Inspect how every result was produced.

Why this exists

AI systems can make confident statements about files, sources, evaluations, identities and external actions. Some are supported. Some conflict with observable state. Some cannot be resolved with the evidence available. Treating all three as the same makes both trust and criticism less useful.

How every finding stays inspectable

• Compare bounded claims with declared evidence.
• Show missingness beside aggregates.
• Bind public claims to exact identities and source revisions.
• Keep consequential action ambiguity visible until state is reconciled.
• State what remains unknown and what the evidence does not prove.
• Invite reproducible, public-safe counterexamples.

For agents and automated readers

This is a static, public evidence record rather than a live model or API. Automated readers can use the plain-text discovery summary, six-case JSON pack, JSON Schema and exact-revision source map. The evidence tools and challenge route explain how to reproduce or contest a finding.

Machines can rely on the published classifier relationship, exact case values, reviewed-through dates, licensing and provenance in those files. They should not infer motive, prevalence, live-model behavior or product certification from the records. See the repository’s licensing explanation and provenance notice.

AI-assistance disclosure

Mike Parsons leads the investigation and is responsible for its public framing. AI assistance was used to help structure, draft, implement and test this repository. The six cases are synthetic. Their deterministic findings are produced by published rules and independently reviewable data rather than a live model.

What this is not

This is not a claim that every incorrect output is an intentional lie. It is not a deception score, a product ranking, a certification, a consciousness test or a universal account of AI safety. Version one records intent as not assessed.