GEThttps://core-api.heysadie.ai/full-call-runs/{id}

Run id

The organization that owns the run

The bank the run belongs to

The candidate replayed

The calls asked for

Random turns asked for on which production is also replayed

Calls the run snapshotted

What the run replays and costs, as estimated at creation

queued: waiting for a worker. running: the candidate is replaying the turns. analyzing: the LLM judge is labeling the differing turns and writing patterns and a verdict. completed: done. cancelled: stopped by POST /full-call-runs/:id/cancel. error: failed, see `error`. queued, running and analyzing can still change; poll until completed, cancelled or error

The differing turns outnumbered what the analysis takes (2000), so it labeled a sample

Why the run failed. Null unless status is error

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 its first job was dispatched to a worker, UTC

When it completed, failed or was cancelled, UTC

Calls with a turn in the run

Turns in the run, one per caller line

Turns whose candidate text equals production, normalized

Turns whose candidate text differs from production

Share, 0 to 1, of calibration replays that equal the stored production text, normalized, 4 decimals. Low agreement means replayed audio itself transcribes differently, so treat differences with care. Null before any replay succeeded

Differing turns by the judge's label

Differing turns by a person's review

Differing turns the candidate got right and production got wrong: by the expected text where the line has a flag, else by the review, else the label

The mirror of fixed: production right, the candidate wrong. These are the regressions a switch would bring

introduced scaled from the judged differing turns to all of them. Null while none is judged

Caller turns the candidate has finished, with a transcript or an error

totals.callerTurns

A few words naming the mistake

One turn of the pattern

turnIds.length

The turns that show the mistake

One sentence on what the candidate gets wrong

The estimate at creation, USD

Sum over the run's priced jobs, USD, 6 decimals: transcription and judge tokens, priced from the catalogue; an unpriced model counts as 0. Null while no job is priced