Eve library

A real, billed canary on the primary lane

Guide

Revision 1

On this page

182.C.36.10. The item says: run real provider canaries with tiny bounded spend and redacted evidence before declaring any production route available. The 2026-09-10 canaries used x-ai/grok-imagine-video because it was the cheapest thing on the list. That proved the transport; it did not prove the route we would actually declare.

So this one runs on alibaba/wan-2.7 — the only family that serves exact-dialogue-performance and multi-character-scene (measured in 182.C.36.03), which makes it the route with no fallback and therefore the one worth the money.

What was spent#

alibaba/wan-2.7 prices duration_seconds at $0.10, and its shortest advertised duration is 2 s, so the smallest render it sells is $0.20. That is what was requested: 2 s, 720p, no reference images, a prompt with no people in it.

Date Recording Model predicted billed
2026-09-11 wan-2.7-2s-720p.json alibaba/wan-2.7 0.20 0.50

Added to the running tally in ../2026-09-10-canary-exchanges/README.md (previously $0.15 of ~$5.00).

How it was built#

The payload was built by the real adapter code, not by hand. The script read the live catalog with parseCatalogBody, compiled our control vocabulary with compileForModel, assembled the body with buildCreateRequest, and passed it through validateCreatePayload and assertSubmittable before anything was sent. A canary built from a hand-written curl proves the API works; this one proves our code can talk to it, which is the claim the item is about.

The compiled body shows the nesting the transport record documents: our negative-direction and aspect-override became negative_prompt and ratio, nested under provider.options["atlas-cloud"].parameters — the slug the endpoints table says serves this model, not one derived from the model id.

What the canary found#

The create returns 202 Accepted, not 200. The 182.C.36.01 provider double had been written answering 200, and a double that answers the wrong success status teaches the suite the wrong success status — this module's own warning turned on itself. The double, the end-to-end test and the scripts were corrected to 202 against this measurement.

Our production code was already right: captureSubmission takes the HTTP status as data and keys on the presence of id, so it never depended on 200. The wrong assumption lived only in the test double and in the canary script, and the canary script's version of it cost $0.20 — it treated the 202 as a failure and exited after the job had been created.

Two smaller wire facts, both confirming earlier records on a different model:

  • 55 bytes of leading whitespace before the JSON. JSON.parse tolerates it; a reader that slices bytes would not. The 2026-09-10 canaries saw the same thing on Grok.
  • x-request-id and x-provider-name are both absent from the create response. The transport record already treats them as nullable rather than as guaranteed.

And one shape worth noting: the create response carries id, polling_url and status and nothing else, while the poll response adds generation_id while still pending — so a generation_id is not evidence that a render finished.

What the render actually was#

The job completed. Create → eight polls → authenticated download, 107 seconds from the first poll to completed. The download answers 401 without the bearer and 200 with it, and returns 2,676,551 bytes of video/mp4 (sha256:361259ae…).

Three of the four things asked for came back different.

asked for delivered
resolution: "720p" 1920×1080
duration: 2 2.020000 s
no audio field sent aac, 44.1 kHz, stereo
— 30 fps, h264
  • The resolution was substituted silently. The request named 720p, the API accepted it, and the file is 1080p. Nothing in the create or poll responses mentions the substitution; it is visible only by probing the bytes. This is the case 182.C.28's encoding-conformance gate exists for, and it only catches it if it compares the delivered file against the request rather than against a policy.
  • The duration is 2.020 s, not 2.000. That is the +20 ms AAC padding 182.C.35.09 measured on our own ffmpeg pipeline, now confirmed on a real provider render. A sync budget that assumes the provider returns the duration it was asked for is already 20 ms down.
  • It produced an audio track nobody asked for. No generate_audio field was sent — the adapter had no value for it — and the take has stereo AAC.

What it cost, and what the catalog said it would#

This is the finding the item exists for.

alibaba/wan-2.7 advertises exactly one pricing SKU: duration_seconds: "0.1". Our own normalizePricing reads it as $0.10 per output second with minimumUsdPerGeneration: null. Two seconds predicts $0.20.

The API reported usage.cost: 0.5. $0.50 — two and a half times the prediction, or $0.25 per delivered second.

That figure is USD and directly comparable: the 2026-09-10 grok canary reported usage.cost: 0.05 for a 1-second 480p render against a 5-cents-per-second listing, matching exactly. So the reporting unit is not in question — the wan listing is.

No combination of this model's single advertised SKU produces 0.50. The likely cause is the 1080p substitution above being billed at a 1080p rate the catalog does not list, but the cause is a guess and the discrepancy is a measurement, and only the second belongs in a cost model.

Consequences, stated rather than implied:

  • Any estimate built from pricing_skus under-predicts this lane by 2.5×. That includes admission decisions, per-job ceilings and budget arithmetic (182.C.37.05), and cost-per-releasable-second (182.C.37.06).
  • 182.C.34.03's rule that cost is a column the subject of the measurement supplies is sharper than it looked: here the subject supplies a number that its own catalog contradicts.
  • This is precisely why the item says to canary before declaring a route available. Declared on catalog arithmetic, this route would have been budgeted at 40% of its real price.