| { | |
| "conventions_md": "config/llm_conventions/arb.md", | |
| "mechanisms": { | |
| "fertility_priors": { | |
| "date": "2026-09-24", | |
| "enabled": false, | |
| "gold": "clear/AVD", | |
| "lambda": 2.0, | |
| "note": "measured 2026-09-24, whole Bible vs Clear gold, lambda=2: WASH, inside the noise floor on both scoring conditions (eflomal F1 .87275->.8729, Delta+0.00015 vs floor 0.0010; +gloss F1 .87415->.8744, Delta+0.00025 vs floor 0.0006). Not a loss -- flat, not worth the extra alignment-time complexity. Plausible unconfirmed hypothesis: arb already has GB070/072=1 (real morphological case), so its genitive/dative marking may already be adequately signaled without extra target fertility -- not verified, just a candidate reason not to spend more time here.", | |
| "source": "measured" | |
| } | |
| } | |
| } | |