Stage 4 / Wave 18 закрыт blocker-pack по временному якорю, полярности домена и допустимости evidence

docs/accounting-assistant/accounting-assistant/03_execution/STAGE_04_WAVE_19_TASK_CARD_2026-03-29.md

@@ -0,0 +1,61 @@
+# Stage 4 / Wave 19 Task Card (2026-03-29)
+
+## Goal
+Build a claim-bound evidence acquisition path for Stage 4 P0 domains so the assistant can produce grounded-positive answers when admissible evidence exists, while preserving honest limited mode when evidence is insufficient.
+
+## In Scope
+- `settlements_60_62`
+- `vat_document_register_book`
+- `month_close_costs_20_44`
+- runtime: anchor extraction, targeted acquisition, evidence packaging, admissibility handoff, grounded eligibility, debug/export traces, reruns
+
+## Out of Scope
+- new domains
+- orchestration redesign
+- Stage 5 investigation mode as core path
+- ontology/schema expansion
+
+## Subwaves
+1. Subwave A — Primary period + controlled temporal expansion
+- keep `primary_period` immutable
+- allow out-of-primary evidence only with linked context + explicit reason
+- expose `allowed_context_window` and expansion reasons in debug
+
+2. Subwave B — Claim-bound anchor extraction
+- introduce `claim_type`, `required_anchors`, `resolved_anchors`, `missing_anchors`
+- compute and expose `claim_anchor_resolution_rate`
+
+3. Subwave C — Targeted evidence acquisition
+- build claim-specific check-paths
+- enrich retrieval items/evidence with targeted checks and context-expansion metadata
+
+4. Subwave D — Positive grounding handoff
+- allow grounded-positive only when temporal/polarity/anchors/admissibility pass
+- preserve limited mode when evidence is not enough
+
+## Required Metrics
+- `claim_anchor_resolution_rate >= 0.95`
+- `targeted_evidence_hit_rate >= 0.80`
+- `limited_mode_correct_retention_rate >= 0.95`
+- `grounded_positive_answer_rate > 0` (positive cases)
+- `false_grounded_answer_rate = 0`
+
+## Run Artifacts (per wave run)
+- `README.md`
+- `run_summary.json`
+- `before_after_metrics.json`
+- `control_case_matrix.md`
+- `claim_anchor_audit.json`
+- `targeted_evidence_acquisition_report.json`
+- `grounded_positive_vs_limited_matrix.md`
+- `chat_export_core8.md`
+- `debug_payloads/`
+- `live_call_inventory.json`
+- `temporal_expansion_audit.json`
+- `evidence_pack_examples/` (`settlement`, `VAT`, `month-close`)
+
+## Current Execution Status
+- Code integration: done
+- Tests/build: done
+- Wave 19 run artifacts folder: `llm_normalizer/docs/runs/2026-03-29_Stage_04_Wave_19_Claim_Bound_Evidence_Acquisition_P0`
+- Preliminary verdict: `WAVE19_ACCEPTED_WITH_LIMITATIONS` (mock core-8 probe; live rerun still required)