Reference

A decision is what the evidence said at collection time. It is never rewritten. It carries two independent verdicts: whether the request may be made, and whether the text may be used for training.

An exclusion is a separate column applied every time an export is built. A quarantine keeps the text and stops using it. A takedown destroys it.

A signal is one contributing observation, recorded with a precedence integer and stored with the decision it contributed to.

An evidence hash covers what the origin said and nothing else. A decision hash additionally covers the engine revision, the parser versions and the outcome.

A grant is a dated, scoped, attributable permission naming an exact origin, a path prefix, a purpose, a rights basis, a reviewer and an expiry.

The crawler fetches six fixed paths per origin before it fetches any page. The list lives in code and does not grow per site.

absent, unavailable and wrong_content_type are three different facts. Collapsing them is how a prevalence number becomes fiction.

A markup response is never read as a policy document, whatever status code carried it. An HTML file parsed as robots.txt yields no rules, which would read as permission to crawl everything.

Signals carry a precedence integer. Robots access sits at the bottom. Risk findings sit at the top, above even the development bypass, because a planted credential is a safety problem.

Explicit denial wins. Ambiguity resolves to review, and nothing resolves upward toward permission.

Conflicts are reported. A person resolves them, and the taxonomy names the shape of each disagreement so there is something to resolve.

Failure codes: 24, in 3 kinds. The code is the contract, and messages change. Safety invariants enforced in code: 51. Decision policy version: 2026-08-15.1.