ГЛОБАЛЬНЫЙ РЕФАКТОРИНГ АРХИТЕКТУРЫ - Рефакторинг этапов Stage 3.8 - юи

This commit is contained in:
2026-04-12 00:54:55 +03:00
parent d65969d2ff
commit 8167bd228d
18 changed files with 1408 additions and 70 deletions
+25 -3
View File
@@ -2592,9 +2592,26 @@ 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.
3. False route drifts and generic responses reduced in regression packs.
4. Manual markup shows increase in correct/grounded labels.
4. Manual markup shows increase in correct/grounded labels.
Status: In validation (functional gates green; manual re-markup trend confirmation pending)
### Stage 3 Exit Audit (2026-04-11)
1. Functional gate:
- `npm --prefix llm_normalizer/backend run build` passed.
2. Regression gate:
- Stage 3 focused suite passed: `9 files / 356 tests`.
- (`addressQueryRuntimeM23`, `assistantWave17RunRegression20260411`, `assistantWave18ManualCommentsRegression`,
`assistantLivingRouter`, `assistantLivingChatMode`, `assistantSoftPolicyReply`,
`assistantBoundaryFallbackReply`, `assistantAnswerPolicyV11`, `assistantSemanticExtractionContract`).
3. Runtime stability gate:
- address lane includes protective fallback for legacy MCP schema error (`SubcontoDt1`) and avoids hard execution drop for debt-lifecycle receivables prompts.
4. Manual smoke:
- reviewer confirmed fixed questions on latest spot checks; domain breadth expansion is tracked separately and is not a Stage 3 blocker.
5. Full backend suite note (non-blocking for Stage 4 kickoff):
- `npm --prefix llm_normalizer/backend run test` reports red integration groups with `5000ms` timeout budget (`assistantEvalHarness`, `assistantP0EvalHarness`, `assistantStage2EvalHarness`, `assistantStage3LifecycleAcceptanceProbe`, part of `assistantEndpoint`/`assistantBroadGuard`).
- one user-facing policy assertion currently diverges when live MCP returns aborted fetch (`This operation was aborted`), producing execution-error soft reply instead of expected clarification template.
- these failures are tracked as stability/infra debt and do not block Stage 4 answer-layer kickoff for already green Stage 3 focused contract pack.
Status: Completed (Stage 3 gate approved; domain expansion moved to Stage 5 backlog)
## Stage 4 (P2): Human-Centric Answer Layer
@@ -2607,7 +2624,12 @@ Goal:
- best next step.
2. Keep claim-to-evidence binding strict.
Status: Planned
Stage 4 kickoff preconditions:
1. Stage 3 functional + regression gates are green.
2. No route/MCP interface changes are required to start Stage 4.
3. Remaining acknowledged risk: domain coverage is still limited and will be expanded iteratively.
Status: Ready to start (kickoff approved on 2026-04-11)
## Stage 5 (P3): Quality Loop Driven By GUI Markup