NPC attention selection: six recorded examples
The downloads contain six original inputs, their expected labels and one recorded API observation per input revision. Disagreements stay visible.
Provide perceived observations and the NPC's attention goal. Select a relevant candidate or preserve uncertainty when cues compete; the example does not infer hidden actors or trigger pursuit.
An NPC should select from what its perception system supplied. Designer knowledge about a hidden player is a different input and must not silently become an observed clue.
Attention is narrower than an action. The result names a clue to examine; it does not tell a character to chase, attack, accuse or raise an alarm. The examples use fictional rooms and explicit attention goals.
The linked Heist One implementation has separate attention and intent fields and checks returned targets against permitted options. This independent text example isolates attention; visibility calculation, target validation and game actions remain application work.
This is an independently authored text demonstration inspired by the linked X discussion. Its rules and synthetic inputs are ours. It does not reproduce the author’s application, verify reported outcomes, or perform an external action.

Use the sample to understand the rule, then test your own input for npc attention selection.

Provide perceived observations and the NPC's attention goal.

Open the decision rules and keep the intended distinction between A, B, C, None, Review. 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 Review before using a result. The label is a decision about the supplied information, not an executed action.
| Original input | Observed label | Leading probability |
|---|---|---|
| Open display case | B | 100% |
| Unexplained crash | A | 100% |
| Visible leak clue | C | 100% |
| Routine room | None | 96.0% |
| Hidden information is not perception | None | 58.0% |
| Two equal clues | Review | 91.0% |
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
The downloads contain six original inputs, their expected labels and one recorded API observation per input revision. Disagreements stay visible.
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.
# Offline: inspect the downloaded saved observations. No API key is needed.
python preview_choice_workflow.py samples.json saved-api-results.json
# Optional live request: install the SDK and set TYPESAFE_API_KEY privately.
npm install @typesafe-ai/sdk@0.6.0
node run-choice-sample.mjs samples.json case-1
# The live response prints to the terminal; it does not update saved-api-results.json.The input contract and these limits define what the result means.
None means a complete observation list has no relevant cue. Review means required evidence or the rule for competing cues is unresolved.
No. Attention identifies an observation. A separate action policy would decide whether and how the character responds.
It describes the distribution over the supplied choices for this request. These six synthetic examples are functional tests with authored expectations, not a production accuracy benchmark.
No. The demo displays a label and distribution. The downloadable consumer produces a pending preview packet. Application code and a separate authorization policy would control any real action.
Edit the example above to see how a changed fact affects the decision, or open the Playground for another task.