iter 7 — the shuffled-null gate lands FIXED

Attack item X-1 · 2026-07-29 · harness harness/structural_gates.py · gate /xi:structural-gates

Why this went now, ahead of the remaining measurement items

X-1 is a standing operator hard rule, not an optional attack item: a shuffled-null negative control must exist as a gate in any evaluation harness, because a walk-forward leak (0.926 OOS, stable in 333 of 333 cells) was once caught only at PR time. Six iterations of this loop had produced measurements and no gate at all. That was a compliance gap in the loop's own work, which makes it the highest-value remaining item regardless of what the suggested order said.

What is shuffled, and why that is the right null here

These probes are redundancy screens, not predictors — there is no train/test split to permute. The relationship under test is "is this candidate reconstructable from the panel", so the null permutes the candidate within each cell. That destroys its association with every panel column while leaving both marginals exactly intact: ties, skew, point masses and the panel's own internal correlation structure all survive. Anything the instrument still reports is therefore its own artifact.

Both directions, because a gate that cannot fail is not a gate

Each leg runs a negative control (shuffled — must report nothing) and a positive control (a genuine near-duplicate — must not pass). If a positive control passes, the script fails on the grounds that the leg is inert. That lesson is from the compute envelope sitting unenforced for a day.

Prediction, written before running

Shuffling destroys dependence, so the true value is 0 — but the statistic is a worst-of-N, so it will not read 0. From iter 4's independent measurement, mean worst-cell ξ at true ξ=0 was 0.0256. I predicted the ξ floor at ≈0.02–0.03 and legacy's worst |ρ| at ≈0.10 (max of ~19 estimates at n=490, sd≈0.045). Both landed.

Result

=== leg 1 — rotation xi cascade ===
  NEGATIVE (shuffled): xi_worst mean 0.0256 · p95 0.0342 · max 0.0385 · PASS 60/60
  POSITIVE (near-duplicate): verdict BAN · xi_worst 0.9815

=== leg 2 — legacy worst |rho| ===
  NEGATIVE (independent): worst |rho| mean 0.0600 · p95 0.1106 · max 0.1561 · PASS 400/400
  POSITIVE (near-copy of duration_us): verdict BAN · worst |rho| 0.9925

Neither probe manufactures redundancy on shuffled data, and both positive controls fired. The ξ null mean of 0.0256 reproduces iter 4's 0.0256 exactly — two independently written harnesses agreeing on the same quantity, which is the cross-validation that makes both believable.

The number this iteration was actually for — the noise floors

The null distribution of a worst-of-N statistic is its noise floor. A band below that floor cannot be cleared; a band far above it is not doing work. Both bands sit comfortably above chance — so whatever else is wrong with these instruments, neither ξ's 0.50 nor legacy's 0.85 is set inside its own noise.

Worth stating precisely, because it is easy to over-read: this is a different question from the panel-R² defect. That line failed because it sat below where real shipped features live (0.65–0.91), not below the noise floor. Iteration 7 answers "is the band above chance?" — yes. It does not answer "is the band in the right place relative to real features?", which remains open.

What landed

/xi:structural-gates, wired into /xi:check-full, with /xi:doctor extended to verify both harnesses and run both gates. Zero free parameters — a gate is not a dial. The full suite runs green: oracle parity exact to 2.44e-15, shuffled null clean in both directions.

Limits

• The legacy panel here is 7 columns, not the real ~19. A worst-of-N floor grows with N, so the real legacy null floor is higher than 0.1106 — extrapolating from the max-of-N scaling, roughly 0.14 at p95. Still ~6× headroom, but the figure quoted is a lower bound.
• X-2 is still open. Future-perturbation invariance was not attempted, and it may not even apply cleanly: these probes have no train/test split, so "perturb test rows, train-side discovery must be unchanged" needs restating for a redundancy screen before it can be a gate. That restatement is the next iteration's problem, and it should not be waved through as if the analogy were obvious.
• Synthetic cells are independent. Same caveat as iter 4 — the real 62 cells are strongly dependent, which will widen the null distribution of the max. The floors here are therefore optimistic.

Evidence: harness/structural_gates_evidence.json. Reproduce: mise run xi:structural-gates