Sandbox
Test keys (ak_test_…) run the whole lifecycle end to end. The preview is real; the physician is simulated; nothing is charged.
What's real and what's simulated
| Step | With a test key |
|---|---|
| Validation and limits | Real. Same rules and errors as live. |
| Identifier screening | Real. Identifying requests get 422 phi_detected. Don't send real personal information to the sandbox either. |
| AI preview, references, quote | Real. The quote shows what the request would cost live. |
| Reviewer selector | Nonempty specialty or physician selections are rejected. The simulator always reports has_eligible_experts: true, regardless of the live roster; sandbox never returns 422 no_matching_expert. |
| Escalation | Real lifecycle, including max_price_cents and Idempotency-Key. No production approval needed, no credit reserved, no charge. |
| Physician answer | Simulated, about a second after escalation. |
| Webhooks | Real, signed decision.answered / decision.canceled events with sandbox: true, sent to your test endpoints. |
| Physicians | Never notified. Sandbox work never reaches a real physician. |
Sandbox decisions have sandbox: true and are only visible to test keys. Cancellation works while the decision is still queued, which in the sandbox is about a second.
The simulated physician
"physician": {
"expert": { "id": "exp_sandbox", "name": "[SANDBOX] Simulated Physician, MD", "specialties": ["general"] },
"answers": [
{ "type": "predicate", "name": "unsupported_claim", "value": true,
"confidence": 0.6, "comment": "[SANDBOX] simulated answer comment: placeholder text so you can test handling of the comment field." },
{ "type": "choice", "name": "error_category", "choice": "unsupported_certainty",
"confidence": 0.6, "comment": "[SANDBOX] simulated answer comment: placeholder text so you can test handling of the comment field." },
{ "type": "score", "name": "error_severity", "level": 2, "label": "Material",
"confidence": 0.6, "comment": "[SANDBOX] simulated answer comment: placeholder text so you can test handling of the comment field." }
],
"answered_at": "2026-11-03T15:01:15Z"
}The simulator copies the AI preview's top answer for each question. These are simulator rules, chosen so you can test your handling of every field; they say nothing about how real physicians answer or rate confidence.
- Expert:
{"id": "exp_sandbox", "name": "[SANDBOX] Simulated Physician, MD"}. - predicate:
valueistruewhen the AIprobabilityis at least 0.5, otherwisefalse. - choice: the AI's
choice, with its original type. - score: the level with the highest AI probability (the lowest index wins a tie), not a rounded
score. - Every answer has
confidence: 0.6andcomment: "[SANDBOX] simulated answer comment: placeholder text so you can test handling of the comment field.". - AI refusal: a deterministic fallback selection (
falsefor a predicate, the first choice, or the first level) withconfidence: 0.0andcomment: "[SANDBOX] synthetic fallback after AI refusal". It is not a model or physician judgment. - Every answer carries a labeled placeholder
comment, so a request withinclude_commentary: trueis always satisfied.
confidence: 0.0 and comment handling, write a question the AI is likely to refuse. To test a false predicate, ask one the AI will answer "no".Going live
Live keys (ak_live_…) follow the same production activation as the rest of the developer platform: card on file, your own identity verification, and approval. See Go to Production. Escalating a live decision on an account without production access returns 403 production_not_active.