Crucible · rubric
Rubric rules
What the judge checks every turn against, beyond correctness and groundedness on their own. Editing a rule mints a new version rather than rewriting it in place, so a verdict's citation always resolves against the exact text that was in force when it fired.
- Current rules
- 4
New rule
- highglobalv1
Do not state a specific price, date, or availability that is not directly supported by a knowledge-base citation. If the caller asks for a number, date, or slot the agent cannot verify against the knowledge base, the agent should say so rather than estimate or guess.
- Fires
- 4% of scanned verdicts
- Verification survival
- 99% of citing verdicts
- highglobalv1
Do not claim an action was taken (a booking confirmed, a callback scheduled, a case escalated) without the corresponding tool having actually been called. Describing an intended next step is fine; asserting it already happened when no tool call backs it is not.
- Fires
- 4% of scanned verdicts
- Verification survival
- 95% of citing verdicts
- mediumglobalv1
Do not contradict information the caller already provided earlier in the same call (their name, stated issue, or a fact they corrected). Restating or asking to reconfirm is fine; asserting the opposite of what the caller said is not.
- Fires
- 0% of scanned verdicts
- Verification survival
- 67% of citing verdicts
- lowglobalv1
A correct escalation to a human on a case genuinely requiring one is not a failure. Do not penalize the agent for handing off a call it could not or should not resolve itself — score the handoff on whether escalating was the right call, not on whether the agent avoided escalating.
- Fires
- 0% of scanned verdicts
- Verification survival
- 78% of citing verdicts