iter 1 — A5: worst-cell substrate validity

SURVIVES 2026-07-10 · attack A5 (seeded #1 blocker) · 62/62 crypto cells grounded · 0/7 worked-example verdict mismatches

Claim under attack (assumed true only to destroy it)

The §B worst-cell ξ bands (CHATTERJEE-THRESHOLD-DECLARATION.md §B — calibrated on even-downsampled 8,000-row raw-column cell-slices over 10 fixed 2-month windows, 2021–2025) transfer to the rotational probe's substrate: stride-20 value-rows over the most recent 80,000 bars per symbol×threshold cell. The probe's own aggregate_worst_cell docstring flags this as an unproven caveat: "A frozen real-CH validity gate … must pass before these numbers are trusted." This iteration IS that gate, run adversarially.

Falsifier

Any of the 7 frozen worked-example pairs (declaration §B anchor table; labels frozen 2026-06-30/07-02, none tuned here) whose §B verdict on the probe substrate differs from the declared verdict. One mismatch = claim BROKEN (bands don't transfer → fix the statistic, never the band).

Mechanism fidelity

Cell grid = all 16 crypto symbols × sorted thresholds (62 cells, incl. 750 — exactly as aggregate_worst_cell builds cells)
Fetch = most recent 80,000 bars, ORDER BY close_time_us DESC LIMIT n, reversed chronological (mirrors fetch_cell)
Substrate = strided indices range(WIN=200, len, STRIDE=20) → 3,990 value-rows/cell; grounding = FLOOR_1000 bar-equivalents
Statistic + verdict = the probe's own xi_max_pair (both directions) and rotation_robust_eval.xi_worstcell_verdict (frozen §B: PASS<0.50 / BAN>0.95 AND breadth≥0.80)
Invariants: ClickHouse SELECT-only · systemd-run scope 2 CPU / 2 GB / no swap · real data only · bands imported, never redefined

Result — all 7 frozen verdicts reproduce

Pair	Declared (§B)	Probe substrate	xi_worst	breadth	worst cell	match
`ofi \| turnover_imbalance`	BAN	BAN	0.9992	1.000	FILUSDT@100	OK
`agg_record_count \| individual_trade_count`	BAN	BAN	0.9994	1.000	LINKUSDT@100	OK
`price_impact \| volume`	BAN	BAN	0.9992	0.984	LINKUSDT@100	OK
`price_impact \| sell_volume`	WATCH	WATCH	0.7686	0.871	XRPUSDT@250	OK
`kyle_lambda_proxy \| ofi`	PASS	PASS	0.3935	0.000	ADAUSDT@750	OK
`bar_cecp_velocity \| ofi`	PASS	PASS	0.0530	0.000	DOGEUSDT@250	OK
`bar_dispersion_entropy \| trade_intensity`	PASS	PASS	0.3984	0.000	ETHUSDT@250	OK

Margin behaviour (honest reading). The substrate shift moves the numbers — it does not move any verdict. The PASS ceiling pair (kyle_lambda_proxy|ofi) drops 0.494 → 0.394 (deeper into PASS); the known-related WATCH pair drops 0.923 → 0.769 (still WATCH, breadth 0.871 keeps the two-dial BAN closed); the three duplicate/broad-BAN pairs stay ≥ 0.9992 with breadth ≥ 0.984. The moat between the PASS cluster (≤ 0.398) and the redundancy cluster (≥ 0.769) is preserved on the value-row substrate.

Evidence (every claim above cites these)

File	What
a5_substrate_validity_attack.py	The attack script — imports the probe's own verdict functions; SELECT-only; frozen pairs + expected verdicts in-file.
a5_results.json	Full per-pair, per-cell ξ values (62 cells × 7 pairs), verdict comparison, cell log.
a5_stderr.log	Run log under the systemd-run 2 CPU / 2 GB scope.
regression_tests/test_a5_substrate_validity.sh	Permanent regression test — re-runs the attack, asserts SURVIVES.

Consequence

The aggregate_worst_cell substrate caveat is now closed by a passed, re-runnable validity gate: the §B bands are usable on the probe's value-row substrate for the labeled anchor set. Clean round 1 of 3 toward convergence.

Opened threads (future iterations)

750-dbps leak: aggregate_worst_cell grounds 750-dbps cells into the ξ cascade even though the asset config marks 750 spearman_only ("thin per slice → ξ unreliable") and aggregate_worst_regime excludes them. FLOOR_1000 grounded all four 750 cells here (80k-bar fetch), and one PASS pair's worst cell IS a 750 cell (ADAUSDT@750, 0.394). Attack candidate for iter 2: is ξ on 750-dbps value-rows stable, or a thin-substrate artifact?
A1 metamorphic invariants, A3 block-length misspecification, A7 negative controls — unattacked.

iter 1 · attack A5 · verdict SURVIVES · one atomic commit: iter 1: A5 worst-cell substrate validity — SURVIVES · append-only (this page is never edited after landing).