Policy, tools and labels

Three authored input populations have different jobs.

Tools catalogue: a tools array, or an object containing one, lists exact tool names with optional descriptions and MCP annotations. The scanner uses annotations before its name/description heuristic and treats unknown effects conservatively. Missing routes cannot be discovered from an incomplete catalogue.

TrustedConfig: epoch and safety are required. Safety contains approval configuration and exact-name tools rules. Allow, guard and deny differ: an allowed mutating tool is still ungated; a guard requires a nonempty target. Optional temporal, consensus, convergence, calibration, linear and budget sections have their own enabled/participation semantics. Safety cannot be disabled. Legacy top-level rules are not the current input shape.

Adequacy labels: declare unique monitor names and finite states. Each state needs an identity, label and evidence value for every monitor. Missing evidence is malformed input. Equal evidence with different labels is a collision; a single label can be vacuous.

Setup manifests additionally need server identity and are not interchangeable with bare scanner catalogues. See the generated example.

The owner retains the complete normative input schema guide. Its numerical warning describes a verifier limitation: the Protect contract itself permits finite decimals in its supported range.

Previous: Assurance CLI reference. Up: Assurance CLI. Next: Verification profiles.