How do you prove an AI agent's output today is the output that passed your checks?

[tantanpq]

More of my build and test artifacts are being produced by AI agents, and my
verification story has not kept up. The agent reports success, the files land
in the repo, and after that the claim "this passed" rests on trust and
scrollback.

I built a small deterministic kernel to close that gap, and I'm testing
whether it's useful to anyone else. The mechanism is deliberately plain:

• The artifact bundle is frozen and every file is bound to its SHA-256 hash.
  A mismatch at read time rejects the run.
• Checks are declared before the run, from five kinds only: JSON equals, JSON
  includes, JSON array-find, text includes, hash equals. The runner makes zero
  model calls; a model can produce candidates, but it never votes on them.
• Attempts form a hash-chained ledger, so a failed attempt can't be quietly
  deleted and replaced with a passing one.
• Verification is clean-room: a second party re-checks the run from the
  manifest hash alone, without the producer's internal state.

Honest limits: it does no math and no statistics, and its verdict ceiling is
PASS_CANDIDATE_ONLY. It certifies that declared checks passed on frozen
files. It does not certify that the files are correct or true.

I'm running a paid pilot this week: the kernel is a $49 download — you run
certification checks on your own bundles with it. If you maintain a project where agent output lands unsealed, I'd like to
hear how you handle this today, and whether a seal like this would mean
anything to you. Reply here and I'll send the order details.