Усилить семантический AGENT gate для 1С

llm_normalizer/backend/dist/services/assistantMcpDiscoveryResponsePolicy.js

@@ -381,6 +381,9 @@ function hasValueFlowActionConflictWithDiscoveryTurnMeaning(input, entryPoint) {
     if (askedDomain !== "counterparty_value") {
         return false;
     }
+    if (askedAction === "net_value_flow") {
+        return true;
+    }
     if (hasExactBankOperationsAddressReply(input, entryPoint)) {
         return false;
     }
@@ -388,9 +391,6 @@ function hasValueFlowActionConflictWithDiscoveryTurnMeaning(input, entryPoint) {
     if (askedAction === "payout") {
         return detectedIntent !== "supplier_payouts_profile";
     }
-    if (askedAction === "net_value_flow") {
-        return true;
-    }
     return false;
 }
 function hasEvidenceLaneConflictWithDiscoveryTurnMeaning(input, entryPoint) {
@@ -534,6 +534,10 @@ function hasSemanticConflictWithDiscoveryTurnMeaning(input, entryPoint) {
         return false;
     }
     if (hasExactBankOperationsAddressReply(input, entryPoint)) {
+        const askedAction = toNonEmptyString(readDiscoveryTurnMeaning(entryPoint)?.asked_action_family);
+        if (askedAction === "net_value_flow") {
+            return true;
+        }
         return false;
     }
     const detectedIntent = toNonEmptyString(input.addressRuntimeMeta?.detected_intent);
@@ -644,6 +648,9 @@ function applyAssistantMcpDiscoveryResponsePolicy(input) {
     const metadataDiscoveryPriority = hasMetadataDiscoveryPriority(input, entryPoint);
     const valueFlowActionConflictWithDiscoveryTurnMeaning = hasValueFlowActionConflictWithDiscoveryTurnMeaning(input, entryPoint);
     const evidenceLaneConflictWithDiscoveryTurnMeaning = hasEvidenceLaneConflictWithDiscoveryTurnMeaning(input, entryPoint);
+    const exactBankOperationsProtectsCurrent = exactBankOperationsAddressReply &&
+        !semanticConflictWithDiscoveryTurnMeaning &&
+        !valueFlowActionConflictWithDiscoveryTurnMeaning;
     if (!entryPoint) {
         pushReason(reasonCodes, "mcp_discovery_response_policy_no_entry_point");
     }
@@ -701,7 +708,7 @@ function applyAssistantMcpDiscoveryResponsePolicy(input) {
     if (exactValueFlowReplyForBusinessOverviewDirectMoneyNeed) {
         pushReason(reasonCodes, "mcp_discovery_response_policy_keep_exact_value_flow_reply_over_business_overview_direct_money_clarification");
     }
-    if (exactBankOperationsAddressReply) {
+    if (exactBankOperationsProtectsCurrent) {
         pushReason(reasonCodes, "mcp_discovery_response_policy_keep_exact_bank_operations_address_reply");
     }
     if (inventoryMarginRankingAddressReply) {
@@ -736,7 +743,7 @@ function applyAssistantMcpDiscoveryResponsePolicy(input) {
         !runtimeMatchedExactReply &&
         !staleMetadataDiscoveryFallbackAgainstExactAddressReply &&
         !exactValueFlowReplyForBusinessOverviewDirectMoneyNeed &&
-        !exactBankOperationsAddressReply &&
+        !exactBankOperationsProtectsCurrent &&
         !inventoryMarginRankingAddressReply &&
         !(deterministicBroadBusinessEvaluationReply && candidate.candidate_status === "clarification_candidate") &&
         ALLOWED_CANDIDATE_STATUSES.has(candidate.candidate_status) &&

llm_normalizer/backend/src/services/assistantMcpDiscoveryResponsePolicy.ts

@@ -537,6 +537,9 @@ function hasValueFlowActionConflictWithDiscoveryTurnMeaning(
   if (askedDomain !== "counterparty_value") {
     return false;
   }
+  if (askedAction === "net_value_flow") {
+    return true;
+  }
   if (hasExactBankOperationsAddressReply(input, entryPoint)) {
     return false;
   }
@@ -544,9 +547,6 @@ function hasValueFlowActionConflictWithDiscoveryTurnMeaning(
   if (askedAction === "payout") {
     return detectedIntent !== "supplier_payouts_profile";
   }
-  if (askedAction === "net_value_flow") {
-    return true;
-  }
   return false;
 }
 
@@ -726,6 +726,10 @@ function hasSemanticConflictWithDiscoveryTurnMeaning(
     return false;
   }
   if (hasExactBankOperationsAddressReply(input, entryPoint)) {
+    const askedAction = toNonEmptyString(readDiscoveryTurnMeaning(entryPoint)?.asked_action_family);
+    if (askedAction === "net_value_flow") {
+      return true;
+    }
     return false;
   }
   const detectedIntent = toNonEmptyString(input.addressRuntimeMeta?.detected_intent);
@@ -870,6 +874,10 @@ export function applyAssistantMcpDiscoveryResponsePolicy(
     input,
     entryPoint
   );
+  const exactBankOperationsProtectsCurrent =
+    exactBankOperationsAddressReply &&
+    !semanticConflictWithDiscoveryTurnMeaning &&
+    !valueFlowActionConflictWithDiscoveryTurnMeaning;
 
   if (!entryPoint) {
     pushReason(reasonCodes, "mcp_discovery_response_policy_no_entry_point");
@@ -940,7 +948,7 @@ export function applyAssistantMcpDiscoveryResponsePolicy(
       "mcp_discovery_response_policy_keep_exact_value_flow_reply_over_business_overview_direct_money_clarification"
     );
   }
-  if (exactBankOperationsAddressReply) {
+  if (exactBankOperationsProtectsCurrent) {
     pushReason(reasonCodes, "mcp_discovery_response_policy_keep_exact_bank_operations_address_reply");
   }
   if (inventoryMarginRankingAddressReply) {
@@ -980,7 +988,7 @@ export function applyAssistantMcpDiscoveryResponsePolicy(
     !runtimeMatchedExactReply &&
     !staleMetadataDiscoveryFallbackAgainstExactAddressReply &&
     !exactValueFlowReplyForBusinessOverviewDirectMoneyNeed &&
-    !exactBankOperationsAddressReply &&
+    !exactBankOperationsProtectsCurrent &&
     !inventoryMarginRankingAddressReply &&
     !(deterministicBroadBusinessEvaluationReply && candidate.candidate_status === "clarification_candidate") &&
     ALLOWED_CANDIDATE_STATUSES.has(candidate.candidate_status) &&

llm_normalizer/backend/tests/assistantMcpDiscoveryResponsePolicy.test.ts

@@ -700,6 +700,67 @@ describe("assistant MCP discovery response policy", () => {
     expect(result.reason_codes).not.toContain("mcp_discovery_response_policy_semantic_conflict_allows_candidate_override");
   });
 
+  it("overrides exact bank operation replies for explicit counterparty received-paid-net questions", () => {
+    const result = applyAssistantMcpDiscoveryResponsePolicy({
+      currentReply:
+        "Exact bank operation answer: found 13 bank operations for Group SVK, showing rows and payment purposes.",
+      currentReplySource: "address_query_runtime_v1",
+      currentReplyType: "factual",
+      addressRuntimeMeta: {
+        detected_intent: "bank_operations_by_counterparty",
+        selected_recipe: "address_bank_operations_by_counterparty_v1",
+        mcp_call_status: "matched_non_empty",
+        capability_route_mode: "exact",
+        answer_grounding_check: {
+          status: "grounded"
+        },
+        assistant_mcp_discovery_entry_point_v1: entryPoint({
+          turn_input: {
+            adapter_status: "ready",
+            should_run_discovery: true,
+            data_need_graph: {
+              business_fact_family: "value_flow",
+              subject_candidates: ["Group SVK"],
+              reason_codes: ["data_need_graph_built", "data_need_graph_bidirectional_value_flow"]
+            },
+            turn_meaning_ref: {
+              asked_domain_family: "counterparty_value",
+              asked_action_family: "net_value_flow",
+              explicit_entity_candidates: ["Group SVK"],
+              explicit_date_scope: "2020",
+              unsupported_but_understood_family: "counterparty_bidirectional_value_flow_or_netting"
+            }
+          },
+          bridge: {
+            bridge_status: "answer_draft_ready",
+            user_facing_response_allowed: true,
+            business_fact_answer_allowed: true,
+            requires_user_clarification: false,
+            answer_draft: {
+              answer_mode: "confirmed_with_bounded_inference",
+              headline: "Group SVK 2020: received 12 093 465 rub., paid 0 rub.; net 12 093 465 rub. toward us.",
+              confirmed_lines: ["received 12 093 465 rub.; paid 0 rub.; net 12 093 465 rub."],
+              inference_lines: [],
+              unknown_lines: [],
+              limitation_lines: [],
+              next_step_line: null
+            }
+          }
+        })
+      }
+    });
+
+    expect(result.applied).toBe(true);
+    expect(result.decision).toBe("apply_candidate");
+    expect(result.reply_source).toBe("mcp_discovery_response_candidate_guarded");
+    expect(result.reply_text).toContain("received 12 093 465");
+    expect(result.reply_text).toContain("paid 0");
+    expect(result.reply_text).toContain("net 12 093 465");
+    expect(result.reason_codes).toContain("mcp_discovery_response_policy_value_flow_action_conflict_allows_candidate_override");
+    expect(result.reason_codes).toContain("mcp_discovery_response_policy_semantic_conflict_allows_candidate_override");
+    expect(result.reason_codes).not.toContain("mcp_discovery_response_policy_keep_exact_bank_operations_address_reply");
+  });
+
   it("overrides an exact ranking-shaped address reply when open-scope ranking still needs organization", () => {
     const result = applyAssistantMcpDiscoveryResponsePolicy({
       currentReply:

llm_normalizer/data/autorun_generators/history.json

@@ -1,4 +1,84 @@
 [
+  {
+    "generation_id": "gen-ag05231858-323f86",
+    "created_at": "2026-05-23T18:58:48+00:00",
+    "mode": "saved_user_sessions",
+    "title": "AGENT | Autonomy business quality pack",
+    "count": 15,
+    "domain": "autonomy_business_quality",
+    "questions": [
+      "Сколько входящих денег за 2020 год по ООО Альтернатива Плюс без разреза по контрагентам?",
+      "Сколько исходящих денег за 2020 год по ООО Альтернатива Плюс без разреза по контрагентам?",
+      "А всего сколько денег пришло в ООО Альтернатива Плюс за 2020, без топов и без контрагентов?",
+      "Теперь дай взрослый обзор за 2020 по компании: входящие, исходящие, нетто, топы, но банк в топах отдельно объясни как финансовый поток.",
+      "скока денег альтернатива заработала за 20 год?",
+      "а это чистая прибыль?",
+      "Какая чистая прибыль по ООО Альтернатива Плюс за 2020?",
+      "А отдельно по СБЕРБАНКУ: он для нас клиент, поставщик или финансовый поток? Дай коротко по подтвержденным строкам.",
+      "какое нетто по деньгам с Группа СВК за 2020 год: сколько получили и сколько заплатили?",
+      "кому мы должны на конец 2020?",
+      "а нам кто должен на конец 2020?",
+      "сколько НДС надо заплатить в налоговую за декабрь 2019?",
+      "Как ты оценишь деятельность компании?",
+      "Какая номенклатура товара реализована с высокой прибылью какая с низкой?",
+      "май 2020"
+    ],
+    "generated_by": "codex_agent",
+    "saved_case_set_file": "assistant_autogen_saved_user_sessions_20260523185848_gen-ag05231858-323f86.json",
+    "context": {
+      "llm_provider": null,
+      "model": null,
+      "assistant_prompt_version": null,
+      "decomposition_prompt_version": null,
+      "prompt_fingerprint": null,
+      "autogen_personality_id": null,
+      "autogen_personality_prompt": null,
+      "source_session_id": null,
+      "saved_session_file": "assistant_saved_session_20260523185848_gen-ag05231858-323f86.json",
+      "saved_case_set_kind": "agent_semantic_scenario",
+      "agent_run": true,
+      "agent_focus": "Expanded targeted AGENT replay for the autonomy milestone: value-flow, business overview, debts, VAT, profit/cashflow distinction, nomenclature margin boundary, and final answer quality must survive realistic business questions.",
+      "architecture_phase": "turnaround_11",
+      "source_spec_file": "X:\\1C\\NDC_1C\\docs\\orchestration\\agent_autonomy_business_quality_20260523.json",
+      "scenario_id": "agent_autonomy_business_quality_20260523",
+      "semantic_tags": [
+        "autonomy_core",
+        "bank_boundary",
+        "bank_classification",
+        "business_answer_quality",
+        "business_evaluation",
+        "business_overview",
+        "cashflow_overview",
+        "cashflow_vs_profit",
+        "clarification",
+        "colloquial_money",
+        "colloquial_total",
+        "counterparty_value_flow",
+        "debt_answer_quality",
+        "direct_profit",
+        "domain_purity",
+        "followup_context",
+        "incoming_total",
+        "limit_honesty",
+        "limited_answer",
+        "net_flow",
+        "next_action",
+        "no_profit_substitution",
+        "no_top_guard",
+        "nomenclature_margin",
+        "outgoing_total",
+        "payables",
+        "profit_vs_cashflow",
+        "receivables",
+        "technical_garbage_guard",
+        "value_flow",
+        "vat"
+      ],
+      "validation_status": "accepted_live_replay",
+      "validated_run_dir": "artifacts\\domain_runs\\agent_autonomy_business_quality_live_semantic_gate_accepted_20260523",
+      "saved_after_validated_replay": true
+    }
+  },
   {
     "generation_id": "gen-ag05231427-70915a",
     "created_at": "2026-05-23T14:27:55+00:00",

llm_normalizer/data/autorun_generators/saved_sessions/assistant_saved_session_20260523185848_gen-ag05231858-323f86.json

@@ -0,0 +1,287 @@
+{
+  "saved_at": "2026-05-23T18:58:48+00:00",
+  "generation_id": "gen-ag05231858-323f86",
+  "mode": "saved_user_sessions",
+  "title": "AGENT | Autonomy business quality pack",
+  "agent_run": true,
+  "questions": [
+    "Сколько входящих денег за 2020 год по ООО Альтернатива Плюс без разреза по контрагентам?",
+    "Сколько исходящих денег за 2020 год по ООО Альтернатива Плюс без разреза по контрагентам?",
+    "А всего сколько денег пришло в ООО Альтернатива Плюс за 2020, без топов и без контрагентов?",
+    "Теперь дай взрослый обзор за 2020 по компании: входящие, исходящие, нетто, топы, но банк в топах отдельно объясни как финансовый поток.",
+    "скока денег альтернатива заработала за 20 год?",
+    "а это чистая прибыль?",
+    "Какая чистая прибыль по ООО Альтернатива Плюс за 2020?",
+    "А отдельно по СБЕРБАНКУ: он для нас клиент, поставщик или финансовый поток? Дай коротко по подтвержденным строкам.",
+    "какое нетто по деньгам с Группа СВК за 2020 год: сколько получили и сколько заплатили?",
+    "кому мы должны на конец 2020?",
+    "а нам кто должен на конец 2020?",
+    "сколько НДС надо заплатить в налоговую за декабрь 2019?",
+    "Как ты оценишь деятельность компании?",
+    "Какая номенклатура товара реализована с высокой прибылью какая с низкой?",
+    "май 2020"
+  ],
+  "metadata": {
+    "assistant_prompt_version": null,
+    "decomposition_prompt_version": null,
+    "prompt_fingerprint": null,
+    "agent_focus": "Expanded targeted AGENT replay for the autonomy milestone: value-flow, business overview, debts, VAT, profit/cashflow distinction, nomenclature margin boundary, and final answer quality must survive realistic business questions.",
+    "architecture_phase": "turnaround_11",
+    "source_spec_file": "X:\\1C\\NDC_1C\\docs\\orchestration\\agent_autonomy_business_quality_20260523.json",
+    "scenario_id": "agent_autonomy_business_quality_20260523",
+    "semantic_tags": [
+      "autonomy_core",
+      "bank_boundary",
+      "bank_classification",
+      "business_answer_quality",
+      "business_evaluation",
+      "business_overview",
+      "cashflow_overview",
+      "cashflow_vs_profit",
+      "clarification",
+      "colloquial_money",
+      "colloquial_total",
+      "counterparty_value_flow",
+      "debt_answer_quality",
+      "direct_profit",
+      "domain_purity",
+      "followup_context",
+      "incoming_total",
+      "limit_honesty",
+      "limited_answer",
+      "net_flow",
+      "next_action",
+      "no_profit_substitution",
+      "no_top_guard",
+      "nomenclature_margin",
+      "outgoing_total",
+      "payables",
+      "profit_vs_cashflow",
+      "receivables",
+      "technical_garbage_guard",
+      "value_flow",
+      "vat"
+    ],
+    "validation_status": "accepted_live_replay",
+    "validated_run_dir": "artifacts\\domain_runs\\agent_autonomy_business_quality_live_semantic_gate_accepted_20260523",
+    "saved_after_validated_replay": true,
+    "save_gate": {
+      "schema_version": "agent_semantic_save_gate_v1",
+      "validation_status": "accepted_live_replay",
+      "validated_run_dir": "artifacts\\domain_runs\\agent_autonomy_business_quality_live_semantic_gate_accepted_20260523",
+      "final_status": "accepted",
+      "review_overall_status": "pass",
+      "business_overall_status": "pass",
+      "steps_total": 15,
+      "steps_passed": 15,
+      "steps_failed": 0,
+      "steps_with_business_failures": 0,
+      "steps_with_business_warnings": 0,
+      "acceptance_gate_passed": true,
+      "saved_after_validated_replay": true
+    }
+  },
+  "source_session_id": null,
+  "session": {
+    "session_id": null,
+    "mode": "agent_semantic_run",
+    "items": [
+      {
+        "message_id": "agent-user-001",
+        "role": "user",
+        "text": "Сколько входящих денег за 2020 год по ООО Альтернатива Плюс без разреза по контрагентам?",
+        "created_at": "2026-05-23T18:58:48+00:00",
+        "reply_type": null,
+        "trace_id": null,
+        "debug": null
+      },
+      {
+        "message_id": "agent-user-002",
+        "role": "user",
+        "text": "Сколько исходящих денег за 2020 год по ООО Альтернатива Плюс без разреза по контрагентам?",
+        "created_at": "2026-05-23T18:58:48+00:00",
+        "reply_type": null,
+        "trace_id": null,
+        "debug": null
+      },
+      {
+        "message_id": "agent-user-003",
+        "role": "user",
+        "text": "А всего сколько денег пришло в ООО Альтернатива Плюс за 2020, без топов и без контрагентов?",
+        "created_at": "2026-05-23T18:58:48+00:00",
+        "reply_type": null,
+        "trace_id": null,
+        "debug": null
+      },
+      {
+        "message_id": "agent-user-004",
+        "role": "user",
+        "text": "Теперь дай взрослый обзор за 2020 по компании: входящие, исходящие, нетто, топы, но банк в топах отдельно объясни как финансовый поток.",
+        "created_at": "2026-05-23T18:58:48+00:00",
+        "reply_type": null,
+        "trace_id": null,
+        "debug": null
+      },
+      {
+        "message_id": "agent-user-005",
+        "role": "user",
+        "text": "скока денег альтернатива заработала за 20 год?",
+        "created_at": "2026-05-23T18:58:48+00:00",
+        "reply_type": null,
+        "trace_id": null,
+        "debug": null
+      },
+      {
+        "message_id": "agent-user-006",
+        "role": "user",
+        "text": "а это чистая прибыль?",
+        "created_at": "2026-05-23T18:58:48+00:00",
+        "reply_type": null,
+        "trace_id": null,
+        "debug": null
+      },
+      {
+        "message_id": "agent-user-007",
+        "role": "user",
+        "text": "Какая чистая прибыль по ООО Альтернатива Плюс за 2020?",
+        "created_at": "2026-05-23T18:58:48+00:00",
+        "reply_type": null,
+        "trace_id": null,
+        "debug": null
+      },
+      {
+        "message_id": "agent-user-008",
+        "role": "user",
+        "text": "А отдельно по СБЕРБАНКУ: он для нас клиент, поставщик или финансовый поток? Дай коротко по подтвержденным строкам.",
+        "created_at": "2026-05-23T18:58:48+00:00",
+        "reply_type": null,
+        "trace_id": null,
+        "debug": null
+      },
+      {
+        "message_id": "agent-user-009",
+        "role": "user",
+        "text": "какое нетто по деньгам с Группа СВК за 2020 год: сколько получили и сколько заплатили?",
+        "created_at": "2026-05-23T18:58:48+00:00",
+        "reply_type": null,
+        "trace_id": null,
+        "debug": null
+      },
+      {
+        "message_id": "agent-user-010",
+        "role": "user",
+        "text": "кому мы должны на конец 2020?",
+        "created_at": "2026-05-23T18:58:48+00:00",
+        "reply_type": null,
+        "trace_id": null,
+        "debug": null
+      },
+      {
+        "message_id": "agent-user-011",
+        "role": "user",
+        "text": "а нам кто должен на конец 2020?",
+        "created_at": "2026-05-23T18:58:48+00:00",
+        "reply_type": null,
+        "trace_id": null,
+        "debug": null
+      },
+      {
+        "message_id": "agent-user-012",
+        "role": "user",
+        "text": "сколько НДС надо заплатить в налоговую за декабрь 2019?",
+        "created_at": "2026-05-23T18:58:48+00:00",
+        "reply_type": null,
+        "trace_id": null,
+        "debug": null
+      },
+      {
+        "message_id": "agent-user-013",
+        "role": "user",
+        "text": "Как ты оценишь деятельность компании?",
+        "created_at": "2026-05-23T18:58:48+00:00",
+        "reply_type": null,
+        "trace_id": null,
+        "debug": null
+      },
+      {
+        "message_id": "agent-user-014",
+        "role": "user",
+        "text": "Какая номенклатура товара реализована с высокой прибылью какая с низкой?",
+        "created_at": "2026-05-23T18:58:48+00:00",
+        "reply_type": null,
+        "trace_id": null,
+        "debug": null
+      },
+      {
+        "message_id": "agent-user-015",
+        "role": "user",
+        "text": "май 2020",
+        "created_at": "2026-05-23T18:58:48+00:00",
+        "reply_type": null,
+        "trace_id": null,
+        "debug": null
+      }
+    ],
+    "agent_run": true,
+    "metadata": {
+      "assistant_prompt_version": null,
+      "decomposition_prompt_version": null,
+      "prompt_fingerprint": null,
+      "agent_focus": "Expanded targeted AGENT replay for the autonomy milestone: value-flow, business overview, debts, VAT, profit/cashflow distinction, nomenclature margin boundary, and final answer quality must survive realistic business questions.",
+      "architecture_phase": "turnaround_11",
+      "source_spec_file": "X:\\1C\\NDC_1C\\docs\\orchestration\\agent_autonomy_business_quality_20260523.json",
+      "scenario_id": "agent_autonomy_business_quality_20260523",
+      "semantic_tags": [
+        "autonomy_core",
+        "bank_boundary",
+        "bank_classification",
+        "business_answer_quality",
+        "business_evaluation",
+        "business_overview",
+        "cashflow_overview",
+        "cashflow_vs_profit",
+        "clarification",
+        "colloquial_money",
+        "colloquial_total",
+        "counterparty_value_flow",
+        "debt_answer_quality",
+        "direct_profit",
+        "domain_purity",
+        "followup_context",
+        "incoming_total",
+        "limit_honesty",
+        "limited_answer",
+        "net_flow",
+        "next_action",
+        "no_profit_substitution",
+        "no_top_guard",
+        "nomenclature_margin",
+        "outgoing_total",
+        "payables",
+        "profit_vs_cashflow",
+        "receivables",
+        "technical_garbage_guard",
+        "value_flow",
+        "vat"
+      ],
+      "validation_status": "accepted_live_replay",
+      "validated_run_dir": "artifacts\\domain_runs\\agent_autonomy_business_quality_live_semantic_gate_accepted_20260523",
+      "saved_after_validated_replay": true,
+      "save_gate": {
+        "schema_version": "agent_semantic_save_gate_v1",
+        "validation_status": "accepted_live_replay",
+        "validated_run_dir": "artifacts\\domain_runs\\agent_autonomy_business_quality_live_semantic_gate_accepted_20260523",
+        "final_status": "accepted",
+        "review_overall_status": "pass",
+        "business_overall_status": "pass",
+        "steps_total": 15,
+        "steps_passed": 15,
+        "steps_failed": 0,
+        "steps_with_business_failures": 0,
+        "steps_with_business_warnings": 0,
+        "acceptance_gate_passed": true,
+        "saved_after_validated_replay": true
+      }
+    }
+  }
+}

llm_normalizer/data/eval_cases/assistant_autogen_saved_user_sessions_20260523185848_gen-ag05231858-323f86.json

@@ -0,0 +1,70 @@
+{
+  "suite_id": "assistant_saved_session_gen-ag05231858-323f86",
+  "suite_version": "0.1.0",
+  "schema_version": "assistant_saved_session_suite_v0_1",
+  "generated_at": "2026-05-23T18:58:48+00:00",
+  "generation_id": "gen-ag05231858-323f86",
+  "mode": "saved_user_sessions",
+  "title": "AGENT | Autonomy business quality pack",
+  "domain": "autonomy_business_quality",
+  "scenario_count": 1,
+  "case_ids": [
+    "SAVED-001"
+  ],
+  "cases": [
+    {
+      "case_id": "SAVED-001",
+      "scenario_tag": "agent_saved_user_sessions",
+      "title": "AGENT | Autonomy business quality pack",
+      "question_type": "followup",
+      "broadness_level": "medium",
+      "turns": [
+        {
+          "user_message": "Сколько входящих денег за 2020 год по ООО Альтернатива Плюс без разреза по контрагентам?"
+        },
+        {
+          "user_message": "Сколько исходящих денег за 2020 год по ООО Альтернатива Плюс без разреза по контрагентам?"
+        },
+        {
+          "user_message": "А всего сколько денег пришло в ООО Альтернатива Плюс за 2020, без топов и без контрагентов?"
+        },
+        {
+          "user_message": "Теперь дай взрослый обзор за 2020 по компании: входящие, исходящие, нетто, топы, но банк в топах отдельно объясни как финансовый поток."
+        },
+        {
+          "user_message": "скока денег альтернатива заработала за 20 год?"
+        },
+        {
+          "user_message": "а это чистая прибыль?"
+        },
+        {
+          "user_message": "Какая чистая прибыль по ООО Альтернатива Плюс за 2020?"
+        },
+        {
+          "user_message": "А отдельно по СБЕРБАНКУ: он для нас клиент, поставщик или финансовый поток? Дай коротко по подтвержденным строкам."
+        },
+        {
+          "user_message": "какое нетто по деньгам с Группа СВК за 2020 год: сколько получили и сколько заплатили?"
+        },
+        {
+          "user_message": "кому мы должны на конец 2020?"
+        },
+        {
+          "user_message": "а нам кто должен на конец 2020?"
+        },
+        {
+          "user_message": "сколько НДС надо заплатить в налоговую за декабрь 2019?"
+        },
+        {
+          "user_message": "Как ты оценишь деятельность компании?"
+        },
+        {
+          "user_message": "Какая номенклатура товара реализована с высокой прибылью какая с низкой?"
+        },
+        {
+          "user_message": "май 2020"
+        }
+      ]
+    }
+  ]
+}

scripts/domain_case_loop.py

@@ -118,6 +118,9 @@ BUSINESS_TOP_LINE_SCAFFOLD_MARKERS = (
     "\u043e\u0433\u0440\u0430\u043d\u0438\u0447\u0435\u043d\u043d\u044b\u0439 \u0431\u0438\u0437\u043d\u0435\u0441-\u043e\u0431\u0437\u043e\u0440",
     "\u0447\u0442\u043e \u043f\u043e\u0434\u0442\u0432\u0435\u0440\u0436\u0434\u0435\u043d\u043e",
     "\u043f\u0440\u043e\u0432\u0435\u0440\u0435\u043d\u043d\u044b\u0435 \u043a\u043e\u043d\u0442\u0443\u0440\u044b",
+    "\u043f\u043e \u0434\u0430\u043d\u043d\u044b\u043c 1\u0441 \u043d\u0430\u0439\u0434\u0435\u043d\u044b",
+    "\u043d\u0430\u0439\u0434\u0435\u043d\u044b \u0441\u0442\u0440\u043e\u043a\u0438",
+    "\u0441\u0443\u043c\u043c\u0443 \u043c\u043e\u0436\u043d\u043e \u043d\u0430\u0437\u044b\u0432\u0430\u0442\u044c \u0442\u043e\u043b\u044c\u043a\u043e",
     "\u0431\u043b\u043e\u043a 1",
     "\u0441\u0442\u0430\u0442\u0443\u0441",
 )

scripts/domain_truth_harness.py

@@ -1156,8 +1156,15 @@ def build_business_review_summary(spec: dict[str, Any], scenario_state: dict[str
     step_outputs = scenario_state.get("step_outputs") if isinstance(scenario_state.get("step_outputs"), dict) else {}
     steps: list[dict[str, Any]] = []
     issue_counts: dict[str, int] = {}
+    review_failures = 0
+    review_warnings = 0
     for index, step in enumerate(spec["steps"], start=1):
         step_state = step_outputs.get(step["step_id"], {})
+        review_status = step_state.get("review_status") if isinstance(step_state, dict) else None
+        if review_status == "fail":
+            review_failures += 1
+        elif review_status == "warning":
+            review_warnings += 1
         business_review = (
             step_state.get("business_first_review")
             if isinstance(step_state, dict) and isinstance(step_state.get("business_first_review"), dict)
@@ -1171,7 +1178,7 @@ def build_business_review_summary(spec: dict[str, Any], scenario_state: dict[str
                 "index": index,
                 "step_id": step["step_id"],
                 "question": step["question_template"],
-                "review_status": step_state.get("review_status") if isinstance(step_state, dict) else None,
+                "review_status": review_status,
                 "direct_answer": business_review.get("actual_direct_answer"),
                 "answer_length_chars": business_review.get("answer_length_chars"),
                 "direct_answer_required": business_review.get("direct_answer_required"),
@@ -1193,6 +1200,7 @@ def build_business_review_summary(spec: dict[str, Any], scenario_state: dict[str
         )
     )
     warnings = sum(1 for step in steps if "business_answer_too_verbose" in step["issue_codes"])
+    semantic_status = "fail" if failed or review_failures else ("warning" if warnings or review_warnings else "pass")
     return {
         "schema_version": "business_first_run_review_v1",
         "scenario_id": spec["scenario_id"],
@@ -1202,24 +1210,32 @@ def build_business_review_summary(spec: dict[str, Any], scenario_state: dict[str
         "steps_total": len(steps),
         "steps_with_business_failures": failed,
         "steps_with_business_warnings": warnings,
+        "steps_with_review_failures": review_failures,
+        "steps_with_review_warnings": review_warnings,
         "issue_counts": issue_counts,
         "overall_business_status": "fail" if failed else ("warning" if warnings else "pass"),
+        "overall_semantic_status": semantic_status,
+        "semantic_gate_passed": semantic_status == "pass",
         "steps": steps,
     }
 
 
-def build_business_review_markdown(business_review: dict[str, Any]) -> str:
+def build_business_review_markdown(business_review: dict[str, Any], *, title: str = "Business-first review") -> str:
     lines = [
-        "# Business-first review",
+        f"# {title}",
         "",
         f"- scenario_id: `{business_review.get('scenario_id') or 'n/a'}`",
         f"- domain: `{business_review.get('domain') or 'n/a'}`",
         f"- title: {business_review.get('title') or 'n/a'}",
         f"- session_id: `{business_review.get('session_id') or 'n/a'}`",
         f"- overall_business_status: `{business_review.get('overall_business_status') or 'n/a'}`",
+        f"- overall_semantic_status: `{business_review.get('overall_semantic_status') or 'n/a'}`",
+        f"- semantic_gate_passed: `{business_review.get('semantic_gate_passed') is True}`",
         f"- steps_total: `{business_review.get('steps_total')}`",
         f"- steps_with_business_failures: `{business_review.get('steps_with_business_failures')}`",
         f"- steps_with_business_warnings: `{business_review.get('steps_with_business_warnings')}`",
+        f"- steps_with_review_failures: `{business_review.get('steps_with_review_failures')}`",
+        f"- steps_with_review_warnings: `{business_review.get('steps_with_review_warnings')}`",
         f"- issue_counts: `{dump_json(business_review.get('issue_counts') or {})}`",
         "",
         "## Human Answer Surface",
@@ -1479,6 +1495,8 @@ def review_export(spec: dict[str, Any], export_path: Path, output_dir: Path) ->
     write_text(output_dir / "truth_review.md", review_markdown)
     write_json(output_dir / "business_review.json", business_review)
     write_text(output_dir / "business_review.md", build_business_review_markdown(business_review))
+    write_json(output_dir / "semantic_audit.json", business_review)
+    write_text(output_dir / "semantic_audit.md", build_business_review_markdown(business_review, title="Semantic audit"))
     acceptance_bundle = write_acceptance_artifacts(output_dir, spec, scenario_state, review_summary)
     return {
         "scenario_state": scenario_state,
@@ -1575,6 +1593,8 @@ def run_live(spec: dict[str, Any], output_dir: Path, args: argparse.Namespace) -
     write_text(output_dir / "truth_review.md", review_markdown)
     write_json(output_dir / "business_review.json", business_review)
     write_text(output_dir / "business_review.md", build_business_review_markdown(business_review))
+    write_json(output_dir / "semantic_audit.json", business_review)
+    write_text(output_dir / "semantic_audit.md", build_business_review_markdown(business_review, title="Semantic audit"))
     acceptance_bundle = write_acceptance_artifacts(output_dir, spec, scenario_state, review_summary)
     print(f"[truth-harness] saved artifacts to {output_dir}")
     print(f"[truth-harness] overall_status={review_summary['overall_status']}")

scripts/test_domain_case_loop_step_state.py

@@ -927,6 +927,91 @@ class DomainCaseLoopStepStateTests(unittest.TestCase):
         self.assertIn("technical_garbage_in_answer", review["issue_codes"])
         self.assertNotIn("business_direct_answer_missing", review["issue_codes"])
 
+    def test_business_first_review_rejects_found_rows_scaffold_as_direct_answer(self) -> None:
+        question = "\u0441\u043a\u043e\u043b\u044c\u043a\u043e \u0432\u0445\u043e\u0434\u044f\u0449\u0438\u0445 \u0434\u0435\u043d\u0435\u0433 \u0437\u0430 2020"
+        step_state = dcl.build_scenario_step_state(
+            scenario_id="semantic_gate_demo",
+            domain="value_flow",
+            step={
+                "step_id": "step_01",
+                "title": "Direct money answer",
+                "depends_on": [],
+                "question_template": question,
+            },
+            step_index=1,
+            question_resolved=question,
+            analysis_context={},
+            turn_artifact={
+                "assistant_message": {
+                    "reply_type": "factual",
+                    "text": "\u041a\u043e\u0440\u043e\u0442\u043a\u043e: \u041f\u043e \u0434\u0430\u043d\u043d\u044b\u043c 1\u0421 \u043d\u0430\u0439\u0434\u0435\u043d\u044b \u0441\u0442\u0440\u043e\u043a\u0438 \u0432\u0445\u043e\u0434\u044f\u0449\u0438\u0445 \u0434\u0435\u043d\u0435\u0436\u043d\u044b\u0445 \u043f\u043e\u0441\u0442\u0443\u043f\u043b\u0435\u043d\u0438\u0439; \u0441\u0443\u043c\u043c\u0443 \u043c\u043e\u0436\u043d\u043e \u043d\u0430\u0437\u044b\u0432\u0430\u0442\u044c \u0442\u043e\u043b\u044c\u043a\u043e \u0432 \u0440\u0430\u043c\u043a\u0430\u0445 \u043f\u0440\u043e\u0432\u0435\u0440\u043a\u0438.\n\u0427\u0442\u043e \u043f\u043e\u0434\u0442\u0432\u0435\u0440\u0436\u0434\u0435\u043d\u043e: 47 628 853,03 \u0440\u0443\u0431.",
+                    "message_id": "msg-1",
+                    "trace_id": "trace-1",
+                },
+                "technical_debug_payload": {},
+                "session_summary": {},
+            },
+            entries=[],
+        )
+
+        review = step_state["business_first_review"]
+        self.assertFalse(review["direct_answer_first_ok"])
+        self.assertFalse(review["answer_layering_ok"])
+        self.assertIn("business_direct_answer_missing", review["issue_codes"])
+        self.assertIn("answer_layering_noise", review["issue_codes"])
+
+    def test_business_first_review_allows_direct_answer_with_evidence_boundary(self) -> None:
+        question = "\u043a\u0430\u043a\u043e\u0435 \u043d\u0435\u0442\u0442\u043e \u043f\u043e \u0434\u0435\u043d\u044c\u0433\u0430\u043c \u0441 \u0413\u0440\u0443\u043f\u043f\u0430 \u0421\u0412\u041a \u0437\u0430 2020"
+        step_state = dcl.build_scenario_step_state(
+            scenario_id="semantic_gate_demo",
+            domain="value_flow",
+            step={
+                "step_id": "step_01",
+                "title": "Counterparty net flow",
+                "depends_on": [],
+                "question_template": question,
+            },
+            step_index=1,
+            question_resolved=question,
+            analysis_context={},
+            turn_artifact={
+                "assistant_message": {
+                    "reply_type": "factual_with_explanation",
+                    "text": "\u041a\u043e\u0440\u043e\u0442\u043a\u043e: \u043f\u043e \u043a\u043e\u043d\u0442\u0440\u0430\u0433\u0435\u043d\u0442\u0443 \u0413\u0440\u0443\u043f\u043f\u0430 \u0421\u0412\u041a \u0437\u0430 2020 \u043f\u043e \u043d\u0430\u0439\u0434\u0435\u043d\u043d\u044b\u043c \u0441\u0442\u0440\u043e\u043a\u0430\u043c 1\u0421 \u043f\u043e\u043b\u0443\u0447\u0438\u043b\u0438 12 093 465 \u0440\u0443\u0431., \u0437\u0430\u043f\u043b\u0430\u0442\u0438\u043b\u0438 0 \u0440\u0443\u0431.; \u043d\u0435\u0442\u0442\u043e \u0432 \u043d\u0430\u0448\u0443 \u0441\u0442\u043e\u0440\u043e\u043d\u0443 12 093 465 \u0440\u0443\u0431.",
+                    "message_id": "msg-1",
+                    "trace_id": "trace-1",
+                },
+                "technical_debug_payload": {},
+                "session_summary": {},
+            },
+            entries=[],
+        )
+
+        review = step_state["business_first_review"]
+        self.assertTrue(review["direct_answer_first_ok"])
+        self.assertTrue(review["answer_layering_ok"])
+        self.assertEqual(review["issue_codes"], [])
+
+    def test_semantic_audit_markdown_exposes_semantic_gate(self) -> None:
+        markdown = dth.build_business_review_markdown(
+            {
+                "scenario_id": "semantic_gate_demo",
+                "domain": "value_flow",
+                "title": "Semantic gate demo",
+                "session_id": "session-1",
+                "overall_business_status": "fail",
+                "steps_total": 1,
+                "steps_with_business_failures": 1,
+                "steps_with_business_warnings": 0,
+                "issue_counts": {"answer_layering_noise": 1},
+                "steps": [],
+            },
+            title="Semantic audit",
+        )
+
+        self.assertIn("# Semantic audit", markdown)
+        self.assertIn("semantic_gate_passed: `False`", markdown)
+
     def test_truth_harness_promotes_business_review_issues_to_findings(self) -> None:
         step_state = dcl.build_scenario_step_state(
             scenario_id="business_surface_demo",