Reference lab · evidence

Context Layer Lab

A green dashboard misses two later problems. I built a deterministic diagnostic that compares newer receipts and schedules before it answers.

Scenario clock: 09:10 UTC, 28 Jul 2026, fixed.

Are all scheduled automations healthy?

The dashboard, 07:30: Yes. (expired at 07:55); Expired at 07:55. It still says yes.
The evidence, 09:10: Checking the evidence in your browser.

Loading the diagnostic.

What arrived between 07:30 and 09:10

07:30 Dashboard says all jobs are healthy.Valid until 07:55. Expired
07:55 Dashboard expires. Nightly Export is due.
08:02 Nightly Export succeeded. Succeeded
08:15 Docs Build is due.
08:30 Weekly Report is due.
08:40 Docs Build failed. Failed
09:05 Weekly Report finished. Output not published. Not published
09:10 The question is asked.The answer uses everything above. Now

Data Sync is due at 17:30, so it is not counted yet.

Synthetic scenario on a fixed clock. Runs in your browser. Nothing is uploaded.

This lab recomputes its answer in your browser with JavaScript. Without it, here is the published answer from the generated eval report: the 07:30 dashboard says healthy, but by 09:10 Docs Build had failed and Weekly Report had not published, so the answer is needs attention. Read the eval report.

Why the answer is no.

The four checks appear once the diagnostic has loaded.

Each scheduled job.

The job table appears once the diagnostic has loaded.

Is this just a timestamp check?

The comparison appears once the diagnostic has loaded.

Try to fool the checker.

The answer is never read from the file. The checker recomputes it from the run receipts and rejects a snapshot that disagrees with itself. Pick an edit. The page applies it to a copy of the public snapshot and runs the same checker.

The limit: if someone rewrites the receipts, their claims, and the answer together, the checker will agree with them. It checks that a snapshot is consistent, not that its sources are true.

Check your own snapshot.

Generate a snapshot from your own records and open it here. It is read in this tab only and checked the same way. A snapshot that does not match its own evidence is rejected, and the page keeps its previous answer.

git clone https://github.com/Kaagemusha/context-layer-lab
cd context-layer-lab
npm ci
npm run diagnose -- --output snapshot.json

Showing the public scenario.

Search the evidence.

The diagnostic can also search the 4 records it used. Current records always rank before expired ones.

Try:

Keyword ranking (BM25F), current records first, up to 5 results. It searches this snapshot only.

What this proves, and what it does not.

It proves the records you can open support the answer, under explicit freshness and schedule rules. The answer is recomputed from them in your browser, and a snapshot that disagrees with itself is rejected.

It does not prove that a source is true, give anyone permission to act, or fill in what the records leave out. Missing evidence is reported as missing.

Claim	How to check it
17 of 17 evaluations pass: 7 record and operational, 10 retrieval.	Read the eval report, or run `npm run eval`.
No model call.	Open your browser's network panel and press Run the public scenario. Nothing is requested.
Same snapshot, same answer.	Run it twice and copy the answer each time. The text is identical.
It fails closed.	Try to fool the checker, above.

Version 1.2.2, released 2 Oct 2026.

Next: what may run.

An answer is not permission to act. Lab 2 of 2 starts from this answer: an agent proposes a retry of the failed Docs Build, but cannot approve it itself.

Open Governed Action Lab