Verified Asset: review page

Hackathon deadline Wed 4 Nov, 9:00 pm Panama (ET). Private working page, not for submission.

1. FINAL: your one-take voice (2:40)

Your own voice, from tonight's one-take recording

Made from your 3:40 voice note. I used your best take of each line, shortened the long pauses, and placed each line where it fits the picture. The burned-in captions now match your words. The eval scene is 3.6 s shorter, so the video runs 2:39.7. Full 1080p file: ~/Videos/verified-asset-demo/verified-asset-final-jeff-voice.mp4 (56 MB).

Check by ear: (1) Your take does not include "letting the model judge made two". The video does not say it, but the on-screen tile still shows "0 vs 2 false reassurances". If you want it spoken, send a 3-second voice note of just that line and I'll drop it in. (2) "SAM 3.1": Whisper heard "Sound 3.1" once, so check that it sounds right. (3) It is quieter than the earlier versions: -15.1 LUFS. That is fine for YouTube.

Earlier versions

Your narration from the phone clips (v7)
ElevenLabs, your cloned voice ("Jeff")

Model eleven_multilingual_v2. Closest to you of the synthetic options by a crude speaker-embedding score (0.79 vs 0.63 for an unrelated voice and about 0.88 for you vs you). I have not listened to it, so judge by ear: it may sound only moderately like you.

VibeVoice (Microsoft, preset voice)

Run by Codex gpt-6-astra, CPU only. This model only offers bundled preset voices; it cannot clone you. Score 0.66, close to the unrelated-voice control, so it does not sound like you. Included for comparison, not as a candidate for your voice. Not listened to by me.

ElevenLabs, generic stock voice (earlier)

These are 720p copies. The full 1080p files are on your Fedora machine in ~/Videos/verified-asset-demo/. Same visuals in both.

What changed in v7: the opening, how-it-works and evaluation scenes are now crisp desktop recordings of the live site, with highlight rings that follow the narration. The phone live-check flow, SAM clip, extension mock page and end card are unchanged. No intro.

Not verified: I sampled frames but haven't watched it end to end, and no second reviewer has seen it. Please look for anything that feels off.

2. The script you recorded from

You said you will record your voice in one shot. Read this in order, calm pace, phone 10 cm from your mouth. The times are where each line sits in the current video, so you can match the pacing. If your take runs different, I re-time the picture to it. Sentences in grey are on screen only: stay quiet there.

0:00Before this hackathon, we had a research prototype and product notes, in a separate repo that isn't part of this entry. The app you're about to see is what we built this week.
0:12Verified Asset checks a second-hand listing's claims against the seller's photos. This isn't an image analyzer. The photo observer never sees the claims. Plain rules decide, and the report is tied to the exact version we checked.
0:31A seller picks claims: bedside table, two-tone wooden bedside table, no visible damage. Each claim asks for the photo that can settle it, with the listing code on paper in the front photo.
0:52The check is now done. The models describe the photos without seeing the claims, and we can see what Muse Spark saw: the photo ID, that the picture was usable, the material, the colours, the visible text of the code with its date, and so on.
1:17Muse Spark is also asking for additional pictures to validate the claim that there was no visible damage. We will now take this close-up picture and perform the check again.
1:30(wait shortened, stay quiet about 4 seconds)
1:34As you can see, the model is checking again for the picture that it has to retake, and successfully detected peeling in a detailed photo.
1:50Buyers open the report with no account. If I edit the listing, the old report is voided. I can also withdraw one.
1:59On a Marketplace-style item page, our Chrome extension reads the report link and shows each verdict. This page is a mock, so no real listing is touched.
2:12Spark observes and plans the missing shot. SAM 3.1 counts and boxes. Glimmer does the local evaluation.
2:22On twelve synthetic cases, blind observation plus rules made no false reassurances; letting the model judge made two. A live check costs about three and a half cents.
2:37Check the claims before you make the trip.

Save as one voice note. If you stumble, stay quiet 3 seconds and repeat the sentence; I keep the best take. Please also say which word you want in "made two" (line at 2:22): "made two" is what is on screen.

3. Your remaining tasks

Ticks are saved only in this browser.

4. What is verified and what is not

Checked (final, your voice): 1920x1080, H.264, 30 fps, AAC; length 2:39.7; -15.1 LUFS, peak -4.4 dBFS. A Whisper re-transcription of the finished audio matches the captions, with no clipped or doubled words in the spots I checked. I sampled 15 frames.

Unverified: the official rules for the video (only the organizers' emails and our Jira notes say 3 minutes, public link, work built since 27 Oct); Gilberto's claim that the live SAM check is deployed today; a full privacy scan of the finished video; any independent review.

Numbers used are from the project's own eval: 12/12 vs 9/12, 0 vs 2 false reassurances, synthetic n=12, about $0.035 per live check.

Links: verified.adwentech.com ยท github.com/cpu-16/verified-asset