promptdojo_

The tool contract — make the dangerous call a validation error — step 4 of 7

Four calls hit the validator during one triage run. The model proposed the first three; call 4 is the rejected rerun re-sent after on-call clicked approve — the harness stamped confirmed: True on it, because that flag comes from a human or an outer policy, never from the model. Walk each call through the checks in order (unknown tool → required fields → allowed values → confirm) and predict the four output lines.

Watch the second call: the model invented severity p4. And note the checks short-circuit — the first failed check names the reject.