| "note": "RE-MEASURED 2026-09-27 (roadmap D4a follow-up) after gram_struct.merged_direction replaced span_extension's own typology.py-only fallback table (D4a) -- fertility_priors.py inherits that same direction_for() call, so it picked up the change automatically, no code edit needed here. gram_struct now resolves an `article` direction for spa (before -- was None under the old WALS/lang2vec-only table, which has no article-order parameter at all), so the 'definite' relation gate -- CLOSED for every prior spa measurement below, including the 2026-09-25 one right above -- is now open. Anchors flagged: 4690->7727 (+3037, all from the newly-open definite gate; verified directly, byte-identical anchor SET size to a from-scratch build_fertility_priors() call under both old and new backends). Whole Bible vs Clear gold, lambda=2, --fertility-typology-fallback: a REAL WIN, clearing the noise floor (0.0012) by ~25x -- eflomal F1 .7742->.8043 (Delta+0.0301), exact_span 175534->182590 (+7056), precision AND recall both up (.822->.847 / .731->.765, not a tradeoff); +gloss F1 .7743->.8046 (Delta+0.0311), exact_span 176293->183333 (+7040). This REVERSES the 2026-09-25 verdict directly above (which was measured under the OLD, article-blind typology table) -- the mechanism itself was never the problem for spa, the missing `article` direction was. ben/asm do NOT flip (re-measured same day, still WASH/LOSS -- see their own updated notes) despite the identical anchor-count jump, so this is genuinely spa-specific, not a blanket unlock. `enabled` flipped true -- first fertility_priors win found via the typology_fallback path (spa/ben/asm's own path), and first ever verdict reversal from a later infrastructure fix rather than a fresh gold language.", |
| "note": "measured 2026-09-24 (whole Bible, Clear gold, WITH the identity guard): net positive on the conflict-aware ledger, 4,435 improvements vs 3,561 real conflicts. Aggregate F1 still reads as a drop (.774->.729) because most of the reported 'regression' is the gold-unclaimed-article metric artifact documented in span_extension.py's fra section, not a real quality loss — see that module's own docstring, not this note, for the full account.", |
