File size: 1,219 Bytes
e00d9cb | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 | {
"conventions_md": "config/llm_conventions/eng.md",
"mechanisms": {
"fertility_priors": {
"date": "2026-09-24",
"enabled": true,
"gold": "clear/BSB",
"lambda": 2.0,
"note": "measured 2026-09-24, whole Bible vs Clear gold, lambda=2 (fixed from hin's own sweep, not independently swept -- see the plan's own sequencing). eflomal F1 .6114->.6150 (Delta+0.0036, noise floor 0.0004), exact_span 53757->54043 (+286); +gloss F1 .61305->.6168 (Delta+0.00375, noise floor 0.0006), exact 54123.5->54438 (+314.5). Clean win, both bars cleared well past the noise floor -- same mechanism as hin (eng has both a free-adposition side for case_marking AND a free-article side for definiteness, so both signal channels apply).",
"source": "measured"
},
"spanext.definite_trigger": {
"date": "2026-09-24",
"enabled": false,
"gold": "clear/BSB",
"note": "measured 2026-09-24 (whole OT, Clear gold): aggregate F1 unchanged (.646->.646); of 401 triggered pairs, 391 had no checkable gold link at all and 0 of the remaining 10 were correct — no evidence of value as built, plausibly because the proper-name rule in compute_definite dominates its output. Kept OFF.",
"source": "measured"
}
}
}
|