Доказать полезность margin-agent через честный boundary replay

scripts/domain_case_loop.py

@@ -256,6 +256,9 @@ GUARDED_INSUFFICIENCY_PRIMARY_MARKERS = (
     "\u0442\u043e\u0447\u043d\u044b\u0435",
     "\u043d\u0435 \u043f\u043e\u0434\u0442\u0432\u0435\u0440\u0436\u0434\u0435\u043d",
     "\u043d\u0435 \u043f\u043e\u0434\u0442\u0432\u0435\u0440\u0436\u0434\u0451\u043d",
+    "\u043d\u0435 \u043f\u043e\u0434\u0442\u0432\u0435\u0440\u0436\u0434\u0430",
+    "\u043d\u0435\u043b\u044c\u0437\u044f \u0447\u0435\u0441\u0442\u043d\u043e \u043f\u043e\u0434\u0442\u0432\u0435\u0440\u0434",
+    "\u043d\u0435\u043b\u044c\u0437\u044f \u043a\u043e\u0440\u0440\u0435\u043a\u0442\u043d\u043e \u043e\u043f\u0440\u0435\u0434\u0435\u043b",
 )
 GUARDED_INSUFFICIENCY_LIMITATION_MARKERS = (
     "\u043f\u0440\u0435\u0434\u0432\u0430\u0440\u0438\u0442\u0435\u043b\u044c\u043d",
@@ -265,12 +268,16 @@ GUARDED_INSUFFICIENCY_LIMITATION_MARKERS = (
     "\u043d\u0435 \u043f\u043e\u0434\u0442\u0432\u0435\u0440\u0436\u0434\u0451\u043d\u043d\u043e\u0435 \u0441\u0430\u043b\u044c\u0434\u043e",
     "\u043d\u0435 \u0434\u043e\u043a\u0430\u0437\u044b\u0432\u0430\u0435\u0442 \u043e\u0441\u0442\u0430\u0442\u043e\u043a",
     "\u043d\u0435 \u0444\u0438\u043d\u0430\u043b\u044c\u043d\u044b\u0439 \u0440\u0435\u0435\u0441\u0442\u0440",
+    "\u0433\u0440\u0430\u043d\u0438\u0446\u0430 \u043e\u0442\u0432\u0435\u0442\u0430",
+    "\u0440\u0430\u0441\u0447\u0435\u0442\u043d\u0443\u044e \u0431\u0430\u0437\u0443",
+    "\u0440\u0430\u0441\u0447\u0451\u0442\u043d\u0443\u044e \u0431\u0430\u0437\u0443",
 )
 GUARDED_INSUFFICIENCY_RESULT_MODES = {"heuristic_candidates"}
 GUARDED_INSUFFICIENCY_TRUTH_MODES = {"limited"}
 GUARDED_INSUFFICIENCY_ANSWER_SHAPES = {"limited_with_reason"}
 BUSINESS_EXPECTED_RESULT_MODES = {
     "clarification_required",
+    "honest_boundary_with_next_action",
     "limited_accounting_answer",
     "evidence_or_honest_boundary",
     "ranking_or_limited_accounting_answer",
@@ -966,6 +973,23 @@ def is_margin_profitability_step(step_output: dict[str, Any]) -> bool:
     question = str(step_output.get("question_resolved") or step_output.get("question_template") or "")
     if is_nomenclature_margin_context(step_output, question):
         return True
+    margin_context_values = [
+        str(step_output.get("scenario_id") or ""),
+        str(step_output.get("target_id") or ""),
+        str(step_output.get("fix_goal") or ""),
+        str(step_output.get("business_mismatch") or ""),
+        str(step_output.get("minimal_patch_direction") or ""),
+        *normalize_string_list(step_output.get("signals")),
+    ]
+    margin_context = " ".join(margin_context_values).casefold()
+    if (
+        "inventory_margin_ranking_for_nomenclature" in margin_context
+        or "inventory_inventory_margin_ranking_for_nomenclature" in margin_context
+        or "margin_false_source" in margin_context
+        or "payment_false_source" in margin_context
+        or ("margin" in margin_context and ("оплат" in margin_context or "банк" in margin_context))
+    ):
+        return True
     tokens = [
         str(step_output.get("expected_business_answer_contract") or ""),
         str(step_output.get("required_answer_contract") or ""),
@@ -978,6 +1002,8 @@ def derive_repair_issue_code(step_output: dict[str, Any], problem_type: str) ->
     violated = normalize_string_list(step_output.get("violated_invariants"))
     if "domain_leak_accounting_route" in violated and is_margin_profitability_step(step_output):
         return "margin_domain_leak_accounting_route"
+    if is_margin_profitability_step(step_output) and problem_type in {"route_gap", "capability_gap", "evidence_gap"}:
+        return "margin_domain_leak_accounting_route"
     for issue_code in (
         "technical_garbage_in_answer",
         "business_direct_answer_missing",
@@ -2141,6 +2167,36 @@ def is_nomenclature_margin_context(step_state: dict[str, Any], question: str) ->
     return has_subject and has_margin_signal and has_rank_signal
 
 
+def is_margin_false_source_boundary_answer(step_state: dict[str, Any], question: str, assistant_text: str) -> bool:
+    tags = set(normalize_string_list(step_state.get("semantic_tags")))
+    question_text = _review_text(question)
+    answer_text = _review_text(assistant_text)
+    has_false_source_question = (
+        "payment_false_source" in tags
+        or (
+            ("марж" in question_text or "прибыл" in question_text)
+            and ("товар" in question_text or "номенклатур" in question_text)
+            and ("оплат" in question_text or "банк" in question_text)
+        )
+    )
+    if not has_false_source_question:
+        return False
+    rejects_source = any(
+        marker in answer_text
+        for marker in (
+            "нельзя",
+            "не подтвержд",
+            "не подтвержда",
+            "не использовать",
+            "не является",
+            "не расчет",
+            "не расчёт",
+        )
+    )
+    names_correct_basis = "выруч" in answer_text and "себестоим" in answer_text
+    return rejects_source and names_correct_basis
+
+
 def build_business_first_review(step_state: dict[str, Any]) -> dict[str, Any]:
     question = str(step_state.get("question_resolved") or step_state.get("question_template") or "").strip()
     assistant_text = str(step_state.get("assistant_text") or "")
@@ -2171,11 +2227,18 @@ def build_business_first_review(step_state: dict[str, Any]) -> dict[str, Any]:
     limited_answer = _has_any_marker(assistant_text, BUSINESS_LIMITED_ANSWER_MARKERS)
     has_next_action = _has_any_marker(assistant_text, BUSINESS_NEXT_ACTION_MARKERS)
     nomenclature_margin_context = is_nomenclature_margin_context(step_state, question)
-    wrong_margin_domain_hits = (
+    raw_wrong_margin_domain_hits = (
         _marker_hits(assistant_text, NOMENCLATURE_MARGIN_WRONG_DOMAIN_ANSWER_MARKERS)
         if nomenclature_margin_context
         else []
     )
+    if raw_wrong_margin_domain_hits and is_margin_false_source_boundary_answer(step_state, question, assistant_text):
+        allowed_false_source_boundary_hits = {"банковск", "списание с расчетного", "списание с расчётного"}
+        wrong_margin_domain_hits = [
+            hit for hit in raw_wrong_margin_domain_hits if hit not in allowed_false_source_boundary_hits
+        ]
+    else:
+        wrong_margin_domain_hits = raw_wrong_margin_domain_hits
     margin_contract_hits = (
         _marker_hits(assistant_text, NOMENCLATURE_MARGIN_EXPECTED_ANSWER_MARKERS)
         if nomenclature_margin_context
@@ -2440,6 +2503,21 @@ def business_expected_result_mode_matches(expected_result_mode: str, step_state:
             and reply_type in {"partial_coverage", "factual", "factual_with_explanation"}
         )
 
+    if expected_result_mode == "honest_boundary_with_next_action":
+        business_review = step_state.get("business_first_review") if isinstance(step_state.get("business_first_review"), dict) else {}
+        return (
+            clean_business_review
+            and bool(assistant_text)
+            and bool(business_review.get("next_action_present"))
+            and (
+                truth_mode in GUARDED_INSUFFICIENCY_TRUTH_MODES
+                or answer_shape in GUARDED_INSUFFICIENCY_ANSWER_SHAPES
+                or step_state.get("balance_confirmed") is False
+                or is_margin_false_source_boundary_answer(step_state, str(step_state.get("question_resolved") or ""), assistant_text)
+            )
+            and reply_type in {"partial_coverage", "factual", "factual_with_explanation"}
+        )
+
     if expected_result_mode == "ranking_or_limited_accounting_answer":
         return (
             clean_business_review
@@ -4357,6 +4435,16 @@ def normalize_analyst_priority_repair_target(raw_target: dict[str, Any], index:
     if not root_cause_layers:
         root_cause_layers = [problem_type]
     issue_code = str(raw_target.get("issue_code") or problem_type or "other").strip()
+    issue_probe = {
+        **raw_target,
+        "scenario_id": scenario_id,
+        "target_id": f"{scenario_id}:{step_id}",
+        "problem_type": problem_type,
+        "root_cause_layers": root_cause_layers,
+        "fix_goal": fix_goal,
+    }
+    if issue_code in {"route_gap", "capability_gap", "evidence_gap"} and is_margin_profitability_step(issue_probe):
+        issue_code = "margin_domain_leak_accounting_route"
     catalog_entry = issue_catalog_entry(issue_code)
     return {
         "issue_code": issue_code,
@@ -5212,6 +5300,18 @@ def build_issue_catalog_snapshot(repair_targets: dict[str, Any], catalog: dict[s
     }
 
 
+def detector_evidence_paths_for_target(target: dict[str, Any]) -> list[str]:
+    explicit = normalize_string_list(target.get("evidence_paths"))
+    if explicit:
+        return explicit
+    refs = target.get("artifact_refs") if isinstance(target.get("artifact_refs"), dict) else {}
+    step_state_path = str(refs.get("step_state_json") or "").strip()
+    if not step_state_path:
+        return []
+    step_state = Path(step_state_path)
+    return [str(step_state.with_name("output.md")), str(step_state.with_name("turn.json"))]
+
+
 def build_detector_candidates(repair_targets: dict[str, Any], catalog: dict[str, Any] | None = None) -> dict[str, Any]:
     source = catalog if isinstance(catalog, dict) else load_issue_catalog()
     issues = source.get("issues") if isinstance(source.get("issues"), dict) else {}
@@ -5225,6 +5325,7 @@ def build_detector_candidates(repair_targets: dict[str, Any], catalog: dict[str,
         detectors = normalize_string_list(entry.get("detectors"))
         if not detectors and issue_code:
             detectors = [f"{issue_code}_detector"]
+        evidence_paths = detector_evidence_paths_for_target(target)
         for detector in detectors:
             key = (issue_code, detector)
             if key in seen:
@@ -5236,7 +5337,7 @@ def build_detector_candidates(repair_targets: dict[str, Any], catalog: dict[str,
                     "detector": detector,
                     "severity": target.get("severity"),
                     "sample_target_id": target.get("target_id"),
-                    "evidence_paths": target.get("evidence_paths") or [],
+                    "evidence_paths": evidence_paths,
                 }
             )
     return {