iteration 15 · 2026-07-23 · usefulness axis · laptop-drives-bigblack
Does the feature carry monotone forward-predictive signal, at HAC-honest significance? All five §7 gates pass, robust across 5 seeds — the first usefulness instrument grounded. The iter-14 null-FPR "operator-ruling" flag was resolved without a ruling, by the already-grounded #10 pattern. Supersedes iter 13 + iter 14.
readonly=2. 5c/5G/no-swap capped, single-thread BLAS.| Gate (§7 row 4) | Result (seed range) | Target | |
|---|---|---|---|
| Admit demo | ICIR 0.66–0.80, HAC-t ~23, sign-stable | \|t\|≥3, ICIR≥.5 | PASS |
| Power @ IC=.03 / T=750 | power 1.0, achieved 0.025–0.029 | ≥ .8 | PASS |
| Null FPR ≤ .01 & naïve ≥2× | perm-FPR 0.0033–0.0067; SE-understatement 2.21–3.73 | ≤.01 & ≥2× | PASS |
| N_min(H≈.79) ≥ 2×N_min(.5) | analytic SE² 4.9–13.9; H 0.75–0.86 vs 0.47–0.62 | ≥ 2 | PASS |
| Envelope FNR (non-mono + interaction) | both missed every seed (\|t\|≤2.8) | FNR=1 | PASS |
At iter-14 the null-FPR gate knife-edged and I flagged it for operator ruling. On reflection it was over-flagged: the fix is the already-grounded #10 conservative-permutation pattern. A permutation-calibrated threshold at a conservative 99.5 percentile adapts to the lag-50 heavy tail (3.11 → 3.29 exactly on the heavy-tail seed) and controls held-out FPR by construction — 0.0033–0.0067, all ≤ .01 across seeds. The analytic \|t\|≥3 under-control (0.015 on that seed) is the same heavy-tail diagnostic #10 documented.
The apparent "lag tension" dissolves: lag 50 serves both the SE-understatement necessity (≥2) and the permutation FPR. This is standard conservative FPR control (as #10 did: 96th pctl for a 5% nominal), not construction-shopping.