ARCH: стабилизировать continuity и защитить exact-ответы от discovery

2026-04-21 18:43:05 +03:00
parent b542b65b81
commit 429bd3d8ec
8 changed files with 603 additions and 18 deletions
@@ -0,0 +1,138 @@
+{
+  "schema_version": "domain_truth_harness_spec_v1",
+  "scenario_id": "address_truth_harness_phase20_continuity_stabilization",
+  "domain": "address_phase20_continuity_stabilization",
+  "title": "Phase 20 continuity stabilization replay",
+  "description": "Targeted AGENT replay for the continuity stabilization slice after assistant-stage1--I0x_DLqDb. The scenario validates that temporal tail words no longer turn into pseudo-counterparties, exact debt snapshots are not overwritten by discovery, and VAT/date follow-ups keep the previous period instead of drifting into unrelated carryover.",
+  "bindings": {},
+  "steps": [
+    {
+      "step_id": "step_01_top_client_all_time",
+      "title": "All-time top client question stays in ranking semantics and does not invent counterparty time",
+      "question": "кто у нас самый доходный клиент за все время?",
+      "required_answer_patterns_any": [
+        "(?i)клиент|контрагент",
+        "(?i)доходн|выручк|заработ"
+      ],
+      "forbidden_answer_patterns": [
+        "(?i)не найден контрагент.*время",
+        "(?i)контрагент с названием\\s+\"?время\"?"
+      ],
+      "criticality": "critical",
+      "semantic_tags": [
+        "value_flow_ranking",
+        "temporal_tail_not_entity"
+      ]
+    },
+    {
+      "step_id": "step_02_top_year_all_time",
+      "title": "Top year question stays in yearly ranking semantics and does not invent pseudo-entity year",
+      "question": "какой у нас самый доходный год?",
+      "required_answer_patterns_any": [
+        "(?i)20\\d{2}|19\\d{2}",
+        "(?i)доходн|выручк|заработ"
+      ],
+      "forbidden_answer_patterns": [
+        "(?i)не найден контрагент.*год",
+        "(?i)контрагент с названием\\s+\"?год\"?"
+      ],
+      "criticality": "critical",
+      "semantic_tags": [
+        "value_flow_ranking",
+        "year_tail_not_entity"
+      ]
+    },
+    {
+      "step_id": "step_03_receivables_as_of_may_2017",
+      "title": "Receivables snapshot for May 2017 answers directly instead of being overwritten by discovery",
+      "question": "кто нам должен денег на май 2017?",
+      "allowed_reply_types": [
+        "factual",
+        "factual_with_explanation"
+      ],
+      "required_direct_answer_patterns_any": [
+        "(?i)долж",
+        "(?i)дебитор",
+        "(?i)задолж"
+      ],
+      "forbidden_direct_answer_patterns": [
+        "(?i)partial_coverage",
+        "(?i)не удалось ответить",
+        "(?i)не найден контрагент"
+      ],
+      "criticality": "critical",
+      "semantic_tags": [
+        "receivables_snapshot",
+        "exact_not_overwritten"
+      ]
+    },
+    {
+      "step_id": "step_04_vat_for_same_period",
+      "title": "VAT payable follow-up keeps the carried May 2017 period",
+      "question": "а какой ндс мы должны примерно заплатить за этот период?",
+      "required_answer_patterns_all": [
+        "(?i)ндс",
+        "(?i)май|2017",
+        "(?i)заплат|к уплате|уплат"
+      ],
+      "forbidden_answer_patterns": [
+        "(?i)текущ(ий|его) период",
+        "(?i)не найден контрагент"
+      ],
+      "criticality": "critical",
+      "semantic_tags": [
+        "vat_followup",
+        "period_carryover"
+      ]
+    },
+    {
+      "step_id": "step_05_payables_today",
+      "title": "Today payables snapshot answers directly and keeps debt semantics",
+      "question": "мы должны комуто денег на сегодня?",
+      "allowed_reply_types": [
+        "factual",
+        "factual_with_explanation"
+      ],
+      "required_direct_answer_patterns_any": [
+        "(?i)должны",
+        "(?i)кредитор",
+        "(?i)задолж",
+        "(?i)долг к оплате|к оплате"
+      ],
+      "forbidden_direct_answer_patterns": [
+        "(?i)partial_coverage",
+        "(?i)не удалось ответить",
+        "(?i)не найден контрагент"
+      ],
+      "criticality": "critical",
+      "semantic_tags": [
+        "payables_snapshot",
+        "exact_not_overwritten"
+      ]
+    },
+    {
+      "step_id": "step_06_receivables_pronoun_followup",
+      "title": "Short follow-up a nam resolves to mirrored receivables instead of fake counterparty time",
+      "question": "а нам?",
+      "allowed_reply_types": [
+        "factual",
+        "factual_with_explanation"
+      ],
+      "required_direct_answer_patterns_any": [
+        "(?i)нам должны",
+        "(?i)дебитор",
+        "(?i)задолж"
+      ],
+      "forbidden_direct_answer_patterns": [
+        "(?i)не найден контрагент.*время",
+        "(?i)контрагент с названием\\s+\"?время\"?",
+        "(?i)partial_coverage"
+      ],
+      "criticality": "critical",
+      "semantic_tags": [
+        "pronoun_followup",
+        "garbage_anchor_forbidden"
+      ]
+    }
+  ]
+}