GEThttps://core-api.heysadie.ai/transcriber-banks/{id}/turns

Turn id

The organization that owns the turn

The tenant the call belongs to

The call the line is from

The assistant that took the call. Null when the call names none

Index of the line in the call's messages, as GET /turn-flags/calls/:callId/lines lists it

The bank the turn belongs to. Null while it is in no bank, and then no candidate replays it

flagged: a person flagged the line as wrong and typed what was said. control: an unflagged caller line of the same call, sampled with a flag, whose expected text is the live transcriber's own. full_call: a caller line replayed by a full-call run, listed under that run

What the live transcriber wrote for the line on the call, frozen when the turn was created

What the caller said, word for word: typed for a flag, the heard text for a control. A full-call turn carries the expected text of a flag on the same line, else null

Tags on what went wrong, e.g. `name` or `spelling`. Empty on controls and full-call turns

The flagger's note. Null when none

The line's place in the call recording. Transcribers hear it with 0.75 s more on each side, so play `startSeconds - 0.75` to `endSeconds + 0.75`, clamped to the recording, to hear what they heard

pending: candidate attempts are queued and not all reported, or the turn was created in no bank. scored: every transcriber its job asked for returned an attempt (a failed attempt counts), or it joined a bank that owed it no attempt. failed: its job ended without an attempt from every transcriber it asked for. A turn taken out of a bank keeps its evaluation, and a backfill does not move a scored turn back to pending: watch `evaluationProgress` for that work

On a control, the flag it was sampled for. Deleting or moving that flag deletes or moves the control. Null otherwise

On a full-call turn, its run. Null otherwise

When two or more candidates match the expected text with different raw text (casing, punctuation), the one an LLM judge found fits the venue best. Null otherwise

On a full-call turn where the candidate differs from production, its label: decided by the expected-text matches where the line has expected text, else by the LLM judge. candidate_right: the candidate has what the caller said and production does not. production_right: production has it and the candidate does not. same_meaning: they differ in a way that does not matter. both_wrong: neither has it. Null otherwise or until the analysis labels it

On a full-call turn, a person's verdict from POST /full-call-runs/:id/reviews, in the same values as `label`. It counts only on differing turns, and outranks the label in `fixed` and `introduced` only where the line has no expected text. Null until reviewed

On a full-call turn, whether production is also replayed on it to measure replay against the live text. False otherwise

Who made it, as the creating request named them in an `X-On-Behalf-Of` header. Null when it sent none

When it was created, UTC

When it last changed, UTC

The transcriber the turn's call ran on, recorded when the turn was created. Its stored text is the turn's production attempt. Null when unknown

The newest open transcribe job holding the turn, for a progress bar. Null when no open job holds it (scored, failed or never queued), and on the POST /turn-flags response, which is written before its job is queued: read the turn again to see it

Tenant id

The tenant's name

Assistant id

The assistant's name

The turn the attempt is on

The transcriber that made the attempt

candidate: a transcriber under test, replayed on the clip. production: the live transcriber's stored text from the call, never replayed, at no cost. calibration: the live transcriber replayed on the clip, on a full-call run's calibration sample

Provider and model, e.g. "deepgram/nova-3". Display only

The transcript. Null when the attempt failed

Why the attempt failed. Null when it succeeded. A failed attempt counts in `errors`, not `scored`

Whether the text equals the expected text after normalizing (lowercase, punctuation dropped except in-word apostrophes, whitespace collapsed; digits and number words stay distinct). Null when the turn has no expected text or the attempt failed

Word error rate against the expected text: word edits divided by the expected word count, 0 for a match, can exceed 1. Null when there is no expected text or text

On a match, how the raw text still differs from the expected: `case`, `punctuation`, or both. Empty otherwise

On a candidate or calibration attempt of a full-call turn, whether its text equals production's, normalized. Null for production attempts, flagged and control turns, and failed attempts

Seconds of audio the transcriber processed for the attempt. 0 for production attempts, which replay nothing

The bank's candidates whose attempt matches, by position

When several match with different raw text, the candidate the judge picked. Null otherwise

The judge's reason for the pick