Docs v2026.10

Technical docs

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

Download the official Model Card and Technical Report for the engines behind every DotCheck score. Tests on data the models did not train on, in plain language.

Measured results

Tests on data the models did not train on. Everyday scoring uses one image engine, one video-frame engine, one audio engine, and eight text heads. Seven share Rask and Comenius; Chinese is Ma Jianzhong.

Primary engines

Vermeer@14.2 image

Pair stills · Niepce + Janssen

Real
0.025
AI
0.983
Balanced acc.
0.9924

Valla@14.3 text · English

Writing · Rask + Comenius + a named tiny per Latin language

Human
0.029
AI
0.984
Balanced acc.
0.979

Muybridge@4.4 video

Pair frames · Plateau + Blau · stamp-cap

Real
0.022
AI
0.976
Balanced acc.
0.9859

Helmholtz@3.1 audio

Sound · trained head

Real
0.018
AI
0.989
Balanced acc.
0.995

English numbers here are the RAID-lite exam (Wikipedia counts; news is watched only). A one-generator English test set is not the public claim.

Text by language

English, Spanish, Portuguese, French, Italian, German, and Dutch share Rask and Comenius, each with a named tiny head. Chinese is Ma Jianzhong. One text model per language. Same absolute floors (human ≤ 0.12 · AI ≥ 0.85 · bal. acc. ≥ 0.90). Unsupported languages and Traditional-Chinese-dominant text are not scored.

Language Engine Human ↓ AI ↑ Bal. acc.
English Valla@14.3 · Sweet 0.029 0.984 0.979
Chinese Valla · Ma Jianzhong 0.047 0.922 0.954
Spanish Valla · Bello 0.029 0.993 0.982
French Valla · Littre 0.021 0.990 0.978
Portuguese Valla · Ribeiro 0.042 0.964 0.958
German Valla · Grimm 0.034 0.966 0.978
Italian Valla · Ascoli 0.016 0.943 0.972
Dutch Valla · Vries 0.056 0.963 0.950

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

Open weights and cards on Hugging Face (Apache-2.0). Everyday scoring still runs through DotCheck:

Chrome, Android, Check, and the Pro API call the same Express path. Same engines. The Technical Report is about those engines, not a fifth model. API keys are minted on Check. Confirm is first-party only - not via the API.