iteration 28 · 2026-07-24 · incrementality axis · bounded slice 1 · laptop-drives-bigblack
The increment/mechanism phase opens. The Huberman-Kandel spanning test turns the feature into a strategy return and asks whether its alpha survives being spanned by the shipped factor set. Slice 1 grounds the regression + coverage + null-FPR + substitution; the power gate sits at the HAC-t boundary because the strategy returns are long-memory — exactly the case §7 routes to a block-bootstrap t (slice 2).
readonly=2. 5c/5G/no-swap capped; single-thread BLAS. Reused the #12 wide slice for the basis returns.| Gate (§7 row 14, slice 1) | Result (seed 20260723) | Target | |
|---|---|---|---|
| Coverage & null-t | coverage 0.955 · null-t mean −0.05 std 0.99 · FP 0.04 | ≥.90 & FP≤.06 | PASS |
| Null-FPR (spanned feature) | 0.0 | ≤ .05 | PASS |
| Substitution | spanned B \|t\| 0.22 (collapses) · incremental C survives 0.833 | B<1 & C≥.8 | PASS |
| Power @ α≈5bps | 0.725 (achieved α 4.19bps, HAC-t median 4.99) | ≥ .8 | boundary |
The power miss is not a magnitude problem and not fixable by more data:
z(f)·r are long-memory + heavy-tailed, so the HAC effective-N grows sub-linearly and the HAC-SE stays unstable window-to-window (~25% of windows dip below t=3 despite a t-median of 5–6).This is exactly what §7 row-14 anticipates: "null FPR≤.05 across Hurst/t(3)/block — else block-bootstrap t." The HAC-t is the wrong standard error for the long-memory strategy return.