Credibility Inspector

This is the interactive UofA Credibility Inspector. Upload your model-credibility evidence and, in a few minutes, get one of two views off the same analysis: a plain-language Reviewer verdict for someone deciding whether to trust a finished package, or the Author (Gap-Finder) gap list for whoever is assembling the evidence. The summary is indicative, not a formal acceptance decision.

Privacy

Your documents leave the Space. They are sent to a hosted model (Together AI) to be read, then discarded. The Space itself stores nothing: each run uses a temporary directory that is deleted when it finishes, and no document is logged by us. But the text does go to a third party, and their terms govern what happens to it there.

If your evidence is confidential, do not upload it here - run the CLI on your own machine, where a local model reads it and nothing leaves your environment. Pasting a public model card sends only text that is already public.

If you generate a package, that file is held only long enough for you to download it and is deleted within 30 minutes.

Steps

1. Upload and detect. Drop your evidence documents, or try the bundled sample, then let the router detect the standard.
2. Confirm the standard. The router picks one from what it read. Switch it if it guessed wrong.
3. Confirm what we understood. The tool shows the credibility factors it read from your evidence. Correct any status it got wrong. That is the only thing you need to touch.
4. Read the result. You land on the Reviewer view (plain-language: what the model was for, what is evidenced, what is missing, the concerns that fired). Toggle to the Author view for the engineer’s gap list, or Save as PDF.
5. Take the package. “Download UofA package” gives you the assurance package itself, not just the report: a JSON-LD graph with its provenance, signed, alongside the rendered report and instructions for re-checking it.

Checking the package yourself

The zip is built by the same code path as the CLI’s uofa import, so what the demo hands you is what the command line would accept:

unzip uofa-pack-*.zip
uofa verify uofa.jsonld --pubkey keys/demo.pub

A valid signature means the file is unmodified since the demo produced it. It is not a review and not an acceptance decision. The key is a demonstration issuer key belonging to the demo, so a package from here is never mistakable for a formally issued one: it does not verify against the default trust anchor, and --pubkey is required on purpose.

The public key travels inside the zip so verification works offline. A trust anchor shipped inside the artifact it validates only proves the artifact is self-consistent, so check it against this copy, which the zip does not control:

keys/demo.pub  sha256:3605a146f4880d9f7a29db6ef5629655091d2ecd0c2b9919cbe49d90d65d83c8

shasum -a 256 keys/demo.pub

If those differ, do not trust the package.

uofa check additionally runs the structural (SHACL) rules. On a model card it will normally report missing bindings such as bindsRequirement or hasValidationResult: a model card is not a complete assurance package, and saying so is the point of the tool. Those findings are about your evidence. C1 Integrity is the check that speaks to the package itself.

⏳ First load can take up to a minute. The demo runs on a GPU that sleeps when idle to keep costs down. On a cold start you may see a "Building" or "Starting" screen below while it wakes, then the app appears. You can also open the demo in its own tab.

↻ Reload the demo

What it measures, and what it does not

The Credibility Inspector surfaces completeness and weakeners in plain terms. It does not issue an acceptance decision and it does not attest your model. For a per-factor write-up or a credibility-assessment consult, leave your email at the end.

Prefer the command line, or want the FDA-co-authored NAFEMS worked example with exact numbers? See the NAFEMS CLI reproduction.