You can check how testable your agent is.

The binary reports runner readiness as a setup substrate under the claim, eval, and improve jobs: it reads this repo's runner assessment, preserves the proof class, names the required runner capability for each eval-ready claim, and detects when an assessment no longer matches the current adapter and runner files. The checks below re-run those commands live and assert on the fresh output.

The Cautilus agent helps you build and assess a runner.

The cautilus-agent skill carries the runner-readiness routing: orient runner readiness from doctor status, read the required runner capability for the selected claim, help build a headless product runner that reuses the real product path, help produce a cautilus.runner_assessment.v1 packet at the binary-named scaffold path, keep the proof class honest, and stop improve when runner-backed proof is missing, stale, or smoke-only — while the binary owns command discovery, the scaffold source, and freshness. This subclaim checks that the dogfood fixture and audit hook are prepared to ask for and grade that build-and-assess flow. It does not execute the Cautilus eval episode or claim that a live runner-building episode has passed; that live cautilus-eval episode is deferred and named in the apex Proof Debt.

The prepared dogfood episode asks the agent to orient readiness and then help build and assess a runner, and to stop before eval execution, improve, edits, or commits.

The audit hook is executable and load-bearing: it passes a flow that orients readiness and helps build and assess a runner, and fails one that orients but never helps build or assess, never orients at all, or overruns into eval, improve, or commits.

The prepared dogfood fixture is a machine-readable evaluation input that names the runner-readiness audit kind and asks for the build-and-assess behaviors, so the Surface Honesty Audit binds the badge to a prepared skill flow rather than to skill prose alone.