# Canary exchanges — real, billed, redacted

Real end-to-end recordings against the live OpenRouter video API, kept as the
contract fixtures for `182.C.11.11`. Spending was approved by the user on
2026-09-10 with a ~$5 cap; the running tally is at the bottom of this file and
each recording states what it cost.

Only the JSON exchanges are committed. The rendered clips are not: they are
hundreds of kilobytes each, the useful part is the measurement, and the bytes
would be biometric-adjacent for any recording that ever involves a person. Each
recording carries the clip's `sha256`, byte count, and full `ffprobe` output
instead.

## `grok-imagine-video-1s-480p.json` — $0.05

`x-ai/grok-imagine-video`, 1 second, 480p, 16:9, prompt with no people in it.
Chosen because it is the cheapest complete exchange in the catalog: 5 cents per
480p second, and it accepts a 1-second duration.

Create → poll → poll → poll → download, ~18 seconds end to end.

### What it confirmed

- **The documented shapes are the real shapes.** The 202 returned exactly
  `{ id, polling_url, status: "pending" }`, and the completed poll exactly
  `{ id, generation_id, polling_url, status, unsigned_urls, usage: { cost, is_byok } }`.
- **The pricing normalizer is right.** `cents_per_video_output_second_480p: 5` ×
  1 second predicted $0.05, and `usage.cost` came back **0.05**. That is
  `182.C.10.05` checked against money rather than against itself.
- **The download really does need the key.** Without `Authorization` the content
  URL answers **401**, which is what `182.C.11.08`'s host allow-list exists to
  protect — we attach our key to a provider-supplied URL.
- **`Content-Type: video/mp4`**, and no `Content-Length` at all, so
  `182.C.11.08`'s streaming byte count is the only size bound that actually
  fires here.

### Two corrections it forced

1. **The headers the CORS list names are not sent.** `182.C.11.05` reads
   `request-id`, `X-Generation-Id` and `X-Provider-Name`, on the strength of the
   catalog response's `access-control-expose-headers`. Measured on this
   exchange, **none of the three appeared** on the create or on any poll — only
   `content-type` and `cf-ray`. The module already captured each as nullable and
   recorded a missing header as absent rather than filling it in, so no code
   changed; but the row now rests on a measurement instead of on an inference,
   and `upstreamProvider` should be expected to be null on this route.

2. **A take can carry an attached picture, and it broke the validator.** The
   clip came back with **three** streams: `h264` 848×480 at 24 fps with 25
   frames, an `aac` track, and an **`mjpeg` stream of a single frame** at the
   same dimensions — an embedded thumbnail. `182.C.11.09` picked the first
   `codec_type: "video"` entry, which happened to be the h264 one. Ordered the
   other way it would have called the take `video_codec_unsupported`, or read
   one frame of a one-second clip as a truncation. Fixed by
   `selectPrimaryVideoStream`, which skips a still-image codec carrying a single
   frame and prefers a stream with a real frame rate over one reporting `0/0`.

### One thing worth remembering about the catalog

`x-ai/grok-imagine-video` carries `generate_audio: null`, and the take came back
**with an aac track**. Null in that field really is "the catalog declines to
state", not "no audio" — which is exactly what the capability vocabulary's
`fallbackWhenAbsent: 'probe'` was built for, now confirmed rather than reasoned.

The delivered duration was 1.041667 s against 1 s requested — 4% over, inside
`182.C.11.09`'s 15% tolerance.

## Follow-up observations on the same job

- The `unsigned_urls` entry was still answering **200** with the full 588 899
  bytes about five minutes after completion. **No expiry was observed**, and the
  TTL is not published — which is exactly why `182.C.11.12` proves the
  release-survives-expiry property structurally rather than by waiting. An
  expiry we have not seen is not one we can rely on not happening.
- `content?index=7` on a job with one output answers **400**
  `{"error":{"message":"Video index 7 out of range (1 videos available)","code":400}}`
  — a permanent failure by our classifier, and a second place the output count
  is stated.

## The duplicate-create recording (`idempotency-duplicate-create.json`)

`182.C.15.03` asks for a guaranteed-idempotent provider submission, so the first
question is whether the provider offers one. It does not, and the recording is
two byte-identical creates sent seconds apart carrying the same
`Idempotency-Key: oshun-hv-dedupe-probe-20260910`:

| Create | Job id                 | Status      | `usage.cost` |
| ------ | ---------------------- | ----------- | ------------ |
| first  | `nakULoPh87xys3evo83G` | `completed` | 0.05         |
| second | `rK48YANErDZClIpnrFU4` | `completed` | 0.05         |

Two ids, two renders, and `total_usage` on `/api/v1/credits` moved 21.018672686
→ 21.118672686 — exactly $0.10. Whatever deduplication this API has, a header
does not reach it, and the header the docs DO name
(`X-OpenRouter-Idempotency-Key`) is one **it** sends on a webhook, not one it
reads on a create.

Three findings came out of the same $0.10, and each one closes a door that a
simpler design would have wanted to walk through:

1. **You cannot list your jobs.** `GET /api/v1/videos` answers with a plain-text
   `404 Not Found` — the router saying no such route, the same evidence shape
   that settled cancellation in `182.C.11.07`. `GET /api/v1/generations` is a
   JSON `Not Found`. A job id you failed to write down is a job you cannot find.
2. **The spend lookups lag.** `GET /api/v1/generation?id=` returned 404 for both
   generations at the moment each job reported `completed` with a cost, and
   `/credits` still showed the old `total_usage`; both settled a few minutes
   later. They audit spend after the fact; they cannot answer "did my submission
   land" while the question is live. (When it does settle, that endpoint is
   rich: `api_type: "video"`, `total_cost`, `provider_name`, `upstream_id`,
   `finish_reason`, `cancelled`, `data_region`, `workspace_id`.)
3. **`user` is not forwarded.** The create body carried
   `"user": "hv-claim-probe-1"` and both generation records report
   `external_user: null`. So `callback_url` — which the guide says is honoured
   per request — is the only channel that returns a value we chose, which is why
   `182.C.15.03` puts the claim key in its path.

`GET /api/v1/activity` exists but answers
`403 Only management keys can fetch activity for an account`, so an
account-level spend feed is available to a different credential than the one
that submits.

### The unbilled half

Four deliberately invalid creates — `resolution: "nope"`, plain, then with a
`user` field, then with an unknown field, then with the `Idempotency-Key` header
— returned **byte-identical** `400` zod envelopes, created nothing and cost
nothing. Two things follow. The API validates the body before it starts any
work, so a 400 is proof that no job was created and the claim is safe to release
and retry; and unknown body fields are ignored rather than rejected, so an
unrecognised field is not an error we will ever see. The enum in the error also
states the resolutions the API accepts outright: `480p`, `720p`, `768p`,
`1080p`, `1K`, `2K`, `4K` — an independent confirmation of the `768p` tier the
catalog shows.

One wire detail worth keeping: both create responses arrived after several lines
of **whitespace padding** ahead of the JSON. `JSON.parse` tolerates it; a reader
that slices bytes would not.

## Running tally

| Date       | Recording                                       | Model                     | USD      |
| ---------- | ----------------------------------------------- | ------------------------- | -------- |
| 2026-09-10 | `grok-imagine-video-1s-480p.json`               | `x-ai/grok-imagine-video` | 0.05     |
| 2026-09-10 | `idempotency-duplicate-create.json`             | `x-ai/grok-imagine-video` | 0.10     |
| 2026-09-11 | `../2026-09-11-wan-canary/wan-2.7-2s-720p.json` | `alibaba/wan-2.7`         | 0.50     |
|            | **Total**                                       |                           | **0.65** |

Budget approved: ~$5.00. Remaining: ~$4.35.

The 2026-09-11 entry is the one that did not cost what the catalog said it
would: `alibaba/wan-2.7` lists a single `duration_seconds: "0.1"` SKU, a
two-second render predicts $0.20, and the API billed **$0.50**. See that
recording's README.
