ops(hier): Phase A law/library decompose + snapshot freeze (replace-diagnose c3)

47 eval-target undecomposed non-news docs (law21+library24+document2) 분해+임베딩 (--skip-analysis, additive). 1005 leaf 생성 fail0, in_corpus 634 무손상 검증. snapshot doc_id_max=25912 chunk_id_max=71164 docs_decomposed 301->348. 측정 drift 0. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
2026-05-25 05:23:38 +00:00
parent fc9e0f1d8f
commit e860baa179
1 changed files with 42 additions and 0 deletions
@@ -0,0 +1,42 @@
+{
+  "version": "v0.2-hier-replace",
+  "label": "snapshot_after_phaseA",
+  "date": "2026-05-25",
+  "plan": "hier-hazy-waffle.md (PR-DocSrv-Hier-Replace-Diagnose-1)",
+  "purpose": "Corpus drift freeze for hier vs prehier (legacy) go/no-go measurement. Frozen AFTER Phase A law/library decomposition so the 47 newly-decomposed eval-target docs' leaves are included.",
+  "snapshot": {
+    "doc_id_max": 25912,
+    "chunk_id_max": 71164,
+    "documents_n": 22097,
+    "prehier_chunks_n": 31327,
+    "hier_leaves_emb": 12697,
+    "docs_decomposed": 348,
+    "hier_in_corpus": 634
+  },
+  "phase_a": {
+    "newly_decomposed_docs": 47,
+    "leaves_created": 1005,
+    "null_embedding_leaves": 9,
+    "skip_analysis": true,
+    "in_corpus_unchanged": "634 (invariant preserved)",
+    "doc_ids": [3776,3853,3854,3855,3856,3865,3867,3868,3878,3879,3886,3887,3888,3897,3916,3917,3920,3921,3980,3981,3982,4041,5205,11495,11496,11500,11503,11504,11505,11514,11515,11591,11617,11620,11625,11627,11644,11645,11646,11647,11669,11689,11691,11711,11712,13305,13652],
+    "note": "47 = eval-target undecomposed non-news docs (law 21 + library 24 + document 2). DOC_MIN_CHARS=4000 gate bypassed via --doc; section analysis deferred (Mac mini, separate axis)."
+  },
+  "eval_set": {
+    "total_cases": 51,
+    "target_docs": 101,
+    "target_docs_decomposed_after_phaseA": "33 + 47 (news 21 + single-chunk excluded by design)"
+  },
+  "model_config": {
+    "embedding": "BAAI/bge-m3 (production)",
+    "reranker": "BAAI/bge-reranker-v2-m3",
+    "search_mode": "hybrid",
+    "knn": "exact (enable_indexscan/bitmapscan=off in eval) — isolates chunking from ivfflat approximation"
+  },
+  "variants": {
+    "prehier": "document_chunks WHERE source_type IS DISTINCT FROM 'hier_section' (legacy 30952 + null-source 375)",
+    "hier_sim_raw": "post-replace simulation: hier leaf if doc decomposed else prehier (per-doc fallback), current builder output incl tiny leaves",
+    "hier_sim_clean": "raw + exclude childless leaves with length(trim(text)) < 30 (A1 held-out finding, measured as variable vs raw)"
+  },
+  "baseline_reference": "PR-Eval-V0_2 production baseline graded NDCG 0.659 (v0_2_baseline_2026-05-23.json) — harness sanity target (production ivfflat path)"
+}