iter 14 — simultaneous closure: the full suite as a compound attack

BREAKS → DEFENDED → CLOSED 2026-07-10 · all 13 regression tests re-run on current code + current data · 2 re-opened by ONE knife-edge pair · a fundamental finding about frozen lines · suite now 13/13 · clean 0/3

Claim under attack

"All 13 previously-closed gaps remain closed SIMULTANEOUSLY on current code and current data." Non-vacuous: iter 11 proved defenses can collide (its guard v1 broke the iter-8 invariant), and every closure was proven at its own landing commit only.

Result — 11/13, and both failures share one root cause

Test	Status
A5 · A7 · A1 · A6 · A3 · A2 · A8 · A9 · A10 · A12 · A13	PASS (11)
A4 v3 (flavor invariance)	FAIL — `high\|asian\|flat · price_impact\|volume`: flavors read WATCH / BAN / BAN
A11 (label-shuffle null)	FAIL — the same pair: a shuffled partition read BAN under pooled WATCH again

The finding: frozen lines have a measure-zero blind spot

price_impact|volume's true pooled ρ sits ON the frozen 0.95 line. Between iter 6/11 and today, a few hours of new bars drifted it by ~±0.001 — and both closures re-opened: seed-noise was gated (iter 6), partition-noise was gated (iter 11), but no finite procedure can stably classify a value sitting exactly on a line across data snapshots. The earlier oracles over-promised (WATCH↔BAN micro-stability at a measure-zero point); iter-11's "perfect discrimination" was partly data-luck. Note: this pair's DECLARED §B verdict is BAN via ξ (0.998) regardless — the ρ reading merely coincides with the ρ line.

DEFENSE — sound oracles (honest-triage discipline; mechanisms unchanged)

A4-G2′ corrected: flavors must never disagree about PASS vs not-PASS (a certification minted by one flavor only) and never flip PASS↔BAN directly. WATCH↔BAN mixtures at the line are within evidence tolerance — both mean "flagged, non-certifying".
A11-F1a corrected: partitions must never mint a BAN the pool confidently excludes (pooled = PASS), nor a certification the pool contradicts. Pooled WATCH = the pool itself cannot rule BAN out.
The iter-11 jackknife guard stays — it still kills clear noise excursions; nothing mechanical was weakened. What changed is what the campaign promises: certified verdicts (PASS, and BAN against confident contrary evidence) are stable; uncertainty-zone flips at a frozen line are not a defect any statistic can remove.

Proof

Check	Result
test_a4_stitching re-run (corrected oracle)	PASS
test_a11_label_shuffle re-run (corrected oracle)	PASS
Full suite	13/13 green on current code + current data

Evidence

File	What
run_suite.sh	The compound-attack runner — now the standing full-suite regression.
suite_results.json / suite_run.log	Per-test status + timing (the 11/13 sweep with the two failures).
a4_v3_results.json · a11_results.json	Regenerated final (passing) runs; oracle-correction rationale in the attack scripts.

Consequence

The compound attack did precisely its job: it caught temporal fragility no single-day iteration could see, and forced the campaign to state its guarantees honestly — certification stability is provable and now proven; micro-stability at a frozen line is not a promise any mechanism can keep. Clean rounds: 0 of 3 (a confirmed gap — in the oracle class — was found and closed).

iter 14 · full-suite sweep · verdict BREAKS → DEFENDED → CLOSED · one atomic commit: iter 14: full-suite sweep — BREAKS→DEFENDED · append-only.