Releases
v2.4.2
Compare
Sorry, something went wrong.
No results found
Release 2.4.2 — Unified CEG Anchor Validation & 30-Item Real-World Audit Benchmark
Commit (v2.4.2 / main): 71ef0e3dfb2fffd84ed3acf98180a2e0403c9ba6
Unified CEG endpoint & dangling-edge validation (pre_submission_linter.py, venue_template_manager.py, academic-writing 1.3.2): Delegated _load_ceg_claim_map() to co-located export_ceg_anchors(), preventing support_status="supported" edges pointing to missing evidence_id targets from suppressing UNGROUNDED_QUANTITATIVE_CLAIM warnings, normalizing whitespace-padded evidence identifiers across dict/list schemas, rejecting directory inputs, and isolating co-located source execution from stale installed packages.
30-item real-world scholarly verification & statistical audit benchmark (tools/benchmarks/run_real_world_audit_benchmark.py, docs/benchmarks/real-world-audit-benchmark-v1.json): Zero-wheel-bloat reproducible evaluation suite across 16 live-captured bibliographic cases (TP=10, TN=6, FP=0, FN=0) and 14 published statistical & gate cases (TP=8, TN=5, FP=1, FN=0), with strict work/signal_counts completeness enforcement and 12-significant-figure cross-libm deterministic replay (content_sha256 b99be3ba0c58256ca29a2150789bc6c2e15371098efd6ccd4e0ab41db288dc9f).
Verified CI Runs:
Skill correctness QA: Run 43204127261 (18 / 18 passed)
Live external contract checks: Run 43204131775 (2 / 2 passed)
You can’t perform that action at this time.