GoldenMatch

Highlights

v1.11.0 ships the negative-evidence + clustered-identity-guard infrastructure as the foundation for v1.12's T3 recovery. DQbench composite essentially unchanged at 66.99 (v1.10 was 66.91) — the indicator-driven demote rule cannot fire on real DQbench T3 because of two structural findings (collision-signal not discriminative; iteration budget exhausted). v1.12.0 (released alongside) reverses v1.11's spec §Non-goals to deliver Path Y, where T3 finally moves.

What shipped

• MatchkeyConfig.negative_evidence: list[NegativeEvidenceField] | None field. Each entry: field, transforms, scorer, threshold, penalty. Default-None for v1.10 cache compat.
• _apply_negative_evidence in core/scorer.py subtracts NE penalties from weighted-matchkey positive score.
• promote_negative_evidence eager rule populates NE on weighted matchkeys for high-identity-prior columns (gated on identity_score >= 0.75 + cardinality_ratio >= 0.5 + col not in matchkey.fields + col not in blocking + col has exact-matchkey counterpart).
• compute_identity_collision_signal lazy indicator (8s budget) + IndicatorContext.identity_collision_signal method.
• rule_demote_clustered_identity at position 7 in DEFAULT_RULES. Threshold 0.75 (raised from 0.5 after T2 false-fire analysis).

Benchmarks (zero-config, no LLM)

| Dataset | v1.10.0 | v1.11.0 |
|---|---|---|
| DBLP-ACM | 0.9641 | 0.9641 |
| Febrl3 | 0.9443 | 0.9443 |
| NCVR | 0.9719 | 0.9719 |
| DQbench composite | 66.91 | 66.99 |

Honest framing

Shipped at parity with v1.10 — the demote rule never executes on real T3 due to iteration budget exhaustion (rule_blocking_too_coarse oscillates), and the collision_rate metric isn't discriminative (T3 = 0.59, T2 = 0.62). v1.12.0 (released alongside) addresses this via Path Y.

Breaking changes

None. Adds optional fields with default-None for v1.10 cache compat.

Full diff: https://github.com/benzsevern/goldenmatch/compare/v1.10.0...v1.11.0