Human authority
Red Zone actions require explicit human approval. Agents cannot mint, return, replay, or approve their own authority token.
Valdris.aiSecurity
The harness fails closed when source identity, artifact integrity, credential separation, review independence, or human authority cannot be established.
Red Zone actions require explicit human approval. Agents cannot mint, return, replay, or approve their own authority token.
Evidence binds the commissioned repository, revision, and route. Work from a different source identity cannot satisfy the run.
Bridge, reviewer, and execution credentials remain distinct. Child processes do not receive finish-line credentials.
Evidence files must remain within the commissioned artifact root. Symlink escapes, mutation, and deletion fail closed.
The delivery identity cannot be every reviewer. Review applicability cannot be silently downgraded.
Authoritative assurance depends on protected execution, trusted signers, provider state, or another commissioned authority outside the delivery agent.
Public release surfaces can be scanned without printing the private restricted-value list. The authoritative workflow must run on a protected default branch with self-review prevention, an independent required reviewer, and pinned unattended identity. Without those controls, the result is operator-supervised evidence, not release authority.
A test pass does not authorize deployment, traffic movement, secret rotation, data migration, app-store promotion, destructive cleanup, failover, or residual-risk acceptance. Those actions remain human decisions.