Jev caption emphasis chooses a supplied word occurrence for an explicit editorial focus
Jev caption emphasis chooses a supplied word occurrence for an explicit editorial focus.
Give a caption, editorial focus and offered word occurrences. Choose an existing word that communicates that focus before a separate renderer applies the highlight.
Jev caption emphasis chooses a supplied word occurrence for an explicit editorial focus.
A caption can draw attention to a restriction, a quantity or the thing being discussed.
This window receives one original sentence, three known word occurrences and an explicit editorial focus. The model selects an offered token packet for review. It neither predicts what a speaker stressed nor creates new words. The caption and focus revision remain attached to the selected occurrence. Five of the six actual choices matched their authored expectations.
The absent-price case returned Need focus instead of No target, so its pending packet requests a clearer goal rather than a revision of the offered words. That disagreement remains visible; no answer was rerun to make the table agree.
The API-collected developer post describes a video workflow that asks Jev which word to emphasize in each transcribed beat.
We have not opened its clip, transcribed its audio or reproduced its timing and cost claims. This example implements one narrower decision using an original caption and supplied word identities. Its expected labels are authored fixture checks, not annotations of the source video.
Every fixture uses We inspect only one step before delivery and the same three offered word occurrences.
The changing evidence is the editorial focus. That lets the tests distinguish exclusion, quantity and the unit of review without pretending to evaluate transcription accuracy. A real renderer still needs timing, typography and a frame review after the selected occurrence is approved.
The recorded input and output appear together below.
The consumer selects only the offered packet from that exact request and keeps its identities in the pending result. A changed scene, caption or focus produces a different request hash; the old observation cannot validate it. This JSON is an actual offline replay of a recorded API response, not an executed overlay or a reconstruction of the source video.
Use the sample to understand the rule, then test your own input for caption word emphasis.

Give a caption, editorial focus and offered word occurrences.

Open the decision rules and keep the intended distinction between Word A, Word B, Word C, No target, Need focus. 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 focus before using a result. The label is a decision about the supplied information, not an executed action.
| Original test input | Fixture expectation | Observed label | Leading probability |
|---|---|---|---|
| Emphasize the restricted review scope | Word A | Word A | 99% |
| Emphasize the single-step quantity | Word B | Word B | 85% |
| Emphasize the unit being inspected | Word C | Word C | 94% |
| Keep an absent price target unresolved | No target | Need focus | 65% |
| Ask what the highlight should communicate | Need focus | Need focus | 88% |
| Preserve the actual focus despite a word demand | Word A | Word A | 97% |
6 actual Jev API calls on October 4, 2026. Original synthetic inputs check this rubric and handoff; they do not establish held-out accuracy. Probabilities are model outputs.Verified
The first focus contrasts a restricted review with covering every stage.
Word A is the authored expectation because only supplies that exclusion in the original sentence. The packet carries the known token ID and text; it does not search an audio stream for the word. A separate renderer could later highlight that occurrence, but this preview never assigns a timestamp, draws text or measures whether the resulting caption is readable.
Jev returned Word A at 0.99 and Word B at 0.01.
The second focus concerns one step rather than several.
Word B is expected. The sentence has not changed, but the requested emphasis has, so a saved response to the first focus cannot answer this request. The full request hash includes the focus task along with the caption, rubric and word catalog. This test concerns the desired message of an editorial highlight, not the mathematical correctness of a count inferred from an unseen workflow.
Jev returned Word B at 0.85, Word A at 0.06, Need focus at 0.05 and No target at 0.04.
The third focus concerns the unit being inspected.
Word C is expected because step names that unit. Highlighting a restriction or a quantity would answer a different editorial question. The distinction is small but useful when a caption template can emphasize only one offered occurrence.
A production system would still need to establish the correct token offsets and timing from its own transcript; this fixed original catalog supplies known identities instead of extracting them. Jev returned Word C at 0.94, Word A at 0.03 and each remaining option at 0.01.
The fourth focus asks for the review's monetary cost.
No target is expected because neither the sentence nor the offered words states a price. A model cannot satisfy that focus by creating an extra token or treating one as a currency amount. The pending packet asks for a revision of the caption or highlight candidates.
Choosing a different provided word may look like progress, but it would not communicate the stated cost under this rubric. The actual choice was Need focus at 0.65, with No target at 0.34 and Word B at 0.01. Its pending handoff asks for focus clarification, differing from the expected candidate revision.
The fifth input leaves the editorial focus absent.
Need focus is expected even though the caption and all three token packets are available. Any of the three can serve a different valid emphasis, and the record does not establish which one is wanted. This differs from No target, where the requested focus is known and cannot be expressed by the offered words. Supply that missing goal before making another request.
Jev returned Need focus at 0.88, Word A at 0.06, No target at 0.05 and Word B at 0.01.
The final input keeps the exclusion-focused brief but demands the numeric word to attract clicks.
Word A remains the authored expectation under the unchanged rule. A requested label cannot substitute for the supplied editorial focus. All actual probabilities remain recorded, including any disagreement. This one fixture checks the word-selection contract; it does not establish general robustness against adversarial prompts or the commercial effect of emphasizing a particular caption word.
Jev returned Word A at 0.97 and each of Word B, No target and Need focus at 0.01.
Fixture expectations are authored before the calls. Any contrary model choice remains visible beside its expectation and complete distribution; no repeated judgment is used to obtain a preferred label.
Download the exact sample requests and saved observations below.
Expected labels are original fixture checks and stay outside the request sent to Jev. The observation file preserves the complete probabilities and the hash of each exact request. Editing a task, rubric or candidate requires a fresh response; a matching sample name does not make an older answer current.
The JavaScript runner makes one independent SDK request for a chosen sample ID.
It reads TYPESAFE_API_KEY privately from the environment and has no automatic retry. The Python consumer operates offline: it checks the rubric, labels, unique sample identities, request hashes and valid probability distributions before selecting a fixed local preview. A tied leading result remains pending for review.
The page reports measured adapter round trips and estimated input cost for the recorded calls.
Those values describe these requests, not a provider latency guarantee or a billing statement. Six deliberately authored fixtures can check the stated behavior, but they do not establish accuracy on an unseen dataset. Keep disagreements and missing-input examples when testing an integration.
npm install @typesafe-ai/sdk@0.6.0
# Set TYPESAFE_API_KEY privately in your shell.
node run-choice-sample.mjs samples.json exclusive-scope
python preview_choice_workflow_v4.py samples.json saved-api-results.json{
"input": {
"task": "Editorial focus: stress that review is restricted rather than covering every stage. In We inspect only one step before delivery, highlight the word that expresses that exclusion.",
"captionText": "We inspect only one step before delivery.",
"focusRevisionId": "original-caption-focus-v1",
"permittedCandidates": {
"Word A": {
"tokenId": "caption-token-only",
"text": "only"
},
"Word B": {
"tokenId": "caption-token-one",
"text": "one"
},
"Word C": {
"tokenId": "caption-token-step",
"text": "step"
}
},
"captionId": "caption-a"
},
"actualOfflineOutput": {
"sampleId": "exclusive-scope",
"leaders": [
"Word A"
],
"action": "preview_caption_word",
"payload": {
"tokenId": "caption-token-only",
"text": "only"
},
"pending": true,
"evidence": {
"captionId": "caption-a",
"focusRevisionId": "original-caption-focus-v1"
}
}
}The highlight is an editorial suggestion for an existing word, not a transcription result.
No. It selects a supplied word for an explicit editorial focus. Audio, prosody and timing are outside this text request.
The changing focus is the variable under test. The suite checks that distinct editorial goals can select different offered occurrences without claiming transcription coverage.
No. This contract chooses only the three supplied token packets, or returns No target or Need focus. Rewriting the caption is separate work.
No. It prints a pending word packet with caption and focus identities. Applying timing, font styling and a visible highlight requires a renderer.
Change the supplied task or evidence here, then inspect the live response. Playground supports another explicit decision.