The reasoning pipeline

Every request to /v1/videos moves through the same six stages, in order, regardless of which downstream model ends up rendering the shot. This is what status is actually tracking as it moves from queued through to succeeded or failed.

Brief decoded

Your prompt (and optional negative_prompt) is parsed into a structured intent — subject, setting, tone, and any explicit call to action. This is also where obviously malformed requests fail fast, before anything downstream is touched.

Context harnessed

The decoded brief is merged with your account's brand dossier (approved claims, prior assets, tone of voice) and the regulatory ruleset that applies to your therapeutic area (ISI, fair balance language, label constraints). This merged context — not the raw prompt — is what every later stage actually reasons over.

Prompts layered

A generation-ready prompt is constructed from the harnessed context: a descriptive still-frame prompt for the opening frame, and a short motion prompt describing how it animates. Keeping these separate is what lets Finsen regenerate a bad keyframe for cents, instead of re-rendering an entire clip.

Claims linked

Every claim implied by the constructed prompt is checked against your dossier's approved claim set. A request that implies an unapproved or off-label claim is rejected here with a compliance_error, before any generation cost is incurred — see Compliance & verification.

Model routed

The request is routed to whichever downstream generation model suits the shot — see Model routing. This is also where resolution and tier-appropriate generation settings are applied.

Output verified

Once rendered, the output is checked again: does the visible result still match the approved claims and fair-balance requirements decided two stages earlier? Only after this second check does the request resolve as succeeded. This is the stage that makes SwishX's output MLR-ready by construction, not by a separate review pass afterward.

Tracking pipeline stage from the API

The status field on a video generation reflects where a request currently sits:

`status`	Meaning
`queued`	Accepted; brief decoding and context harness not yet started.
`processing`	Somewhere between context harnessing and model routing.
`verifying`	Rendered; running the output-verification stage.
`succeeded`	Passed verification. `output` is populated.
`failed`	Rejected or failed at some stage. See `error` for which one.

See it live

The Usage & Analytics page on the API Platform shows average time spent in each stage across your account's requests.