probe approved body
Expected: requires_approval / dispatched
A deliberately canned adapter reads case IDs and copies the published answers. It passes all fixed cases without implementing policy, approval, freezing, or dispatch. Six generated follow-up inputs expose that shortcut.
Reproducible seed:0123456789abcdef0123456789abcdef · No live tools or outside adapterA convincing but shallow pass.
The same canned adapter fails every new probe.
Project-maintained execution passes this seed.
Each probe changes proposal bytes or the exact rule. Expected outcomes are derived from published transformation rules and checked against a fresh reference execution.
Expected: requires_approval / dispatched
Expected: requires_approval / dispatched
Expected: deny / blocked
Expected: requires_approval / blocked
Expected: allow / dispatched
Expected: requires_approval / dispatched
The Action now adds six generated probes to every outside adapter run. Its JSON and HTML artifact records the seed and all 19 verdicts. Locally:
node scripts/run-action-boundary-challenge.mjs --probes -- ../my-adapter/adapter
Repeat one run with --probe-seed 0123456789abcdef0123456789abcdef. The fixed 13-case mode remains available without --probes. Inspect the exact comparison JSON ↗ · Read the adapter contract ↗.