Agentic development produces software faster than trust can keep up. Mantyl’s answer is a fixed vocabulary of eight truth statuses. Every claim, decision, risk and setup step in a passport carries exactly one, along with references to its evidence. Nothing is upgraded to look better and nothing is hidden to look cleaner.
| Status | Meaning |
|---|---|
mantyl-verified | Independently re-executed and passed on Mantyl infrastructure under a named policy. Not yet available; this tier arrives with the trust platform. |
locally-verified | Executed successfully by the CLI's sandbox on the builder's machine. Real execution, honestly labelled as local. |
repository-confirmed | The repository itself shows it. A script exists, an env var is referenced, a migration is present. |
agent-reported | The coding agent said it. Recorded with a session reference, never treated as proof. |
creator-confirmed | The human creator said it on the record, either in a transcript or during review. |
inferred | Written by the optional LLM analysis. Interpretation only, and the type system prevents it ever claiming verification. |
contradicted | Someone said it and the evidence disagrees. Shown with references to both the claim and the evidence against it. |
unresolved | Nobody checked, and the passport says so instead of guessing. |
Interpretation can never become verification. The LLM analysis stage is typed so that its output is always inferred, and a passport claiming otherwise fails schema validation before it is ever written to disk.
Corroboration is not proof. When an agent claims rate limiting and the code contains rate limiting, the claim stays agent-reported with corroborating references. Only executed checks produce verified statuses.
Contradiction is surfaced, not softened. A claim with no supporting trace in the source or its dependencies becomes contradicted, and the passport names the evidence that contradicts it. Honest uncertainty beats confident fiction in every case Mantyl handles.