metadata
license: other
tags:
- dotcheck
DotCheck
We train detection models, measure them on holdouts we publish, and ship the product that runs them.
This Hub org is the open side of that work: Apache-2.0 open weights (.npz heads on frozen upstream backbones), model cards, and the same gate numbers we use in production. Scores are estimates, not verdicts. The tables are what we measured. Use them as you will.
See what's real.
| Hub id | Wire | Backbone / stack | Artifact |
|---|---|---|---|
vermeer-image-v14_2 |
inhouse@14 |
Pair stack: max-side 256 + native 224 | siglip2_base_patch16_224_linear_head_v14_2.npz |
valla-text-v12_2 |
inhouse-text@12 (+ lang heads) |
TMR + Fakespot + logistic heads | text_stack_head_v12.npz + lang |
muybridge-video-v4_4 |
inhouse-video@4 |
Pair stack + stamp-cap max | siglip2_base_patch16_224_linear_head_v4_4.npz |
helmholtz-audio-v3_1 |
inhouse-audio@3 |
Dasheng-Base + logistic head | dasheng_base_linear_head_audio_h3e.npz |
- Collection: DotCheck Engines
- About:
DotCheck/dotcheck - Inventory:
REGISTRY.mdin this about repo /hub/REGISTRY.mdin source
Product
Everyday scoring is the DotCheck product, not a Hub demo:
- Site: https://dotcheck.ai
- Guest Check: https://dotcheck.ai/check
- Pro API: https://dotcheck.ai/api (
Authorization: Bearer dc_…) - Method and gates: https://dotcheck.ai/docs · https://dotcheck.ai/how
Open weights and cards live here. Clip assembly that joins video frames with audio (Covenant) stays in the product path.
Holdout claims (SSOT)
Measured holdout floats live in product Data.json / MODEL-CHOICE.md (image bal 0.9924, EN text bal 0.979, video stamp-cap bal 0.9859, audio bal 0.995). Do not cite stale bake-off YAML.