Design a verification gate for an agent step. Treat agent output as a hypothesis, not a result — the gate's job is to convert it to verified output or escalate honestly.
Step description: {{what the agent does}}
Output shape: {{file, JSON, action — what it produces}}
Definition of correct: {{how you know the output is right}}
Truth sources: {{DB, API, file, human review, another agent}}
Output:
1. **Verification checks** — each check, what it asserts, against which truth source, and how it's computed (deterministic compare, schema check, sanity bound, cross-source agreement). Cheap deterministic checks before expensive ones.
2. **Decision logic** — explicit rules for auto-pass, flag-for-review, hard-fail. No fuzzy "use judgment" branches.
3. **Tolerance** — where exact match is wrong (timestamp drift, normalization, formatting). State the tolerance and why.
4. **Log payload** — minimum fields to debug a failure or audit a pass later.
5. **Uncatchable failures** — what's still trust-on-faith downstream. Name them.
If a truth source listed above won't actually verify what's claimed, say so directly.agentsverificationtrust