Hugging Face's logo

Join the conversation

Join the community of Machine Learners and AI enthusiasts.

Sign Up
Back to feed
Nicholastempleman 
posted an update 2 days ago
Post
71
Can you catch a one-character change in an AI measurement record?

Open this specific historical record in CSOAI's free browser verifier: https://councilof.ai/gspc-verify/?card=%2Fsigned%2Fcards%2Fa31be261cd078a7b5eb42dae5e3b6794c996efc83be20dbe5df7448c4a5c7bbe.json

The verifier automatically loads and checks this record. Inspect its recorded 19 August date and original outcome. Then change accuracy from 0.3 to 0.4 without changing its signature, and verify again. Our own browser check on 8 October returned VALID for the original and INVALID for the altered copy. No public tally was submitted.

The record is from 19 August. Signature validity establishes integrity under the identified key, not measurement truth, freshness, model safety or certification. Independent participant results are still to be established.

Steps and limits: https://huggingface.co/blog/Nicholastempleman/signed-ai-record-integrity-challenge

Tell us what you could reproduce and what was confusing; report checks that cannot complete too. Please do not paste confidential records into public replies.

CSOAI update, prepared with AI assistance.
dipankarsarkar

Reproduced it outside your verifier, a few lines of Python on your preimage_rule: original VALID, accuracy 0.4 INVALID. The id field catches it too, it is sha256 of the preimage.

The key pins. d4cb0eaa... is #card-attestation-1 in did:web:csoai.org.

The confusing part is one step over, in the DID document itself. Your root and its councilof.ai mirror serve different _gspcBoardKeyNote text today.

The mirror says single-key custody.
csoai.org, the copy a did:web resolver actually reads, still says 3-party Coinbase cb-mpc, "the private scalar never exists as a whole number". That is the wording C-2026-0925-01 retracted on 25 Sep.

All seven keys are identical in both files, so a split-brain check that compares keys would pass this. Does the C-2026-0819-05 check diff the whole document, or only the canon keys?
In this post