Screen-Time
Proof your product was seen. Brands pay creators for screen time. Screen-Time measures it against the contract:
it tracks the brand's product through every frame of a sponsored video, rejects look-alikes, hears the brand name, and
checks on-screen terms frame by frame.
Problem statement: Build for Your World. Before a brand pays a creator for a sponsored Reel, Screen-Time checks the
delivered video against the deal's terms and shows the moment behind each verdict.
Built by Aldo Tile (Aetrai) for the Meta Global AI Developer Hackathon, startup track, on Meta's Model API (SAM 3.1,
Muse Spark and Muse Voice Transcribe). Everything in this repo was built during the hackathon, Oct 28 – Nov 4, 2026,
with Claude Code as the coding assistant: most commits carry its Co-Authored-By line.
Try it
Live at https://screen-time-inky.vercel.app. The main campaign is a LaCroix Strawberry Peach test campaign run by
Aetrai on real Reels: 8 friends filmed 9 of the 11 Reels, the builder filmed the 2 look-alike clips, and we burned in
the captions from each Reel's own speech (Rights). Ash's "re-cut" is her other take, filmed the same day
and uploaded after her first. Michael K's Reel came in through his request link, uploaded on his behalf for the demo
recording. Every report on the site is the stored result of a full run. Judges can start at
/judge: five steps through a live sample run, a pay case, a hold case,
the accuracy page and the follow-up agent.
- No password: the sample run, the example reports, and the follow-up agent on those reports. The sample run and the
agent are capped per visitor and per day. Past a cap, you see the newest finished run instead.
- Demo password (email [email protected] for it): saving a campaign, checking your own video, and approving the re-cut link
the agent queues. That link is a real upload link for the creator.
What it measures
Measured on the 11 LaCroix test Reels as delivered on the site (9 people; 1,931 frames with the can in 2,800), scored
from the site's own run reports (data/runs) frame by frame at 15 fps against labels drafted by two independent Claude
vision passes (George's on our earlier zoomed edit of his Reel), then checked frame by frame by a person in /label
(11 of 11 clips, Aldo, Nov 3; no ranges changed). Details, intervals and the reproduce command (pnpm accuracy --labels data/labels/lacroix) are in docs/ACCURACY.md:
| SAM 3.1 alone | SAM 3.1 + Muse Spark (what the app reports) |
|---|
| Precision | 95.0% | 98.4% (95% interval 96.7–99.6%) |
| Recall | 99.3% | 91.1% (70.3–100.0%) |
| F1 | 97.1% | 94.6% |
| Seconds counted on the look-alike Reel | 4.2 s | 0.0 s |
- What Muse Spark adds: it rejected both look-alikes SAM alone counts (a LaCroix Pamplemousse can, a different flavor,
and a Poppi can). Per frame, nearly all of the precision gain is the one look-alike Reel, which the builder filmed.
It costs recall on George's Reel: Muse Spark could not confirm his small can, so it counted 0 s against 11.2 s
labelled, held the fee and flagged the report. Object by object, hosted Muse Spark is right on 11 of 12 objects,
SAM alone on 10 of 12.
- Where it is wrong: George's Reel (above); and Michael's Reel (not Michael K's), counted 3.0 s where the labels say
2.6 s, so it said pay against a 3 s term where the labels say hold (it flagged the report as close). Pay or hold
agrees with the labels on 9 of 11 Reels.
- Until Nov 4 the published figures scored a zoomed edit of George's Reel that we made (2.3x, following the can); they
now score the Reel as delivered.
- Cost: $0.106–0.230 of Meta API per Reel on these runs ($1.56 for all 11), with the product photo and
transcription. SAM 3.1 is $0.20 per 1,000 frames; Muse Spark and Voice Transcribe are counted at the fixed amount
the budget reserves per call, not the billed amount.
- An earlier test set of 20 clips the builder filmed himself or took from Pexels (2026-10-28: precision 97.7%,
recall 95.6%) is kept in docs/ACCURACY.md; its campaigns are archived and hidden from the site.
Who else checks creator videos
Brands already check creator videos, by hand and with tools. What each source says, as read on 2026-11-03:
여기까지가 README 앞부분입니다. 나머지는 원문에서 읽을 수 있습니다. README 전문 보기 ↗
