{
 "mechanisms": {
  "fertility_priors": {
   "date": "2026-09-25",
   "enabled": false,
   "gold": "sword/ChiUns",
   "lambda": 2.0,
   "note": "measured 2026-09-25 (roadmap M1), whole Bible vs its own SWORD/ChiUns gold, lambda=2: the intended negative control (isolating, no case/articles) did NOT behave as predicted -- it is the LARGEST regression of the six languages measured. cmn genuinely has GB074=1/GB075=0 (a real 'before' adposition direction, e.g. Chinese 的 as a free genitive-type particle is a plausible real phenomenon here) AND GB089=GB090=0 (subj_incomplete=True), so 7126 anchors were flagged -- NOT a bug in isolation, but the finite_verb channel firing for an isolating language with no subject-doubling convention is the more likely culprit. eflomal F1 .6834->.6752 (Delta-0.0082, floor 0.0004), exact_span 100207.5->94932 (-5275.5, the biggest single move measured); +gloss F1 .6849->.6774 (Delta-0.0075), exact -4947. CONTROLS RUN (the required one-language check for this M1 batch): a typology-blind null-prior at the same evidence scale (7125 anchors, lambda=2) is WORSE (F1 .6665, Delta-0.0169, exact -7712.5) and an inverted placebo (wrong anchors, 6638) is BETTER but still negative (F1 .6782, Delta-0.0052, exact -1625). Ranking baseline > invert > real-ferf > null-prior: correct targeting measurably REDUCES the damage relative to blind targeting, but does not reverse it into a win -- for cmn, ANY added fertility pressure appears to be a net cost, not just misplaced evidence. GAP found during this measurement, FIXED same day (2026-09-25, immediately after M1): `fertility_priors._subject_indexing_incomplete()` took no `iso`/typology_fallback parameter and read raw Grambank codes only; for a language with NO Grambank entry at all, `grambank={}` and the function returned True (incomplete) UNCONDITIONALLY -- i.e. 'no data' was silently treated as 'confirmed no subject indexing'. This inflated spa/ben/asm's finite_verb flagging identically to cmn's (all four got within 1 of the same anchor count, 7125-7126, despite different corpora) even though only cmn's True value is an actual measured Grambank fact. Fixed by checking code PRESENCE before checking value (`tests/test_fertility_priors.py`'s new regression tests). cmn's OWN verdict above is UNAFFECTED by the fix -- cmn's GB089=GB090=0 is real, present Grambank data, not an absence, so its anchor count and result are unchanged; the fix only changed spa/ben/asm (see their own re-measured entries).",
   "source": "measured"
  }
 }
}