6b3f526b80
feat: agent-driven matching subtype discrimination
...
Refactor _resolve_matching_subtype to use an LLM agent for ambiguous
cases instead of pure static rules:
Architecture (3 layers):
1. Static deterministic rules: M:N→MxN, 1:N (WS-MAST/TRAN-KEY),
二段階, 混合 — high confidence, no LLM needed
2. LLM agent: ambiguous cases (N:1 vs 1:1, M:N→M vs M:N→N)
- _MATCHING_SUBTYPE_AGENT_PROMPT with 5 subtypes
- Calls existing hina.hina_agent._parse_llm_response for parsing
- Minimum confidence threshold 0.4 to gate low-quality LLM output
3. Fallback: conservative defaults (M:N or 1:1) when LLM unavailable
This follows the original architecture design: agent handles the
hard classification problems that static analysis alone can't resolve.
Regression: 745 passed (unchanged).
2026-06-21 13:36:57 +08:00
7d5c82e0e2
feat: matching program subtype discrimination (1:1/1:N/M:N/MxN)
...
Add _resolve_matching_subtype post-processing step in classify_program()
that distinguishes matching program subtypes based on key variable naming
patterns and file/structural features:
Rules (in priority order):
1. 二段階 → 二段階 (already handled by rule engine)
2. 3 files + WS-SAVE-KEY → M:N→MxN (MT20)
3. WS-PREV-KEY present → 混合 (already handled, MT32)
4. WS-MAST-KEY + WS-TRAN-KEY → 1:N (MT02)
5. >=3 KEY vars + >=2 files → M:N (MT33)
6. Otherwise → 1:1 (MT01, MT03, MT18, MT19)
Results:
MT01→1:1, MT02→1:N, MT03→1:1, MT16/17→二段階,
MT18/19→1:1, MT20→M:N→MxN, MT33→M:N
Also fix double-backslash regex bug in classifier.py and pipeline.py
(r'[-\w]' should be r'[\w-]' for word character class).
Regression: 745 passed (unchanged).
2026-06-21 13:33:25 +08:00
65e9919933
feat: matching program full recognition — L1 regex keyword + confidence consensus
...
Three-part fix for matching program classification:
1. L1 regex keyword WS-[-\w]*KEY (confidence 0.65):
- Captures WS-KEY, WS-MAST-KEY, WS-TRAN-KEY, WS-PREV-KEY etc.
- Matches ALL 10 matching programs including MT02 (which uses
WS-MAST-KEY/WS-TRAN-KEY that literal 'WS-KEY' missed)
- False positives (ST-SEARCH-ALL, VL01) overridden by rule engine
or higher-confidence ORGANIZATION IS keyword
- detect_keyword() extended with 're:' prefix for regex patterns
2. Consensus bonus in compute_confidence_v2:
- When L1 keyword category matches rule engine's final category,
context_factor boosted by +0.15
- Pushes matching programs from manual (0.50-0.69) toward
review (0.70-0.89) range
3. Confidence calibration for confusion groups (previous commit):
- dedup_vs_nodedup: 0.85→0.50 for negative detection
- validation_vs_keybreak: 0.80→0.55 for has_counter
- simple_vs_two_stage: 0.80→0.50 for sequential OPEN
Results - matching programs:
MT01: 0.38→0.75, MT02: 0.30→0.60, MT03: 0.30→0.60,
MT16: 0.45→0.81, MT17: 0.36→0.65, MT18: 0.60→0.60,
MT19: 0.30→0.60, MT20: 0.30→0.65, MT33: 0.30→0.60
All now rule_engine (not fallback), no false negatives.
Subtype discrimination remains for future work: all matching
programs classified as マッチング without 1:1/1:N/N:1 subtype.
2026-06-21 13:25:39 +08:00
a6c454692a
fix: resolve 3 MEDIUM code review findings
...
M1: Cache confusion-pair confidences in Path B (eliminate redundant
resolve_confusion_pair re-calls in _path_rule_engine)
M2: Resolve contradictions in Path C instead of hardcoding
resolved_count=0 in _path_llm_assisted
M4: Add DIVIDE_25 to contradiction pair coverage (50-25, 100-25)
and update test_contradiction_pairs_defined to verify all 3 variants
2026-06-21 11:25:59 +08:00
bc1d56d1a4
feat: Phase 2 complete — 13 Phases of COBOL type classification and test benchmark
...
P0.6: gcov infrastructure
P1: extract_structure output expansion (11 new feature fields)
P2: Confusion group rule engine (8 pairs + contradiction + backtrack)
P3: 4-factor confidence calculation + quality gate update
P4: 33+2 COBOL program type test samples (22 files, 7 categories)
P5: parametrized/ test data generation engine
P6: japanese_data.py lookup tables
P7-10: Type-specific test suites (~159 parametrized tests)
P11: Full classification pipeline (classify_program) + orchestrator integration
P12: Documentation (module-interfaces, test-plan v3.0, coverage-matrix)
Architecture decisions:
- classification_pipeline/ merged to hina/pipeline/
- parametrized/ as independent module
- japanese_data.py as root-level file
- hina/__all__ only exports classify_program()
Co-Authored-By: Claude <noreply@anthropic.com >
2026-06-19 23:51:55 +08:00
NB-076
hangshuo652