Reading a verdict

Every endpoint returns a check object. Its result depends on the kind of check; the rest is the same for all.

The check object

Field	Meaning
`id`	`chk_…`. Use it with `GET /v1/checks/{id}`.
`object`	Always `check`.
`mode`	`live` or `test`, from the key.
`source`	Where it came from: `api`, `sdk`, `mcp`, `watch`, `comfyui` or `web`.
`label`	Your `label`, if you sent one.
`created_at`, `finished_at`	ISO 8601, UTC.
`kind`	`check`, `rank`, `keyframe`, `preflight`, `extend` or `preview`.
`state`	`queued`, `running`, `done` or `error`.
`verdict`	Once done: `pass`, `fail` or `unsure`. For a ranking, the verdict of the best take.
`headline`	One sentence, e.g. "Don't use this take: 2 problems found."
`result`	What was found (below).
`cost`	What it cost: `{"usd": 0.04, "display": "$0.04"}`. Test checks cost 0.
`seconds`	Seconds of video checked.
`report_url`	An HTML report with evidence frames, kept for your retention period.
`signed`	The signed record: `record`, `signature`, `kid`. See Signed verdicts.
`error`	If `state` is `error`: `{type, message, fix}`. You weren't charged.
`review`	Set from the review screen: `approved`, `rejected`, `escalated` or null.

Status

pass: usable. There may be warn findings worth a glance (e.g. black frames at the very start, which are a trim, not a defect); they don't fail the take.
fail: don't use it as is. At least one problem was found with confidence of 60% or more.
unsure: there's a possible problem below 60% confidence. Raccord says "not sure, take a look" rather than guessing. Show the take to a person.

Findings

result.findings is a list, worst first.

Field	Meaning
`check`	Which check found it (table below).
`severity`	`fail` or `warn`.
`title`	One plain sentence a producer understands.
`fix`	The one change most likely to resolve it.
`start`, `end`	Seconds into the clip. A finding covering the whole clip has `start` 0 and `end` equal to the clip length.
`confidence`	0 to 1. How the stated confidence holds up is measured on the accuracy page.
`evidence`	Frames that show it: `frame`, `time`, `caption`, and a `region` (x, y, w, h from 0 to 1) when it's about part of the frame. `source` names a reference photo when the evidence is one; `crop` is true when only the region is shown.

What each check catches

`check`	Finds
`blank_frames`	Black or single-colour frames. At the very start or end it's a `warn` (a trim).
`frozen_frames`	The picture holding still when it should move, and stutter.
`flicker`	Brightness flickering across the frame or in one area. Camera shake is removed first.
`delivery_spec`	Wrong resolution or frame rate against the `size` and `fps` you sent, and variable frame rate.
`identity`	Nobody matching the references, the wrong person, or a face that changes partway (drift). Needs `reference[NAME]`.
`wardrobe`	A character's outfit changing colour partway, or not matching the costume photos you sent as `costume[NAME]`.

result also repeats status and headline, and includes checks_run (which checks ran), engine_version, elapsed_seconds, clip (resolution, frame rate, length, codec and the file's SHA-256, matching signed.record.inputs) and identity: for each character, how closely their face matched over time (times, scores, match_threshold).

Other kinds

rank: result.takes best first, each with rank, path (your file name), status, reason, a score (only the order matters) and its full verdict. result.top_two names up to two takes to use, passing takes only, so it's empty when none passes. result.errors lists files that couldn't be read.
preflight: result.findings about the prompt and photos; result.characters counts usable photos per character.
preview: result.stop (true or false), headline, reasons, and the raw signals.