iteration 14 · 2026-07-23 · usefulness axis · laptop-drives-bigblack
📈 #4 Rank IC — period-IC rewrite CHECKPOINT · 4/5 robust
The iter-13 plan is executed: the significance test now lives on the period-IC series. Four of five gates are robust across 5 seeds; the fifth (null-FPR ≤ .01) knife-edges because of a genuine near-white-data lag tension — flagged for operator ruling (like the #0 PBO descope). Supersedes iter 13.
4 / 5
gates robust across 5 seeds
2.2–3.7×
SE-understatement (≥2 ✓)
SE² 5–14
N_min(H) ratio (≥2 ✓)
FPR .006–.014
null-FPR straddles .01 ⏳
Preflight (resource-only): load1 0.72 · 42 GiB available · si/so ~0 · ClickHouse active, readonly=2. 5c/5G/no-swap capped, single-thread BLAS.
State across 5 seeds
| Gate (§7 row 4) | Result (seed range) | Target | |
|---|
| Admit demo | ICIR 0.66–0.80, HAC-t ~23, sign-stable | \|t\|≥3, ICIR≥.5 | robust |
| Power @ IC=.03 / T=750 | power 1.0, achieved 0.025–0.029 | ≥ .8 | robust |
| Naïve-iid SE-understatement | 2.21–3.73 | ≥ 2× | robust |
| 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 | robust |
| Envelope FNR (non-mono + interaction) | both missed every seed (interaction \|t\|≤2.6) | FNR=1 | robust |
| Null FPR ≤ .01 | perm-calibrated held-out {.010, .014, .007, .011, .006} | ≤ .01 | 2/5 fail |
What was resolved (verify-before-report)
- Period-IC significance (not bar-level, which is near-white on BTC 3s data). Period-lag frozen at 50 by the measured Hurst (H≈0.81 long memory ⟹ large HAC bandwidth; the SE-understatement plateaus ≥2× at lag 50 for every persistent feature — not lag-shopped).
- SE-understatement necessity grounds (#3's framing): naïve/HAC ≥2× on the real persistent period-IC (median 2.24), robust across seeds.
- N_min(H) via a textbook identity: N_min(persistent)/N_min(iid) = (measured SE-inflation)² ≈ 5. The direct contiguous-window curve is reported but drift-confounded (non-monotonic — long-memory local means swamp a fixed signal; that drift is the penalty), so the analytic identity is the clean evidence.
- Interaction-only exemplar fixed: an rr-independent balanced ±1 mask → the marginal cancels robustly (missed every seed).
The one open item — a genuine near-white-data tension
The null-FPR ≤ .01 gate knife-edges (permutation-calibrated held-out FPR 0.006–0.014; 2/5 seeds > .01). This is intrinsic: the SE-understatement ≥2 necessity requires a large period-lag (50), but the lag-50 HAC-t period-null is heavy-tailed (the #10 class), so 1%-FPR control is noisy and straddles the bound. A small lag would control FPR but kill the ≥2× necessity. On near-white BTC 3s-returns, the two cannot both hold cleanly.
I will not construction-shop a more conservative FPR margin to force a pass — that would violate the discipline held for #9/#10. This is the #0-PBO-descope situation: the §7 threshold assumes more autocorrelation than the data provides.
Resolution (operator ruling)
- Operator ruling on the null-FPR bound for near-white-return usefulness instruments: (a) relax the 1%-FPR bound to a robustly-achievable level (like the #0 PBO descope), (b) accept a conservative permutation-controlled margin, or (c) two frozen lags (small for FPR-size, 50 for the SE-understatement).
- Absent a ruling, the honest terminal candidate is GROUNDED-with-documented-envelope on the 4/5 robust gates + the FPR heavy-tail noted as the near-white-data operating boundary — pending operator confirmation.