Raccord Check
Is this take good?
The quality gate for AI-generated video. Every take from any model, checked against what was intended, with the evidence and the fix.
Don’t use this take: 1 problem found.
Maya · match to reference
Maya’s face changes from 0:04.2 to the end of the take and no longer matches the reference.
Fix: re-render with a different seed and Maya’s reference images attached.
Illustrative verdict, in the format the working build produces.
The problem
Every second is billed. Many takes are thrown away. Someone has to watch them all.
$0.40
per second of Veo 3.1 video with audio, at Google Cloud list price.1
400 → 15
clips generated versus clips used for one AI-made TV ad.2
70,000+
clips sorted by a team of five in 30 days for one holiday ad.3
Models now try to keep characters consistent. On the vendor pages we reviewed, none checks the finished take against your references.4 So a person still does.
- Google Cloud, Vertex AI pricing, Veo 3.1 with audio, 720p/1080p, November 2026. cloud.google.com/vertex-ai/generative-ai/pricing
- The Keyword, 16 Jun 2025, on Kalshi’s NBA Finals ad: 300–400 clips generated, 15 used. One production. thekeyword.co
- TheWrap, 3 Nov 2025, on Coca-Cola’s 2025 AI holiday ad. thewrap.com
- Reviewed November 2026: Runway Gen-4 (runway.com), Kling Element Library (kling.ai), Veo 3.1 Ingredients (blog.google), LTX Elements (ltx.io). Not a claim that none exists.
What it does
One question. One answer. For every take.
Catch
Character and wardrobe drift against your references, plus black frames, freezes and flicker. Each problem comes with the frame, the time and one fix.
Rank
Give it a batch of takes. It puts the best two in front of you, so people watch fewer clips.
Stop early
Checks the start frame before you pay for motion, stops asking once a take passes, and halts drifting renders on open models.
Measured so far
All three checks met their v1 accuracy targets.
Tested on 5 November 2026 on 420 held-out clips that were not used while tuning. These are planted defects in synthetic clips and animated public-domain portraits, not real AI-generated footage.5
124 of 124
technical defects caught: black, frozen and flickering pictures. 0% of bad takes passed, 0% of good takes failed.5
82 of 82
character identity defects caught: wrong person, face swap and gradual face drift. 0% of bad takes passed, 0% of good takes failed.5
69 of 72
wardrobe changes caught: outfit colour changes. 4.2% of bad takes passed and 8.3% of good takes failed, inside the v1 targets of under 5% and under 10%.5
259 of 263
problems reported at 90–100% confidence were real.5
0.42 s
of CPU time per second of video, on an Apple M5 laptop with no GPU.5
~0.23 s
per image for the start-frame check, on the same laptop. Measured in tests, not in the benchmark.6
Not yet done: comparing Raccord Check with a general AI model asked the same question. Accuracy on real AI-generated footage will be measured with pilot customers.
- Kolam Labs internal benchmark, 5 Nov 2026: 420 held-out clips (seed 7, not used while tuning), engine 0.1.0, Apple M5 with 16 GB, CPU only. Planted defects in synthetic clips and animated public-domain portraits, plus good takes designed to tempt false alarms. Not real AI-generated footage.
- Kolam Labs internal tests, November 2026: start-frame (keyframe) check timing on the same laptop. A test measurement, not part of the benchmark.
Where it fits
Wherever money is about to be spent on a take.
- Before render
The prompt, the references and the start frame.
- During render
Stop a drifting take on models you run yourself.
- After render
A verdict per take, and the best of a batch.
- At delivery
A signed record of what was checked.
Built to read output from any model
- Veo
- Kling
- Runway
- Seedance
- LTX
- Wan
- Hailuo
- and the next one
It never generates video itself. Model names are trademarks of their owners, used only to describe compatibility.
For pipelines and agents
Wherever generated video lands, it can be checked.
A paid API priced per second checked, with an MCP server so AI agents can call it directly.
- REST API
- MCP server for AI agents
- ComfyUI node
- Watch folders
- Command line
- Review screen
How we build it
Honest by design.
Plain words
Every problem is one sentence a producer understands, with the time it happens.
Honest confidence
When it isn’t sure, it says “not sure, take a look” instead of guessing.
Signed verdicts
Each verdict records what was checked and by which version, and can be re-run.
Private by default
Clips are deleted after checking unless you keep them, and never used for training without consent.
Early access.
For studios, platforms and pipelines generating video at volume.
hello@kolamlabs.aiLaunch expected Q4 2026