One corrupt signal.
A defense you can verify.
Trace one corrupted worker result through detection, review and recovery. Inspect every custody checkpoint as the replay advances.
Controlled software fault · no physical damage
Integrity before intelligence
Public replay has no access to private live hardware. Temperature, ECC and power UNKNOWN.
The error stays in the loop.
Semgrep
Local source scan pinpoints the fail-open fallback. Before/after scans bind the exact target and rule hashes.
Akash
Inference handoff is prepared. Actual Akash response, model, usage and latency remain NOT_TESTED.
ClickHouse
Cloud readback receipt: 55 rows, 55 unique occurrence IDs, HTTP 200. Inspect the separate analytics branch below.
Worker → integrity gate
Select a checkpoint to inspect the recorded evidence.
LoadingLoadingLoadingVerification pending
Browser recomputes committed leaves and every MMR prefix. Python also checks classifications, delta-G* and source bytes.
NOT_COMPUTEDNOT_COMPUTEDSelect a checkpoint
Loading execution evidence…
Video, transcript and evidence
The narrated sponsor walkthrough shows historical executions. The separate 12-leaf local review-action proof can be independently recomputed. A recorded replay does not invoke providers again.
Watch /video/ — captions & transcript ↗ How to evaluate — sponsor & customer use cases ↗ Recompute local FCG ↗