| { | |
| "conventions_md": "config/llm_conventions/fra.md", | |
| "mechanisms": { | |
| "fertility_priors": { | |
| "date": "2026-09-25", | |
| "enabled": false, | |
| "gold": "clear/LSG", | |
| "lambda": 2.0, | |
| "note": "measured 2026-09-25 (roadmap M1), whole Bible vs Clear gold (NT only, LSG), lambda=2: LOSS, not a wash -- eflomal F1 .87305->.8660 (Delta-0.0071, noise floor only 0.0003), exact_span 42269->41498 (-771); +gloss F1 .8751->.8683 (Delta-0.0068, floor 0.0004), exact -693.5. Consistent with the already-documented fra articles/case_marking regression (span_extension.py, section 5.2b) -- a third, independent mechanism (this time an eflomal-time prior, not a deterministic grab) also hurts fra. Not investigated further here; a real candidate-check/relation-gating weakness specific to fra, not yet root-caused for this mechanism.", | |
| "source": "measured" | |
| }, | |
| "spanext.base_mechanisms": { | |
| "date": "2026-10-09", | |
| "enabled": false, | |
| "gold": "clear/LSG", | |
| "note": "MEASURED 2026-10-09 (measure_flag runs DP7, fra_lsg vs Clear LSG, NT gold, word grain): span extension as shipped LOSES — raw F1 .8751 without vs .8329 with (precision .8695 -> .7897, recall .8808 -> .8811), convention-aware .8866 -> .8824 (-0.0042), exact spans -434. possession_affix adds nothing for fra (inert), so the loss is the article / case-marking widenings. Base mechanisms off for every fra edition; re-measure if a per-edition E6 verdict says otherwise.", | |
| "source": "measured" | |
| } | |
| } | |
| } | |