Gate sensitivity

Planted-truth sweep, 200 replicate seeds per point, n = 1500 items, two models, per-draw flaky success probability q = 0.5, gate policy --require-sign-stable --min-effect-vs-noise 1.0. Seeds are derived from the label sensitivity-v1; the table regenerates byte-identically.

A pass rate at true gap 0 is a false pass: the gate approving a claim where no true difference exists. g80 is the smallest swept true gap the gate passes at least 80% of the time — the gate's detectable-gap resolution at that noise level.

k	flaky fraction	gap 0.0	gap 0.005	gap 0.01	gap 0.02	gap 0.05	g80
2	0.02	0.000	0.060	0.220	0.535	0.990	0.05
2	0.05	0.010	0.015	0.060	0.235	0.825	0.05
2	0.15	0.010	0.010	0.040	0.075	0.470	> 0.05
4	0.02	0.020	0.635	1.000	1.000	1.000	0.01
4	0.05	0.010	0.290	0.860	1.000	1.000	0.01
4	0.15	0.000	0.150	0.505	0.995	1.000	0.02
8	0.02	0.020	0.835	1.000	1.000	1.000	0.005
8	0.05	0.005	0.375	0.960	1.000	1.000	0.01
8	0.15	0.000	0.145	0.555	0.990	1.000	0.02

The gap-0 column is the measured false-pass rate. Rulings scope applies: a gate pass is a statement about this synthetic noise model at these parameters, not a certificate for any real benchmark (NO_K1_CERTIFICATE).