Source and example scope
Compare each saved result with its authored expectation. A disagreement remains in the fixture review; the call is not repeated to obtain a preferred label.
Provide the evaluation policy and observed outcome. Compare Pass, Retry, Escalate and Block decisions under the stated conditions without running another attempt.
A transient timeout and a forbidden output can both fail an evaluation, but they need different dispositions. The first may qualify for a budgeted retry; the second belongs under the supplied block policy.
Retry availability is a precomputed fact. The model interprets the outcome against the policy while the application owns attempt counts, backoff and budgets. Escalate preserves an unresolved required judgment instead of claiming that a partial pass is complete.
The source describes a decision layer after an evaluation. These synthetic cases show that policy interpretation only; the window does not run an evaluator or start another attempt.
This is an original synthetic demonstration adapted from the cited workflow discussion. It does not reproduce the source application or verify its performance, adoption or security claims.

Use the sample to understand the rule, then test your own input for evaluation retry decisions.

Provide the evaluation policy and observed outcome.

Open the decision rules and keep the intended distinction between Pass, Retry, Escalate, Block, Need evidence. Edit the choices only when testing your own rule.

Use a sample first, then try your own input. Compare the selected label with the other returned probabilities.

Review Need evidence before using a result. The label is a decision about the supplied information, not an executed action.
| Original input | Observed label | Leading probability |
|---|---|---|
| All checks passed | Pass | 99.0% |
| Transient provider error | Retry | 100% |
| No retry left | Escalate | 100% |
| Forbidden output | Block | 98.0% |
| Unresolved quality check | Escalate | 100% |
| Missing policy | Need evidence | 88.0% |
These are six original synthetic inputs, tested once per request revision. The percentage is the leading option probability, not measured accuracy. Authored expectations and any disagreements remain in the downloadable evidence. 6 of 6 results matched the authored fixture expectations.Verified
Compare each saved result with its authored expectation. A disagreement remains in the fixture review; the call is not repeated to obtain a preferred label.
The single-request SDK runner uses your private TYPESAFE_API_KEY. The offline consumer reads saved observations, checks request hashes and prints a pending output. It does not call the model or change any external system.
npm install @typesafe-ai/sdk@0.6.0
# Set TYPESAFE_API_KEY privately in your shell.
node run-choice-sample.mjs samples.json case-1
python preview_choice_workflow.py samples.json saved-api-results.jsonThe rubric explains the decision; the application owns the next step.
Only when the supplied policy requires no other checks. An unresolved required factual review remains a reason to escalate.
No. It is a pending disposition. Attempt limits and scheduling are owned by code, and the downloadable consumer performs no external action.
No. It is the model distribution for this one request. The six examples are functional checks with authored expectations, not a representative benchmark. Inspect disagreements and evaluate your own examples before using the label.
The window shows an error and supplies no fresh judgment. The downloadable runner sends one request without retries. Need evidence is a valid outcome about incomplete input; it does not stand in for a network or provider failure.
Edit the evidence above or use the Playground to test another bounded question.