| { | |
| "mechanisms": { | |
| "fertility_priors": { | |
| "date": "2026-09-27", | |
| "enabled": false, | |
| "gold": "clear/IRVAsm", | |
| "lambda": 2.0, | |
| "note": "RE-MEASURED 2026-09-25 after the same subject_indexing fix (anchors 7126->4690). Whole Bible vs Clear gold (IRVAsm -- same base_text bug as ben, fixed in the original M1 run), lambda=2, --fertility-typology-fallback: now a clear LOSS, well past the noise floor -- eflomal F1 .5572->.5544 (Delta-0.0028, floor 0.0022), exact_span 14676->14396 (-280); +gloss F1 .5573->.5548 (Delta-0.0025, floor 0.0019), exact -259.5. (Pre-fix: F1 inside the noise floor both ways, Delta-0.00015/+0.00005, but exact_span already down -445/-425.5.) Same pattern as ben: the spurious finite_verb anchors were masking a real, larger loss from the possessor/dative-only signal once isolated. 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); asm also does NOT flip -- eflomal F1 .5572->.5550 (Delta-0.0022, right at the 0.0022 floor -- borderline, not a confident win either way), exact_span 14656->14344 (-312); +gloss F1 .5570->.5548 (Delta-0.0022), exact 14675->14386 (-289). Net: essentially the same magnitude loss as the 2026-09-25 measurement, just from a different anchor mix -- kept `enabled: false`.", | |
| "source": "measured" | |
| }, | |
| "spanext.typology_fallback": { | |
| "date": "2026-09-24", | |
| "enabled": false, | |
| "gold": "clear/IRVAsm", | |
| "note": "measured 2026-09-24 (whole Bible, Clear gold, WITH the identity guard): same pattern as ben, smaller scale (115 vs 258 real conflicts before the guard). Kept OFF.", | |
| "source": "measured" | |
| } | |
| } | |
| } | |