== coreultra9-gfni-rot-validate.log metric=wall 160 gated leg(s), 0 refused, reps=[1, 2, 3, 4], rungs=[256, 320, 378, 384, 392, 400, 448, 512, 640, 768] -- arm order, as banked (position derived from LEG timestamps within each rep and rung) -- rep 1 driver NARROW-REP arms=fold,force,auto,autoalt rep 2 driver NARROW-REP arms=force,auto,autoalt,fold rep 3 driver NARROW-REP arms=auto,autoalt,fold,force rep 4 driver NARROW-REP arms=autoalt,fold,force,auto 0 (rep, rung) group(s) disagree with the driver's banked order arm pos0 pos1 pos2 pos3 auto 10 10 10 10 autoalt 10 10 10 10 fold 10 10 10 10 force 10 10 10 10 BALANCED: every arm in every position equally often (counts [10]) -- box state, min / median / max over all gated legs -- foreign_cpu before leg (% of one core, other processes): 0.0 / 18.8 / 157.8 foreign_cpu after leg: 7.8 / 31.2 / 214.1 load_before (Win32_Processor LoadPercentage, whole box): 0.0 / 4.0 / 20.0 load_after: 0.0 / 3.0 / 12.0 five legs with the most foreign CPU either side: m384-auto-r2 91/214, m512-autoalt-r1 158/95, m378-auto-r2 147/103, m384-auto-r1 69/139, m640-fold-r2 94/133 first leg ts=2026-09-19T00:18:16.5903848Z last leg ts=2026-09-19T00:39:17.4330910Z -- decisions per arm (path / stripe width W / window count), and whether every rep agreed -- m=256 auto=fold autoalt=fold fold=fold force=W 512, 7 win m=320 auto=fold autoalt=fold fold=fold force=W 512, 7 win m=378 auto=fold autoalt=W 32, 7 win fold=fold force=W 512, 7 win m=384 auto=fold autoalt=W 128, 8 win fold=fold force=W 512, 7 win m=392 auto=fold autoalt=W 256, 8 win fold=fold force=W 512, 7 win m=400 auto=fold autoalt=W 256, 8 win fold=fold force=W 512, 7 win m=448 auto=W 512, 10 win autoalt=W 512, 10 win fold=fold force=W 512, 7 win m=512 auto=W 512, 10 win autoalt=W 512, 10 win fold=fold force=W 512, 7 win m=640 auto=W 512, 11 win autoalt=W 512, 11 win fold=fold force=W 512, 7 win m=768 auto=W 512, 11 win autoalt=W 512, 11 win fold=fold force=W 512, 7 win -- PAIRED fold minus force per rep (positive = force is cheaper) -- m n fold med force med med diff med % min diff max diff sign p 256 4 1.82 2.13 -0.296 -16.24% -0.332 -0.268 0/4 0.1250 n.s. 320 4 2.10 2.16 -0.055 -2.62% -0.066 -0.018 0/4 0.1250 n.s. 378 4 2.40 2.21 +0.179 +7.47% +0.151 +0.209 4/4 0.1250 n.s. 384 4 2.43 2.21 +0.207 +8.52% +0.154 +0.218 4/4 0.1250 n.s. 392 4 2.45 2.25 +0.224 +9.14% +0.145 +0.250 4/4 0.1250 n.s. 400 4 2.47 2.20 +0.244 +9.91% +0.201 +0.309 4/4 0.1250 n.s. 448 4 2.69 2.24 +0.452 +16.80% +0.416 +0.476 4/4 0.1250 n.s. 512 4 3.00 2.30 +0.709 +23.59% +0.658 +0.723 4/4 0.1250 n.s. 640 4 3.65 2.36 +1.271 +34.84% +1.258 +1.325 4/4 0.1250 n.s. 768 4 4.32 2.49 +1.841 +42.58% +1.715 +1.854 4/4 0.1250 n.s. sign = reps where force was cheaper; p = exact two-sided binomial. At n=4 the floor is p=0.125. unanimous fold: [256, 320] mixed signs: none unanimous force: [378, 384, 392, 400, 448, 512, 640, 768] => crossover (unanimous signs) is between m=320 and m=378 -- PAIRED autoalt (base) minus auto (change) per rep (positive = the change is cheaper) -- m n autoalt med auto med med diff med % min diff max diff sign p 256 4 1.81 1.81 -0.007 -0.36% -0.028 +0.009 2/4 1.0000 n.s. 320 4 2.09 2.08 +0.002 +0.07% -0.017 +0.016 2/4 1.0000 n.s. 378 4 7.06 2.35 +4.710 +66.73% +4.544 +4.795 4/4 0.1250 n.s. 384 4 3.25 2.40 +0.850 +26.15% +0.836 +0.864 4/4 0.1250 n.s. 392 4 2.57 2.43 +0.143 +5.57% +0.112 +0.177 4/4 0.1250 n.s. 400 4 2.58 2.46 +0.118 +4.57% +0.062 +0.143 4/4 0.1250 n.s. 448 4 2.32 2.34 -0.006 -0.24% -0.034 +0.010 1/4 0.6250 n.s. 512 4 2.36 2.36 -0.007 -0.30% -0.023 +0.028 2/4 1.0000 n.s. 640 4 2.48 2.50 -0.018 -0.75% -0.028 +0.025 1/4 0.6250 n.s. 768 4 2.60 2.62 -0.015 -0.58% -0.018 -0.012 0/4 0.1250 n.s. the narrowing rule's verdict: read only where the two arms DECIDED differently (section 3). -- auto over min(fold, force) OF THE SAME REP (the brief's acceptance ratio), median and range of per-rep ratios -- m=256 n=4 median 0.992 range 0.987-1.001 m=320 n=4 median 0.991 range 0.972-0.997 m=378 n=4 median 1.078 range 1.054-1.090 m=384 n=4 median 1.085 range 1.069-1.093 m=392 n=4 median 1.088 range 1.072-1.094 m=400 n=4 median 1.121 range 1.095-1.129 m=448 n=4 median 1.041 range 1.031-1.059 m=512 n=4 median 1.029 range 1.026-1.033 m=640 n=4 median 1.056 range 1.055-1.065 m=768 n=4 median 1.055 range 1.048-1.061 -- position effect, two-way (arm level then rep level divided out): what running Nth is worth, per rung -- m pos0 pos1 pos2 pos3 pos0-pos3 256 -0.15% -0.42% +0.11% +0.44% -0.59% 320 +0.01% +0.02% -0.02% -0.08% +0.09% 378 -0.34% -0.40% +0.08% +0.45% -0.79% 384 -0.03% -0.37% +0.19% -0.06% +0.03% 392 +0.04% -0.43% +0.25% -0.25% +0.29% 400 +0.13% -0.56% +0.40% -0.13% +0.26% 448 -0.07% -0.05% +0.09% -0.09% +0.02% 512 -0.24% +0.38% -0.46% +0.32% -0.56% 640 +0.04% -0.16% -0.45% +0.52% -0.48% 768 +0.12% +0.09% -0.39% +0.43% -0.31% across rungs, median by position: -0.01% -0.26% +0.09% +0.13% read the GAP between positions, not the levels (they are centred). A FIXED order banks pos0 minus the position each arm always held as a systematic term it cannot see.