iter 8 — pooled-path order invariance (closing the iter-3 thread)

SURVIVES 2026-07-10 · the last open thread · 36 verdict checks × 4 orderings × 2 statistics · 0 changes · clean 1/3

Claim under attack (assumed true only to destroy it)

evaluate_pooled's CI-gated verdicts — the binding ρ gate and the advisory ξ gate, both replicate-and-agree gated since iter 6 — are invariant to within-symbol row reordering: the SAME logical data in a different row order cannot change any verdict. This is the iter-3 open thread: the pooled path's ξ still falls back to the POSITIONAL tie-break (pooled matrices carry no identity column), and row order also enters through the bootstrap index streams. Iter 6's MC gate should absorb it — never proven end-to-end until now.

Attack construction

Base + 3 hash-deterministic within-symbol JOINT reorderings (candidate and panel rows of each symbol permuted identically — logical data unchanged, segment structure preserved; real rows only).
The probe's own evaluate_pooled with all iter-6 hardening active (segmented bootstrap + replicate-and-agree), identical seeds across orderings.
6 regimes spanning vol × session × trend, pooled across 16 symbols (n_pooled 1,770–3,776); 6 candidates × 5 panel columns mixing heavy-tie counts and continuous features; max-over-panel exactly as production runs it.
Falsifier F1: any (regime, candidate) whose ρ or ξ verdict differs across the 4 orderings.

Result — verdicts are order-invariant

Check	Result
ρ (binding) verdicts across 4 orderings — 36 checks	0 changes
ξ (advisory) verdicts across 4 orderings — 36 checks	0 changes
Reading	The iter-6 replicate-and-agree gate absorbs order perturbations at verdict level — the positional tie-break in this path is not verdict-material. (Bit-level invariance is not a meaningful demand of a bootstrap path; verdicts are the output.)

Evidence

File	What
a8_pooled_order_attack.py	The attack — falsifier pre-declared in-file; probe's own evaluate_pooled under test.
a8_results.json	Per-regime, per-candidate verdict quadruples for both statistics.
a8_stderr.log	Run log under the 2 CPU / 2 GB scope.
regression_tests/test_a8_pooled_order.sh	Permanent regression test — re-runs the attack, asserts SURVIVES.

Consequence

The iter-3 open thread is closed, and with it every item of the A1–A7 attack library and every thread opened along the way has now been attacked at least once. Campaign totals so far: 8 iterations, 7 real mechanism findings fixed, 8 permanent regression tests. Clean rounds: 1 of 3 — two more consecutive clean rounds trigger ROBUST-ENOUGH → operator review.

Open threads

None known. Next firings: extend the attack library (novel harnesses) or accumulate clean rounds.

iter 8 · pooled-path order invariance · verdict SURVIVES · one atomic commit: iter 8: pooled-path order invariance — SURVIVES · append-only.