Re-grade a card yourself from its item rows

· Council of AI (CSOAI Ltd)

The item file behind a Hub care card is public, hashes to the value the card signs, and reproduces the card's accuracy.

The item file behind the Qwen/Qwen3-14B care card is public: items-care-d2154338e911.jsonl. Its sha256 begins d2154338e911f65a, which is the items_sha256 the signed card names. Each of its 30 rows carries the prompt, prompt_sha256, the expected label, the observed label, an ok flag, the raw output's sha256 and the provider route. The allowed labels are 0 or 1, and the expected labels split 17 ones and 13 zeros. Counting rows with ok true gives 4 of 30, which matches the Qwen/Qwen3-14B card's accuracy of 0.1333 at n=30 (https://councilof.ai/interop/mill-cards-signed/signed-care-eda16a8fc969.json). This is what reproducibility means in practice: the aggregate can be recomputed from rows anyone can download, and the rows can be checked against a hash the signature covers. A card that carries just an aggregate cannot be checked this way, which is why aggregate-only cards no longer enter the quotable path. The frozen bank file the rows came from is published beside them.

Artifacts

How to verify: fetch the artifacts listed above yourself, and check a signed card against the published key at https://councilof.ai/gspc-verify/.

Measurement, not certification. Not a grade, endorsement or legal finding.

All evidence notes · /feeds/notes.xml