ГЛОБАЛЬНЫЙ РЕФАКТОРИНГ АРХИТЕКТУРЫ - Рефакторинг этапов 3.1 - Добавлен строгий контракт семантического извлечения (валидность, качество, reason-codes, рекомендация применять canonical). Подключен semantic guard в predecompose/runtime: мягкое отклонение плохого canonical: отдельный safe-кейс для сырого полезного raw-фрагмента; fallback теперь выигрывает только когда это оправдано. Подключty semantic arbitration в tool-gate
This commit is contained in:
@@ -2349,6 +2349,23 @@ Plan (Stage 3):
|
||||
- Coverage critic threshold before final answer.
|
||||
- Reason-code taxonomy normalized.
|
||||
|
||||
Implemented in current pass (Stage 3.1 kick-off):
|
||||
1. Added strict semantic extraction contract for address predecompose:
|
||||
- `address_semantic_extraction_contract_v1` built from `source_message + canonical_message + predecompose contract`.
|
||||
- Captures: extraction snapshot, guard hints, quality, validity, apply-canonical recommendation, reason codes.
|
||||
2. Added semantic guard integration in predecompose runtime:
|
||||
- New rejection path for low-value/unsafe canonical rewrites (`normalized_fragment_rejected_semantic_guard`).
|
||||
- Fallback preference updated: deterministic fallback can win over LLM canonical when semantic contract rejects rewrite.
|
||||
3. Added semantic arbitration integration in tool-gate:
|
||||
- `resolveAddressToolGateDecision(...)` now honors `apply_canonical_recommended`.
|
||||
- New skip reason: `llm_predecompose_semantic_guard_rejected`.
|
||||
4. Extended runtime/debug observability:
|
||||
- semantic contract propagated through address tool-gate, deep debug payload, and living chat debug payload.
|
||||
- orchestration contract now includes semantic guard fields (`semantic_contract_valid`, `semantic_apply_canonical_recommended`, `semantic_reason_codes`).
|
||||
5. Added focused regression tests:
|
||||
- `assistantSemanticExtractionContract.test.ts` (new)
|
||||
- Existing router/chat regressions revalidated (`assistantLivingRouter.test.ts`, `assistantLivingChatMode.test.ts`).
|
||||
|
||||
Acceptance (Stage 3):
|
||||
1. LLM outputs strictly validated schema for extraction/decomposition (no free-form).
|
||||
2. Deterministic guards can block or downgrade answers when evidence insufficient.
|
||||
|
||||
Reference in New Issue
Block a user