A one-character integrity test for a signed AI measurement record

Can you detect when a published AI measurement record changes?

Try this specific record in CSOAI's free browser verifier. This is a small, reproducible integrity exercise using an existing historical record. It is not a new benchmark run or a model certification.

Try the challenge

1. Open the specified published record directly. The verifier automatically loads and checks this record.
2. Inspect the original outcome and the recorded 19 August 2026 date. If loading or a check cannot complete, report that outcome too.
3. Change only the example's accuracy value from 0.3 to 0.4. Keep its signature unchanged.
4. Verify again. Report the outcome you actually observe, including any check that cannot complete.

In our own browser check on 8 October 2026, the original returned VALID. Editing the score cleared the earlier result; re-verification returned INVALID, with both a card-ID mismatch and signature failure. We did not submit a public tally. These are operator checks, not independent participant results or a full security audit.

The published example is record a31be261cd078a7b5eb42dae5e3b6794c996efc83be20dbe5df7448c4a5c7bbe, recorded 19 August 2026, for axis swag-30 and model qwen2.5:7b.

Correction, 8 October 2026: the initial instructions used a random-card button. This version links the specified August record directly, so the 0.3 value refers to that record.

What the checks mean

A content-derived record ID helps detect a changed record body. A cryptographic signature binds the signed content to an identified key. Altering a score while retaining the original signature should therefore change the integrity outcome.

A valid signature does not establish that the measurement is true, current, reproducible, safe or independently certified. This exercise does not rerun the August measurement. Interpreting any signed record also requires understanding the signing key, signed scope and verification method.

Keep INVALID separate from UNCHECKABLE: a failed integrity check and a check that cannot complete are different outcomes. Report which individual check failed or remained unavailable rather than treating every problem as the same result.

For context, read the methodology and refutation ledger. A linked methodological claim still needs its own evidence; this challenge does not certify those pages.

Tell us what you could reproduce

Share the original and edited outcomes, what was confusing, and any reproducible failure. Corrections are useful results. Please do not paste confidential records into public replies.

Have a public claim whose evidence changes? Tell us the claim, its public source and the decision you need to support. This is a qualification enquiry; any paid pilot needs separately agreed delivery, scope, price and acceptance terms.

Published by CSOAI with AI-assisted preparation. The browser self-test described above was performed by our operator workflow; independent user verification remains to be established.