Iteration 16 — FRAGILE↔COND screen S-05 12 PASS · 2 OUT B-04 TRIGGERED — THE LAB'S HUNGRIEST BOUNDARY

In plain words: the boundary that separates lucky features from reliably-conditional ones had its interrogation — and this is the boundary the certification lab's own priority table marks as hungriest (it has exactly one trusted instrument there today). The screen enforced a sharp principle: evidence must be about a FEATURE, not about the whole panel. A verdict machine needs to say "feature X held in 7 of 10 regimes," not "the market got weirder." Twelve instruments cleared that bar: distribution-difference tests in three geometries, selection-stability measures (with the correlated-features fix), rank-trajectory overlap, region-resolved dependence with its significance test, recurrence coupling, sequential betting evidence, per-feature factor-loading congruence, and the worst-regime floor. Two elegant classics went out on exactly that principle — covariance shrinkage intensity and whole-matrix similarity are honest diagnostics of regime instability, but they describe the panel, not any feature; both are retained as context, neither as gauge. Twelve passers means the fourth handoff batch fires next — into the lab's most-starved slot.

 The per-feature bar for FRAGILE↔CONDITIONAL evidence:
   "feature X held in k of N named regimes"     ← moves a status    → PASS
   "the panel's covariance conditioning moved"  ← describes weather → OUT (context)
 A status machine grades features, not climates. Panel scalars stay
 in the toolbox as context — they never gate.