Verification
PreviewWhy Pimsy does not assume its first answer is correct.
Verification is the difference between an agent that produces output and an agent that produces results. Every step can declare a check; the task cannot complete while a required check is failing.
Verification strategies#
Verifier agents#
For high-stakes work the orchestrator spawns a verifier with read access to the artifact and the trace but no authorship stake in the result. Its instruction is to find the strongest reason the result is wrong.
producer agent ──▶ artifact ──▶ verifier agent
│
┌────────────────────┤
▼ ▼
issues found no issues
│ │
▼ ▼
revision step mark verified
│
└──▶ re-verify (max N rounds, then escalate)Reporting uncertainty#
When verification cannot be completed — a source is offline, a test environment is unavailable, evidence remains contested — Pimsy returns the result marked unverified with the reason attached. It does not upgrade an unverified result to a confident one.
{
class="tok-str">"status": class="tok-str">"completed_with_caveats",
class="tok-str">"verification": {
class="tok-str">"passed": false,
class="tok-str">"checks": [
{ class="tok-str">"name": class="tok-str">"tests_pass", class="tok-str">"result": class="tok-str">"pass" },
{ class="tok-str">"name": class="tok-str">"citations_resolvable", class="tok-str">"result": class="tok-str">"partial", class="tok-str">"detail": class="tok-str">"class="tok-num">2 of class="tok-num">14 sources unreachable" }
]
},
class="tok-str">"confidence": class="tok-str">"medium",
class="tok-str">"caveats": [class="tok-str">"Two cited sources could not be re-fetched at delivery time."]
}Last updated 2026-09-14

