iter 13 — PENDING diagnosability: silent kernel failures made visible

BREAKS → DEFENDED → CLOSED 2026-07-10 · closes the iter-12 opened thread · the defense immediately diagnosed the iter-12 mystery · clean 0/3 (reset)

Claim under attack (assumed true only to destroy it)

"A kernel that fails on every cell is distinguishable in the probe's verdict output from one that was evaluated and merely produced insufficient evidence — a broken kernel cannot silently masquerade as an uncertain candidate." PENDING is the correct fail-safe outcome, but for the terminal head-to-head (7 NEW kernels to be written) a silent compute failure would be indistinguishable from honest uncertainty — an audit hole.

Attack + result

The probe's own aggregate_worst_cell on the proven-failing petrosian_fd + the proven-working bartels_rank_vn_ratio (2 cells × 40k bars).
F1: failing kernel with zero output carries no failure record — FIRED (v1): candidate_matrix's bare except → NaN recorded nothing.
F2: working kernel carries a false failure flag — never fired.

DEFENSE (reporting mechanism — verdicts unchanged, bands untouched)

candidate_matrix:
+   failures[nm] = f"{ErrType}: {msg}"          # on exception
+   failures[nm] = "all-NaN output (kernel ran, produced no finite values)"
+   C.attrs["kernel_failures"] = failures       # zero signature ripple
aggregate_worst_cell:
+   per-kernel rollup {n_cells_failed, first_error} -> final[nm]["kernel_failures"]
+   per-cell stderr: "[cell] kernel FAILURE  @ : "

Both silent modes covered: kernels that RAISE and kernels that run but yield all-NaN.

Proof — and the diagnosis the terminal phase needed

Kernel	Verdict	kernel_failures (post-fix)
`petrosian_fd`	PENDING (unchanged — fail-safe)	{n_cells_failed: 2, first_error: "all-NaN output (kernel ran, produced no finite values)"} — it does NOT raise: it runs and yields NaN on the crypto substrate. The iter-12 mystery, solved by the defense itself.
`bartels_rank_vn_ratio`	BAN (unchanged)	null — no false flag (F2 clean)

Same attack re-run: SURVIVES, 0 breaches. §B declaration self-test: PASS.

Evidence

File	What
a13_pending_diagnosability_attack.py	The attack — falsifiers pre-declared in-file.
a13_results.json / a13_verdict.json	Post-fix falsifier evaluation + the probe's full verdict output with failure records.
a13_stderr.log · a13_stderr_postfix.log	v1 (BREAKS) and post-fix (SURVIVES) runs under the 2 CPU / 2 GB scope.
regression_tests/test_a13_pending_diagnosability.sh	Permanent regression test — re-runs the attack, asserts SURVIVES.

Consequence

The 10th real finding — an audit hole rather than a wrong number: broken kernels can no longer hide inside PENDING. The defense instantly earned its keep by solving iter-12's mystery (petrosian_fd = all-NaN, not an exception), which is exactly the diagnostic the terminal phase's 7 new kernels will rely on. Clean rounds: reset to 0 of 3. (Runtime note: 7 loop firings arrived batched this round and were collapsed into this single iteration per the one-iteration-per-firing contract.)

iter 13 · PENDING diagnosability · verdict BREAKS → DEFENDED → CLOSED · one atomic commit: iter 13: PENDING diagnosability — BREAKS→DEFENDED · append-only.