Model documentation is revised silently. Parameters are added, defaults change, preview notices are removed, and there is usually no changelog. So a specification claim that says only what is true has a shelf life it does not disclose.
Four fields fix it. This is not a standard, it is just the minimum set that makes a claim re-checkable by someone who is not you.
| Field | Why |
|---|---|
| Model identifier | wan3.0-video, not "Wan". Families contain products with different specs, and cross-product contamination is the most common source of wrong figures |
| The value | The figure itself |
| Source document | Which page. Not "the docs" — the specific one, because vendors contradict themselves |
| Date read | The only field that lets anyone tell whether your claim is stale or the vendor changed it |
Written out, a row looks like:
wan3.0-video— maximum output resolution 1080P. Source: Alibaba Cloud Model Studio, Wan 3.0 video generation API reference. Read 27 August 2026.
Longer than "Wan 3.0 supports up to 1080P". The difference is that a reader can disprove it in one click, which is the entire point.
Without a date, a wrong claim and a stale claim are indistinguishable, and they need completely different responses. A wrong claim means the author did not check. A stale claim means the author checked and the world moved.
With a date, a reader who finds a discrepancy learns something useful either way: if your date is old, the spec changed and they now know roughly when; if your date is recent and you still disagree, one of you is reading a different document, which is a situation that genuinely occurs.
Dating also makes maintenance tractable. You can sort by date, re-check the oldest rows, and know your coverage. Undated claims cannot be triaged, so in practice they are never re-checked — which is how a spec table quietly becomes a historical document.
Not every fact has the same backing, and flattening that is how third-party inference gets laundered into vendor fact. Three tiers cover almost everything:
First-party. The vendor's own documentation. State the document.
Corroborated third-party. Multiple independent sources agree and the vendor's own
artefacts are consistent with it. Say so, and say how many. The billing rule that reference
video is charged for its own seconds is an example: three independent gateway
documentations state the same formula, and Alibaba's response schema carries an
input_video_duration field that would be pointless otherwise. That is strong. It is still
not first-party, and the sentence should say which it is.
Single third-party. One source, not corroborated, vendor silent. Usable, flagged as such, and never phrased as though the vendor said it.
The cost of tiering is a few words per sentence. The benefit is that when one of your third-party claims turns out to be wrong, it fails alone instead of taking your first-party claims with it.
The rows most worth having are the ones that cost you something to state — no 4K tier, no 21:9 ratio, 30 fps rather than 24, closed weights, a benchmark lead that sits inside its own confidence interval.
Two reasons, one principled and one practical.
The principled one: a spec table that only contains favourable facts is not a spec table. The reader cannot tell whether an absent capability was checked and found missing, or never checked.
The practical one: unflattering rows are the ones nobody else publishes, which makes them the ones that get cited. When someone asks whether a model does 4K, the page that can answer no, here is the parameter table, checked on this date is the only page that answers the question. Pages that promise the capability cannot serve that query at all.
Wan 3.0 sat first on the Artificial Analysis text-to-video With Audio board in the snapshot I read on 25 August 2026, at Elo 1,240.
The honest version of that sentence includes four more things: the board (there are several), the snapshot date (rankings move), the sample count (5,728 comparisons, the smallest in the top five), and the confidence interval (±9). The gap to second place was three points. Three is inside nine, and the leaderboard itself listed the rank as a range of 1–2.
So the correct claim is "statistically tied for first on that board in that snapshot", not "the world's best video model". The first is defensible for as long as the snapshot is cited. The second is not defensible today and will read as marketing in a month.
Every specification on wan-3.run is published in this format — identifier, value, source link, date read — including that benchmark, which the site reports as a tie rather than a win.