iter 4 — A6: thin-cell ξ stability at the FLOOR_1000 boundary

BREAKS → DEFENDED → CLOSED 2026-07-10 · attack A6 (the 750-dbps thread from iter 1; KB §4 small-noisy-cell) · 8,680 thin windows · 77 frozen-line crossings found · sub-slice stability guard added · clean 0/3

Claim under attack

"FLOOR_1000 bar-equivalent grounding (frozen floor = 50 value-rows) makes per-cell ξ reliable — a thin-but-grounded cell cannot cast a noise-driven §B vote (a breadth count at ξ≥0.50, or an xi_worst contribution)." Aimed directly at the promotion gate: §B PASS requires xi_worst < 0.50 over ALL grounded cells, so ONE noisy thin cell flips PASS→WATCH; the theoretical null sd of ξ at n=60 is ≈0.082, so a 0.50-crossing from a true 0.2–0.3 needs only 2.5–3.5σ.

Attack construction (real contiguous sub-slices — nothing generated)

Each 80k-bar cell (3,990 value-rows) yields 20 disjoint contiguous windows of 60 value-rows (1,200 bar-equivalents — just above the frozen floor). Each window is byte-for-byte what a legitimately thin cell (newly-listed symbol, short-history 750-dbps cell) would present. Falsifiers: F1 window ξ ≥ 0.50 while full-cell < 0.40 (the 0.10 buffer excludes genuine near-line ambiguity); F2 window ξ > 0.95 while full-cell < 0.85.

ATTACK RESULT — 77 crossings: both §B directions contaminated

Falsifier	Hits	Pair	Worst case
F1 (false PASS-line votes)	48	`kyle_lambda_proxy\|ofi` — certified-orthogonal	window ξ 0.7174 vs full-cell 0.359 (NEARUSDT@250)
F2 (false BAN-line votes)	29	`price_impact\|volume`	window ξ 0.9508 vs full-cell < 0.85

30 of 62 cells produced crossings; max |ξ_window − ξ_fullcell| = 0.75. The floor alone does not make thin votes safe — in either direction.

DEFENSE — sub-slice agreement stability guard (bands untouched)

xi_cell_vote_stable(x, y, keys):        # rotation_robust_eval, iter 4
    stable ⟺ every contiguous sub-slice (2 halves + 3 thirds + 4 quarters) sits on the
             SAME SIDE of BOTH frozen lines (0.50, 0.95) as the full-cell reading

xi_worstcell_verdict_guarded(stable_votes, n_unstable):
    §B band arithmetic (FROZEN) on STABLE votes only     # BAN never rests on noise
    PASS with any unstable cell → PENDING                # never certify unresolved evidence

Parameter-free beyond the frozen lines — the k∈{2,3,4} contiguous partitions are a fixed construction (KB §5 subsample-consistency); FLOOR_1000 itself untouched.
Fail-safe fusion follows the established PENDING-SIBLING precedent: evidence-quality guards may degrade PASS to PENDING; they may never manufacture PASS or BAN.
Wired into aggregate_worst_cell: vote stability is judged on the argmax panel column (the pair actually driving the cell's vote).

Defense iteration recorded honestly: the first guard (halves+thirds) left ONE locally-consistent window through (BTCUSDT@750, ξ 0.578 vs truth 0.399 — all 5 sub-slices agreed); adding quarters closed it. Evidence: a6_v2_stderr.log (1 breach) vs a6_v2_stderr_k4.log (0 breaches).

Proof the defense closes the gap (same attack, re-run)

Check	Result
G1 — quarantine of noise-crossing thin votes	76/76 crossings judged UNSTABLE (none admitted)
G2 — declared §B verdicts through the guard, 62 full cells	7/7 reproduce (BAN·BAN·BAN·WATCH·PASS·PASS·PASS); all three PASS pairs 62/62 stable — zero collateral PENDING
G3 — end-to-end thin-cell scenario (crossing window appended as a 63rd cell)	PASS pair → PENDING (honest "can't certify"), never WATCH/BAN; BAN pair stays BAN (29/29)
A5 substrate-validity regression	SURVIVES (7/7)
§B declaration self-test	PASS

Evidence

File	What
a6_thin_cell_attack.py / a6_results.json	v1 attack — 77 crossings, per-cell window min/max table.
a6_thin_cell_attack_v2.py / a6_v2_results.json	Same surface vs the defended mechanism — G1/G2/G3, 0 breaches (final).
a6_stderr.log · a6_v2_stderr.log · a6_v2_stderr_k4.log	Run logs under the 2 CPU / 2 GB scope — incl. the intermediate 1-breach guard round.
regression_tests/test_a6_thin_cell.sh	Permanent regression test — re-runs v2, asserts SURVIVES.

Consequence

The promotion gate can no longer be swayed by a thin cell in either direction: noisy thin votes are quarantined, BAN requires stable evidence, and unresolved cells block certification (PENDING) instead of polluting it. This also retires the 750-dbps thread opened in iter 1 (thin 750 cells now quality-gate themselves). Clean rounds: 0 of 3.

Open threads

A3 block-length misspecification · A2 low-power blind spot · A4 cluster bootstrap — unattacked.
evaluate_pooled positional tie-break default (advisory path) — carried from iter 3.

iter 4 · attack A6 · verdict BREAKS → DEFENDED → CLOSED · one atomic commit: iter 4: A6 thin-cell floor stability — BREAKS→DEFENDED · append-only.