The clip and the test conditions
The source video is an original Harbor Notes editorial fixture, not footage of a real product or customer. Its six scenes show a fictional launch brief. Several consequential fields appear only on screen: the review window, the brief owner, and its final status. The narrator explicitly says she extends the review window without reading the final number aloud. That makes the clip useful for testing whether a model follows the visual state instead of producing a plausible answer from speech alone.
On September 20, 2026, we selected Qwen3.8-Omni-Flash in Qwen Studio, left its displayed Thinking mode on, uploaded the MP4, and asked three questions in the same guest conversation. We did not supply a transcript, frame extracts, or the fixture's answer key to Qwen. Alibaba Cloud documents video and audio input with text output for this model; our observations below come from the actual Qwen Studio replies, checked against the source video and its retained scene boundaries. We did not run repeated trials or measure speed, cost, or accuracy over a dataset.
It chose the saved value, not the superseded one
Our first question asked for the final review window, the value it replaced, a timestamp, and the visual or spoken support. Qwen answered 72 hours, identified 48 hours as the superseded value, and placed the evidence at roughly 00:26 to 00:33. The source's review scene runs from 00:26.019 to 00:33.762. Its answer landed in the right scene, though the rounded range is not a frame-exact citation. Watch that scene.
Qwen pointed to the screen's INITIAL WINDOW and FINAL REVIEW WINDOW labels and the SUPERSEDED and SAVED tags. It separated the narrator's statement that she extended the window from the visual evidence for the exact 72-hour value. The narration does not supply either number, so a transcript-only answer could not justify this choice.
It found the final state without calling it publication
The second question asked who owned the brief, what its final publication status was, and whether the video showed approval. Qwen located the last scene at approximately 00:41 to the end. The source scene begins at 00:40.992 and ends with the 50.672-second clip. It read Mina Chen under OWNER and Needs owner review under STATUS, with the adjacent NOT PUBLISHED tag. It correctly treated READY TO REVIEW as a pending check, not approval. Watch the final scene.
The screen identifies the person, while the voice explains that the brief is waiting for review. Qwen used both, without promoting the brief to a published state.
It refused a percentage the clip did not contain
For the third question, we asked how much the 72-hour window increased qualified demo requests, in percent, and demanded a timestamp for any number. The video sets Qualified demo requests as a goal at about 00:18 to 00:26. It never reports a baseline, an observed result, or a percentage lift. A later screen says No unsupported percentages. Qwen declined to supply a figure and cited the goal and guardrail scenes separately. See the goal and the claim guardrail.
One sentence in Qwen's otherwise careful answer went beyond the clip. It said the brief “never goes live.” The final screen shows NOT PUBLISHED and Needs owner review at the time the recording ends. It cannot establish what would happen later. Refusing to supply a percentage was correct; the timestamp did not justify that extra claim.
What this result is useful for
On this one clip, Qwen3.8-Omni-Flash handled three distinct evidence checks: choosing a saved visual state over a discarded one, retrieving the right final scene, and refusing an unsupported numerical outcome. The source holds each relevant screen for several seconds with large, readable text. We did not test brief flashes, dense charts, long meetings, or disputed speech; the three questions also shared one conversation, so they are not independent benchmark samples.
To repeat the meaningful part of this test on your own recording, ask for the answer, a short timestamp range, the exact screen label or spoken line that supports it, and any earlier value that was replaced. Then open the cited range. For this fixture, the decisive check is 00:26.019 to 00:33.762: 48 hours is marked SUPERSEDED, while 72 hours is marked SAVED. For a fuller workflow on checking transcripts against visual evidence, see our source-linked video article guide.
