iter 20 — non-finite consistency, and CONVERGENCE

SURVIVES → ROBUST-ENOUGH — OPERATOR REVIEW 2026-07-10 · clean round 3 of 3 · the loop stops attacking and awaits the operator's ruling (handoff H8)

The final attack

Claim: no real production column or working kernel output carries ±Inf that materially corrupts the h_norm stage — the ONE statistic a code audit shows does not isfinite-mask (it drops only NaN), while BANning on min-over-cells.

Check	Result
Inf census — every float production column, 60-day all-symbols window (one ClickHouse pass)	ZERO Inf
Inf census — 5 working kernels' real outputs × 2 cells	0/10 kernel-cells carry Inf
Latent asymmetry	Documented + armed as a tripwire: `test_a20_nonfinite.sh` fails loudly the day real Inf appears, with the KB-consistent fix named (isfinite-mask h_norm like every other stage)

CONVERGENCE — three consecutive clean rounds

Round	Attack	Verdict
iter 18	expanded-suite sweep — all 16 regression tests on further-hardened code + fresh data	SURVIVES (16/16)
iter 19	750-inclusion consistency (the oldest carried observation)	SURVIVES (7/7 invariant, raw + guarded)
iter 20	non-finite handling consistency	SURVIVES (zero Inf in reality)

What 20 iterations bought — the campaign scoreboard

14 real findings found & fixed: tie-rank null bias (ξ=⅓ on permuted binaries) · stable-sort temporal leak (+0.244 null bias) · positional-key order dependence (breadth votes crossing 0.50 under re-ordering) · thin-cell noise votes (0.72 vs truth 0.36) · stitched-bootstrap distortion (3.6× the documented residual) · MC knife-edge verdicts · the LIVE ξ low-power blind spot (14 shipped pairs, dCorr to 0.87) · worst-of partition-selection BAN minting · pandas library rot of the h_norm degenerate contract · 2 audit holes (silent kernel failures; missing driving-cell attribution) · an input-hygiene crash (duplicate slate names) · 2 oracle-class over-promises whose correction yielded the knife-edge theorem and resolution-localized redundancy.
17 regression tests + the run_suite_v2.sh compound suite — all green on current code and current data.
Guarantees stated precisely: certification stability (PASS and confident verdicts invariant across seeds, partitions, orderings, epochs, resolutions) · byte-level reproducibility · fail-safe PENDING with full diagnosability · advisory surfacing of the ξ low-power class (banding = operator's, H7).
Every A1–A7 item, every KB §6 control, every §B stage, all three transfer axes, the production kernel path, and every carried observation attacked at least once. Coverage of ~30M row-reads, SELECT-only, all compute under 2 CPU / 2 GB, zero synthetic data, zero band changes.

What happens now (operator handoff H8)

The loop STOPS ATTACKING and idles. Rule on the verdict: (a) declare the rotational probe operational → terminal phase (write the 7 missing kernels with ≤1e-9 oracle proofs, then the legacy-vs-rotational head-to-head on the 10 candidates); or (b) direct further attacking. Optional at review: the dCorr band question (H7) and the probe-file split (953/1000 lines).

Evidence

File	What
a20_nonfinite_attack.py / a20_results.json / a20_stderr.log	The final attack + census results.
regression_tests/test_a20_nonfinite.sh	The armed tripwire.
CLAUDE.md	The full append-only ledger, rows 0–20 + ★ ROBUST-ENOUGH.

iter 20 · non-finite consistency · SURVIVES · clean 3/3 · ROBUST-ENOUGH — OPERATOR REVIEW · one atomic commit: iter 20: non-finite consistency — SURVIVES (ROBUST-ENOUGH) · append-only.