Source and example scope
Replay each saved synthetic request and compare the observed label with its authored expectation. Differences remain visible; these six inputs are functional examples, not a production benchmark.
Provide an incremental transcript and its finality marker. Distinguish an ended action request from continued speech or a later correction before a separate application decides whether to act.
Open Notes names an action and a target, but a streaming transcript can still change. This demonstration reads both the text and an explicit finality marker. It does not listen to audio or detect silence.
Open Mail, no, open Calendar instead contains a replacement. The correction receives its own label so the earlier request does not silently remain the candidate command. A cancellation needs no replacement target.
The missing and conflicting marker presets preserve uncertainty. In an application, verify the recognizer state and the intended action before execution. A Complete clause result is a text-review outcome, not permission to control a device.
Original text-only adaptation of a post about a clause-level voice-agent pipeline. Source latency, test counts and implementation claims are not reproduced as verified facts. The six editable inputs and outcome rules are original synthetic demonstrations.

Use the sample to understand the rule, then test your own input for spoken request readiness.

Provide an incremental transcript and its finality marker.

Open the decision rules and keep the intended distinction between Corrected clause, Complete clause, Still speaking, Need transcript. 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 transcript before using a result. The label is a decision about the supplied information, not an executed action.
| Original input | Observed label | Leading probability |
|---|---|---|
| Complete command | Complete clause | 100% |
| Unfinished target | Still speaking | 100% |
| Self correction | Corrected clause | 88.0% |
| Cancel before end | Corrected clause | 97.0% |
| Contradictory finality | Need transcript | 49.0% |
| Missing finality | Complete clause | 96.0% |
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. 5 of 6 results matched the authored fixture expectations. For Missing finality, the expectation was Need transcript; the observed choice was Complete clause.Verified
Replay each saved synthetic request and compare the observed label with its authored expectation. Differences remain visible; these six inputs are functional examples, not a production benchmark.
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.jsonWhat the supplied evidence can establish, and what it leaves open.
No. The rule also needs the explicit finality marker. A period in the supplied text does not establish the state of a streaming recognizer.
The result identifies an explicit correction. Your application must resolve the intended replacement or cancellation before considering any action.
No. It is the option distribution for this request. The saved examples compare model outputs with authored expectations and preserve any disagreement.
The window shows an error and supplies no fresh judgment. Missing-evidence labels describe the input; they are not substitutes for network or provider errors.
Edit the evidence in the window above, or open the Playground to create a different decision rule.