iter 12 — kernel path end-to-end: the probe's real production path

SURVIVES 2026-07-10 · first end-to-end run of real registry kernels through the fully-hardened cascade · twin positive control fires BAN · clean 1/3

Claim under attack (assumed true only to destroy it)

Iterations 1–11 attacked the machinery using production columns as candidate proxies. The probe's REAL path — registry kernels through candidate_matrix → aggregate_worst_cell — was never exercised with the iters-2–11 hardening active. The claim: it works end-to-end, and a kernel whose shipped twin sits in the production panel is flagged, never certified orthogonal.

Attack construction

The probe module imported AS-IS; KERNEL_PROBE_DIR → the real mql5 registry (76 kernels).
Kernels: petrosian_fd + bartels_rank_vn_ratio (shipped twins bar_petrosian_fd / bar_bartels_rank_vn_ratio live in the panel — built-in positive controls) plus sevcik_fd, hvg_clustering; 3 cells × 40k bars.
F2: run failure or malformed hardened verdict records. F3: a twin kernel certified PASS (WATCH/BAN/PENDING all mean "not certified").

Result — the cascade's core promise holds on its real path

Kernel	Verdict	Detail
`bartels_rank_vn_ratio` (twin)	BAN	ρ_worst 0.9963 vs its shipped twin · ξ_worst 0.9288, breadth 1.0 · eliminated at ρ · stability counts + dcorr advisory all present — the positive control fires end-to-end
`sevcik_fd`	WATCH	ρ_worst 0.8896 (fractal-dim family correlation — plausible, flagged not certified)
`petrosian_fd` (twin)	PENDING	zero measurable output — the kernel raised on the crypto substrate and `candidate_matrix` filled NaN; the fail-safe held (PENDING dominates PASS: never certified). See opened thread below.
`hvg_clustering`	PENDING

0 breaches on the pre-declared falsifiers: every verdict record carries the full hardened field set; no twin was certified.

Opened thread (observation, not a broken claim)

candidate_matrix swallows kernel exceptions into all-NaN silently — a failing kernel is indistinguishable from an infeasible one (no failure reason recorded). PENDING is the correct fail-safe outcome, but diagnosability suffers: 2 of 4 kernels produced nothing on this substrate and the verdict JSON cannot say why. Reporting-mechanism candidate for a future iteration.

Evidence

File	What
a12_kernel_path_attack.py	The end-to-end attack — falsifiers pre-declared in-file.
a12_results.json / a12_verdict.json	Falsifier evaluation + the probe's own full verdict output (meta + final).
a12_stderr.log	Run log under the 2 CPU / 2 GB scope.
regression_tests/test_a12_kernel_path.sh	Permanent regression test — re-runs the path, asserts SURVIVES.

Consequence

The last unexercised surface — the real production path — now has end-to-end evidence: the hardened cascade flags a shipped-twin kernel (BAN via ρ 0.9963) and fail-safes unproducible kernels to PENDING. Clean rounds: 1 of 3.

iter 12 · kernel path end-to-end · verdict SURVIVES · one atomic commit: iter 12: kernel path end-to-end — SURVIVES · append-only.