OpenAI Build Week · Work & Productivity
The visit broke.
The thread didn’t.
Break a synthetic care-team handoff on purpose. Continuity Relay asks GPT‑5.6 to preserve the patient context, clinician intent, and one next safe workflow step—then exposes every source and receipt a reviewer needs to challenge it.
01Preserve contextOnly registered source facts
02Carry intentWhat the clinician meant to finish
03Name one next stepWorkflow action, never treatment
04Repair the breakOwnership and evidence stay visible
Failure injection workbench
Make the handoff fail where a judge can see it.
Choose one frozen synthetic case. The browser sends only its ID and the failure mode.
1 · Choose case
Frozen synthetic fixtures
Ready. No model call has been made.
Ground truth
Voice intake dropped mid-handoff
Synthetic patientSYN-RP-204
Visit stateClinician review pending
Clinician intent
Continuity packet
No theatrical fallback.
If the live model or validator fails, this surface says so. It never swaps in a hidden canned answer.
Handoff summary
Grounded source IDs
Judge the boundary, not the adjectives
The model proposes. Deterministic code owns truth.
- BrowserscenarioId + failureModeNo free text. No patient data.
- Serverloads frozen factsHashes the grounded input.
- GPT‑5.6returns strict JSONResponses API, store: false.
- Validatorchecks every sourceUnknown citations fail closed.
- Humanreviews the proposalNo autonomous care action.