateEvidence(candidates)[0];
+    expect(detectProblemUnitType(cluster)).toBe("lifecycle_anomaly_node");
+  });
+
+  it("collapses duplicate problem units by signature", () => {
+    const units: ProblemUnit[] = [
+      {
+        schema_version: "problem_unit_v0_1",
+        problem_unit_id: "pu-1",
+        problem_unit_type: "broken_chain_segment",
+        title: "broken",
+        mechanism_summary: "m1",
+        business_defect_class: "failed_edge:statement_to_document",
+        severity: { score: 0.7, grade: "high" },
+        confidence: { score: 0.6, grade: "medium" },
+        affected_entities: ["Document:DOC-1"],
+        affected_documents: ["Document:DOC-1"],
+        affected_postings: [],
+        affected_accounts: [],
+        affected_counterparties: [],
+        affected_contracts: [],
+        failed_expected_edge: "statement_to_document",
+        evidence_pack: ["cand-1"],
+        entity_backlinks: [{ entity: "Document", id: "DOC-1" }],
+        snapshot_limitations: []
+      },
+      {
+        schema_version: "problem_unit_v0_1",
+        problem_unit_id: "pu-2",
+        problem_unit_type: "broken_chain_segment",
+        title: "broken",
+        mechanism_summary: "m2",
+        business_defect_class: "failed_edge:statement_to_document",
+        severity: { score: 0.8, grade: "high" },
+        confidence: { score: 0.7, grade: "high" },
+        affected_entities: ["Document:DOC-1"],
+        affected_documents: ["Document:DOC-1"],
+        affected_postings: [],
+        affected_accounts: [],
+        affected_counterparties: [],
+        affected_contracts: [],
+        failed_expected_edge: "statement_to_document",
+        evidence_pack: ["cand-2"],
+        entity_backlinks: [{ entity: "Document", id: "DOC-1" }],
+        snapshot_limitations: []
+      }
+    ];
+
+    const collapsed = collapseDuplicates(units);
+    expect(collapsed.duplicate_collapses).toBe(1);
+    expect(collapsed.problem_units.length).toBe(1);
+    expect(collapsed.problem_units[0].evidence_pack).toEqual(["cand-1", "cand-2"]);
+    expect(collapsed.problem_units[0].severity.score).toBe(0.8);
+  });
+
+  it("assembles problem units and summary with bounded scaffold fields", () => {
+    const assembled = assembleProblemUnits({
+      route: "hybrid_store_plus_live",
+      evidence: [
+        buildEvidence({
+          evidenceId: "ev-1",
+          sourceId: "DOC-1",
+          payload: {
+            failed_expected_edge: "statement_to_document",
+            anomaly_patterns: ["period_close_risk"]
+          },
+          confidence: "high"
+        }),
+        buildEvidence({
+          evidenceId: "ev-2",
+          sourceId: "DOC-2",
+          payload: {
+            anomaly_patterns: ["settlement_tail"]
+          },
+          confidence: "low"
+        })
+      ]
+    });
+
+    expect(assembled.candidate_evidence.length).toBe(2);
+    expect(assembled.problem_units.length).toBeGreaterThan(0);
+    expect(assembled.problem_unit_summary.schema_version).toBe("problem_unit_summary_v0_1");
+    expect(assembled.problem_unit_summary.units_total).toBe(assembled.problem_units.length);
+    expect(Array.isArray(assembled.problem_unit_summary.unit_types)).toBe(true);
+    expect(typeof assembled.problem_unit_summary.severity_distribution.low).toBe("number");
+    expect(typeof assembled.problem_unit_summary.confidence_distribution.medium).toBe("number");
+    expect(typeof assembled.problem_unit_summary.duplicate_collapses).toBe("number");
+  });
+});

llm_normalizer/backend/tests/retrievalProblemUnitDualPayload.test.ts

@@ -0,0 +1,101 @@
+import { afterEach, describe, expect, it, vi } from "vitest";
+
+const PROBLEM_UNITS_FLAG = "FEATURE_ASSISTANT_PROBLEM_UNITS_V1";
+const ORIGINAL_PROBLEM_UNITS_FLAG = process.env[PROBLEM_UNITS_FLAG];
+
+function restoreFlag(): void {
+  if (ORIGINAL_PROBLEM_UNITS_FLAG === undefined) {
+    delete process.env[PROBLEM_UNITS_FLAG];
+  } else {
+    process.env[PROBLEM_UNITS_FLAG] = ORIGINAL_PROBLEM_UNITS_FLAG;
+  }
+}
+
+async function normalizeWithFlag(flagValue: "0" | "1") {
+  process.env[PROBLEM_UNITS_FLAG] = flagValue;
+  vi.resetModules();
+  const { normalizeRetrievalResult } = await import("../src/services/retrievalResultNormalizer");
+  return normalizeRetrievalResult("F1", ["R1"], "hybrid_store_plus_live", {
+    status: "ok",
+    result_type: "list",
+    items: [
+      {
+        source_entity: "Document",
+        source_id: "DOC-1",
+        risk_score: 4
+      }
+    ],
+    summary: {
+      broad_query_detected: false
+    },
+    evidence: [
+      {
+        evidence_id: "ev-1",
+        claim_ref: "requirement:R1",
+        source_type: "retrieval_item",
+        pointer: {
+          fragment_id: "F1",
+          route: "hybrid_store_plus_live",
+          source: {
+            namespace: "snapshot_2020",
+            entity: "Document",
+            id: "DOC-1",
+            period: "2020-06"
+          },
+          locator: {
+            field_path: "risk_score",
+            item_index: 0
+          }
+        },
+        failed_expected_edge: "statement_to_document",
+        anomaly_patterns: ["period_close_risk"],
+        confidence: "medium"
+      }
+    ],
+    why_included: ["test"],
+    selection_reason: ["test"],
+    risk_factors: ["test"],
+    business_interpretation: ["test"],
+    confidence: "medium",
+    limitations: [],
+    errors: []
+  });
+}
+
+describe.sequential("retrieval dual payload compatibility for problem units", () => {
+  afterEach(() => {
+    restoreFlag();
+    vi.resetModules();
+  });
+
+  it("keeps legacy payload intact when FEATURE_ASSISTANT_PROBLEM_UNITS_V1 is OFF", async () => {
+    const result = await normalizeWithFlag("0");
+    expect(Array.isArray(result.items)).toBe(true);
+    expect(result.items.length).toBe(1);
+    expect(result.raw_entities).toBeUndefined();
+    expect(result.candidate_evidence).toBeUndefined();
+    expect(result.problem_units).toBeUndefined();
+    expect(result.problem_unit_summary).toBeUndefined();
+  });
+
+  it("adds Stage 2 dual payload fields when FEATURE_ASSISTANT_PROBLEM_UNITS_V1 is ON", async () => {
+    const off = await normalizeWithFlag("0");
+    const on = await normalizeWithFlag("1");
+
+    expect(on.items).toEqual(off.items);
+    expect(on.raw_entities).toEqual(off.items);
+    expect(Array.isArray(on.candidate_evidence)).toBe(true);
+    expect(on.candidate_evidence?.length).toBeGreaterThan(0);
+    expect(Array.isArray(on.problem_units)).toBe(true);
+    expect(on.problem_units?.length).toBeGreaterThan(0);
+    expect(on.problem_unit_summary?.schema_version).toBe("problem_unit_summary_v0_1");
+    expect(on.problem_unit_summary?.units_total).toBe(on.problem_units?.length);
+    expect(on.summary.problem_units_enabled).toBe(true);
+    expect(on.summary.candidate_evidence_count).toBe(on.candidate_evidence?.length);
+    expect(on.summary.problem_units_count).toBe(on.problem_units?.length);
+    expect(on.summary.problem_unit_duplicate_collapses).toBe(on.problem_unit_summary?.duplicate_collapses);
+    expect(on.summary.problem_unit_types).toEqual(on.problem_unit_summary?.unit_types);
+    expect(on.summary.problem_unit_severity_distribution).toEqual(on.problem_unit_summary?.severity_distribution);
+    expect(on.summary.problem_unit_confidence_distribution).toEqual(on.problem_unit_summary?.confidence_distribution);
+  });
+});

llm_normalizer/backend/tests/sessionBackwardCompat.test.ts

@@ -1,5 +1,6 @@
 import { describe, expect, it } from "vitest";
 import { AssistantSessionStore } from "../src/services/assistantSessionStore";
+import { createEmptyInvestigationState } from "../src/services/investigationState";
 
 describe("assistant session backward compatibility", () => {
   it("lazy-upgrades legacy session objects without investigation_state", () => {
@@ -35,4 +36,42 @@ describe("assistant session backward compatibility", () => {
     expect(ensured.items.length).toBe(0);
     expect(ensured.investigation_state?.schema_version).toBe("investigation_state_v1");
   });
+
+  it("preserves optional stage2 problem_unit_state in session clone flow", () => {
+    const store = new AssistantSessionStore();
+    const sessionsMap = (store as unknown as { sessions: Map<string, unknown> }).sessions;
+    const sessionId = "legacy-session-3";
+    const baseState = createEmptyInvestigationState(sessionId, "2026-03-26T10:00:00.000Z");
+
+    sessionsMap.set(sessionId, {
+      session_id: sessionId,
+      updated_at: "2026-03-26T10:00:00.000Z",
+      items: [],
+      investigation_state: {
+        ...baseState,
+        status: "active",
+        problem_unit_state: {
+          active_problem_units: ["pu-1", "pu-2"],
+          resolved_problem_units: ["pu-0"],
+          problem_unit_backlinks: [
+            {
+              problem_unit_id: "pu-1",
+              entity_backlinks: [
+                {
+                  entity: "counterparty",
+                  id: "cp-1"
+                }
+              ]
+            }
+          ],
+          focus_problem_types: ["broken_chain_segment"]
+        }
+      }
+    });
+
+    const session = store.getSession(sessionId);
+    expect(session).toBeTruthy();
+    expect(session?.investigation_state?.problem_unit_state?.active_problem_units).toEqual(["pu-1", "pu-2"]);
+    expect(session?.investigation_state?.problem_unit_state?.focus_problem_types).toEqual(["broken_chain_segment"]);
+  });
 });