Docs v2026.7

Technical docs

How our scoring models work: measured results, clear limits.

Download the official Model Card and Technical Report for the models behind every DotCheck score. Held-out test results, explained in plain language.

Measured results

Published held-out results for Docs v2026.7. Everyday scoring uses one image model, one video-frame model, one audio model, and seven language-specific text models on a shared feature stack.

Primary models

Vermeer (image) inhouse@12
real average 0.000 · AI average 0.938 · balanced accuracy 0.9950
Valla (text English) inhouse-text@9
human average 0.045 · AI average 0.909 · balanced accuracy 0.980
Muybridge (video) inhouse-video@2
bag p90 real average 0.004 · AI average 0.940 · balanced accuracy 0.985
Helmholtz (audio) inhouse-audio@2
real average 0.004 · AI average 0.984 · balanced accuracy 0.986

Text by language

Shared TMR + Fakespot features; one text model per language. Same absolute floors as English (human ≤ 0.12 · AI ≥ 0.85 · bal. acc. ≥ 0.90). Unsupported languages are not scored.

Language Model Human ↓ AI ↑ Bal. acc.
English Valla inhouse-text@9 0.045 0.909 0.980
Spanish Valla inhouse-text-es_v2 0.088 0.953 0.972
Portuguese Valla inhouse-text-pt_BR_v1 0.103 0.945 0.928
French Valla inhouse-text-fr_v2 0.082 0.976 0.975
Italian Valla inhouse-text-it_v1 0.069 0.986 0.972
German Valla inhouse-text-de_v1 0.064 0.968 0.958
Dutch Valla inhouse-text-nl_v1 0.101 0.978 0.933

Portuguese (pt) uses the Brazilian-trained text model. Full methods and limits are in the PDFs above.

Need programmatic scores? See the Pro API docs (keys from Dashboard).

Model cards and open weights on Hugging Face (Apache-2.0). Everyday scoring still runs through DotCheck: huggingface.co/DotCheck · vermeer-image-v12 · valla-text-v9 · muybridge-video-v2 · helmholtz-audio-v2 · DotCheck Engines collection.