Relay's run view: two claim actions, each confirmed applied by read-back, with a version trace and audit timeline.
A run in the hosted demo. Both actions confirmed by read-back.

All 12 hosted eval scenarios pass, including timeouts before and after the write.

Layman's Version

Relay is where the Corrobo idea started. It runs an AI agent's actions against a fake claims system, injects failures (timeouts after a write, timeouts before one, stale versions, duplicate requests) and shows what happens next.

After every attempt it reads the record back and decides: done, safe to retry, replan, ask a person, or stop and investigate.

Technical Version

Flow

  1. execute
  2. read back
  3. reconcile
  4. decide

Failures are injected around a synthetic source-of-record adapter. Every run persists its trace, so the review queue and the eval page show the same evidence the decision was made on.

Why it's frozen

I kept Relay unchanged as a reference and built Corrobo separately as a reusable library, so the demo stays exactly what it was.

Relay's review queue for actions that need a person.
The review queue.

Limitations

  • Built as a demo, not production software. The hosted version is read-only.
  • No real claims system or insurer integration.