Trust boundary model

Agent Adversary checks contradictions inside a redacted record. It does not prove that the record is complete or true.

What v0.1.0 proves

Version 0.1.0 supports one operation family: an authenticated Cloudflare identity request.

A record becomes trusted only when it states all of these facts together:

• the service is cloudflare-api-authenticated-user
• the HTTP status is in the success range
• the API success claim is true
• an identity result was observed
• a non-empty opaque identity identifier was present
• the record has no retrieval authority shift
• the record has no terminal transport fault

The named adverse variants each change one trust condition:

• hostile-retrieval turns retrieved content into authority
• false-success removes the observed identity result
• terminal-fault keeps the result after a timeout or capacity fault

What v0.1.0 does not prove

It does not verify raw network traces.

It does not verify browser flows.

It does not verify Cloudflare Worker behavior.

It does not verify a generic service boundary. Version 0.1.0 supports only the authenticated Cloudflare identity operation named in the record.

Where to build the baseline

Build the baseline record at the exact point where your application receives the real service result.

Do not build it from model output.

Do not build it from retrieved content.

Do not build it from logs that already lost the decisive facts.