API reference

Base URL

https://api.physionlabs.ai/v1

Authentication

Send the API key as a bearer token to authenticated API endpoints.

Authorization: Bearer gk_live_...

Conventions

Send and receive JSON unless an upload endpoint requests video bytes.
Include an API error’s request_id when contacting support.
Treat evaluations as asynchronous jobs.
Stop polling at completed, partial, or failed.
Read Retry-After before retrying a rate-limited request.

Findings

result.glitches holds one entry per finding. Both kinds carry id, type and description; the rest depends on type, so branch on it.

field	`visual_glitch`	`prompt_misalignment`
`region`	where in the clip, with per-frame boxes	—
`prompt_segment`	—	the span of your prompt this is about
`severity`	—	1-5

A field that does not apply is absent, not null, so a visual glitch has no severity key at all. Both SDKs model this as a union you narrow on type:

if (finding.type === "prompt_misalignment") finding.severity;
else finding.region;

if finding.type == "prompt_misalignment":
    finding.severity
else:
    finding.region

severity says how far a prompt requirement was from being realized: 1 is a minor mismatch and 5 means the requirement is absent entirely. It reads in the opposite direction from a score, which is the one thing worth getting right before you threshold on it. Every finding you receive already cleared our reporting threshold.

{
  "id": "gl_pm_r2",
  "type": "prompt_misalignment",
  "severity": 2,
  "description": "Action of washing a dirty white plate \u2014 The plate is dirty and white, and the robot holds it under water, but the 'washing' action is mostly just rinsing under the tap rather than scrubbing or using soap, though the prompt doesn't specify scrubbing.",
  "prompt_segment": {
    "text": "wash a dirty white plate",
    "char_start": 12,
    "char_end": 36
  }
}

How long a run took

Every evaluation carries timing, and timing.e2e_ms is the only number in it: milliseconds from your submission to the terminal answer, which is what you waited — our queueing, fetching the video, and the model’s own time included.

{ "timing": { "e2e_ms": 43000 } }

null when nothing was measured, which is every run settled before we recorded it. Null and never 0: a run whose latency nobody recorded and a run that took no time are different claims.

Fields not described here

A response may contain fields this reference does not name. They are not part of the v1 surface, they can change or disappear without a version bump, and the SDKs drop them rather than surfacing them — the Python client does not put them in model_extra. Read what is documented. GET /v1/status accepts unauthenticated requests. Upload PUT requests use a temporary upload URL and do not receive the Galileo API key. The endpoint pages in this section are generated from the Galileo v1 OpenAPI contract. The Node.js and Python SDK types come from the same contract revision.

Getting started

Guides

SDKs

Resources

API reference

Base URL

Authentication

Conventions

Findings

How long a run took

Fields not described here

​Base URL

​Authentication

​Conventions

​Findings

​How long a run took

​Fields not described here

Base URL

Authentication

Conventions

Findings

How long a run took

Fields not described here