| "note": "RE-MEASURED 2026-09-25 after the same subject_indexing fix as spa (anchors 7126->4690). Whole Bible vs Clear gold (IRVBen -- found and fixed a real config/gold_langs.json bug in the original M1 run: base_text was recorded as bare 'IRV', the parquet's real value is 'IRVBen', so every prior score against this entry silently read 0/0 verses and f1=0.0), lambda=2, --fertility-typology-fallback: now a clearer LOSS than the pre-fix measurement -- eflomal F1 .5994->.5985 (Delta-0.0009, floor 0.0007), exact_span 18427.5->18271 (-156.5); +gloss F1 .5999->.5993 (Delta-0.0007, floor 0.0005), exact -144. (Pre-fix: F1 roughly flat, Delta 0.0000/+0.00045, but exact_span already down -316.5/-282 -- 'not a win' either way, just for a different reason.) Counterintuitively, removing the spurious finite_verb anchors made the aggregate F1 WORSE, not better -- those spurious fills were incidentally near-neutral-to-helpful in aggregate, while the real remaining possessor/dative-only signal is net negative for ben on its own. Still no win under either measurement. RE-MEASURED AGAIN 2026-09-27 (D4a follow-up, spa's own note above explains why): anchors 4690->7727 (definite gate now open too), but UNLIKE spa this does not flip ben -- still a wash/small loss: eflomal F1 .5994->.5983 (Delta-0.0011, still inside floor 0.0007... barely past it), exact_span 18453->18247 (-206); +gloss F1 .5992->.5993 (Delta+0.0001, essentially flat), exact 18478->18283 (-195). Net: the extra definite-relation anchors are close to neutral for ben, same conclusion as before -- kept `enabled: false`.", |
| "note": "measured 2026-09-24 (whole Bible, Clear gold, WITH the identity guard): still net negative on the conflict-aware ledger, 87 improvements vs 217 real conflicts (down from 427/1,355 before the guard — an ~84% reduction in harm, but not a flip to positive). Kept OFF; revisit if the pronoun-stealing gap span_extension.py documents ever gets a further fix.", |
