Loading
Loading
Brands pay creators for screen time. Screen-Time measures it against the contract. Meta SAM 3.1 tracks your product in every frame of a sponsored video, Muse Spark rejects look-alikes, Muse Voice Transcribe hears your brand, and Screen-Time checks each on-screen term frame by frame.
The sample runs the real pipeline, live on a 15-second Reel. No password needed to run the sample or open the example reports; checking your own video needs the demo password.
Product on screenSAM 3.1 + Muse Spark
0.0s
of 15.1 s, counted so far
LaCroix Strawberry Peach canCountedThe contract
0 of 3 checkedOne LaCroix test campaign, start to finish. Every step below is live on this site.
The problem. Brand deals come with rules: three seconds on screen, say the name, disclose the ad. By hand, someone scrubs every Reel, or nobody checks.
The creator link. Aldo, who built Screen-Time, sends his own Reel through his creator link: no account, no password.
The live run. SAM 3.1 tracks the can, Muse Spark checks each find against the product photo, Voice Transcribe listens, and every term of the contract is checked: a Poppi on the table is rejected as a look-alike.
Add a creator’s video and say what your product looks like. Every number in the report links to the moment that proves it.
Brands upload the file the creator delivered, or ask the creator to send it. Platform terms bar downloading other people’s videos without permission, so the brand checks the file the creator delivers.
The LaCroix test campaign’s three terms, as written.Open the campaign
Counting every can in the shot would pay a creator for someone else’s product. Muse Spark checks each object SAM 3.1 finds against your description and product photo, and says why in words you can check.
Screen-Time’s count against the labels, per frame, on the 11 Reels as delivered on this site.
Labels: drafted by two independent Claude vision passes at 5 frames a second, then checked frame by frame by a person on all 11 clips.
How it was measured, and where it is wrong| Score | SAM 3.1 alone | + Muse Spark |
|---|---|---|
| Precision | 95.0% | 98.4% |
| Recall | 99.3% | 91.1% |
Alone, a look-alike counts as the brand’s product. Recall is lower with Muse Spark because on George’s Reel it could not confirm the product SAM 3.1 tracked, so it was not counted and the report flags it.
What a check costs
$0.115for the whole check of Ash’s 15-second Reel, with Muse Spark and Voice Transcribe (227 frames).
One LaCroix test campaign with a $300 fee and the product photo: every Reel with its verdict, what it pays or holds, its seconds on screen and the moment that proves every term.
Open the LaCroix campaignPay $300. 21.7 s on screen, LaCroix said, “this is an ad” out loud: all three terms met. Tap a term to play the moment that proves it.
A can nobody can confirm. George’s can is tiny and blurred: Muse Spark can’t confirm it’s the right one, so nothing counts and the report asks a person to look.
The follow-up agent. Peter never says it’s an ad. Muse Spark with tools drafts the note and a re-cut request; nothing is sent until you approve.
The other take. Ash’s first take never says LaCroix, so the fee is held; her other take says it, is checked again, and pays.
The campaign. Pay or hold for every creator, with dollar totals and a CSV for finance. Screen-Time never moves money.
Check it yourself. Precision and recall per frame against labels a person checked, with the labels and Screen-Time’s count playing side by side.
Built during the hackathon. The pipeline, the contract checker, the follow-up agent and the accuracy harness, all open source, at cents a Reel.
The demo is free. Checking one video costs about $0.10–0.25 in Meta API fees.
SAM 3.1 tracking costs $0.20 per 1,000 frames.
Play a Reel and watch the person’s labels and Screen-Time’s count run side by side, frame by frame. Pick any row below to load its Reel.
Not Michael K’s Reel
seconds so far, of 12.4 s
Whole Reel: 2.6 s labelled, 3.0 s counted; precision 84.8%, recall 100.0%.
Red is where they disagree: solid where Screen-Time counted a frame the person did not label (an over-count), outlined where it missed a labelled frame. Per frame, at 15 frames a second.
The test set: 11 labelled vertical videos, including a clip with only a look-alike in shot, which must count 0 seconds. Screen-Time counted 0.0 s on them. All 11 results
A test campaign run by Screen-Time with friends, who agreed to be shown. Not affiliated with, sponsored or endorsed by LaCroix or Poppi; their names and cans belong to their owners.