DotCheck/muybridge-video-v4_4

Muybridge@4.4 is DotCheck’s live video pair stack (inhouse-video@4): Muybridge 3.8 on the transport view plus Blau 1.4 on the native center crop, combined by one stack, then stamp-cap (max) across stamps. This repo is new. It does not replace muybridge-video-v2 (archive).

Scores are estimates under this model, not proof of authorship. Public Layer A is homework at cut 0.5.

Field	Value
Hub id	DotCheck/muybridge-video-v4_4
Wire id	inhouse-video@4
Label	Muybridge@4.4
Artifact	siglip2_base_patch16_224_linear_head_v4_4.npz
Arms	…_v3_8.npz (transport) · …_v1_4.npz (center) · shared …_v6_6.npz
Spines	same 13.12 + 6.6 as Vermeer@14.2
Backbone	google/siglip2-base-patch16-224 (Apache-2.0)
Stamps	pair2: n=2 inner 33%/67% when duration ≤ 180s; else min(12, max(2, floor(dur/60)))
Fold	stamp-cap max (not bag p90)
Encode	same pair views as stills (max-side 256 + native 224)
Output	p ∈ [0,1] — estimated P(AI)
Serve	CPU FastAPI POST /v1/analyze-pair kind=video

Model description

Two shared SigLIP2 towers and a stack combiner. Same spines as Vermeer@14.2. Live on inhouse-video@4. Clip fold is stamp-cap max, not bag p90.

Product clip scores that fuse frames with audio use a private seating path (Covenant: Muybridge + Helmholtz). Do not read the homework table as a fused eyes-and-ears claim.

In this repo: README.md, LICENSE, NOTICE, CITATION.cff, the 4.4 stack .npz, the video arm heads, the shared center arm, and the two spine safetensors.

Architecture

clip → pair2 stamps
  each stamp raster (min side ≥ 224; else refuse)
    → transport: max-side 256, aspect kept, JPEG q70
          → 13.12 spine → v3_8 logit
    → center: native 224×224 crop, JPEG q85
          → 6.6 spine  → v1_4 logit
    → stack_w · [z_v, z_j] + stack_b → sigmoid → p_stamp
  clip p = max(p_stamp)

Inference

Open weights: the stack .npz, video arm heads, shared center arm, and spines (Apache-2.0). Product scoring: Check or Pro API.

Not AutoModel.from_pretrained("DotCheck/…"). Everyday scoring runs through DotCheck.

Training data

Holdout real n=602 / AI n=180. See eval/MUYBRIDGE_4.md.

Evaluation

Homework Layer A at thr 0.5. Source: RULER_video_v4_4.json and video_gates_v4_4.json.

Metric	n	Measured
mean P(AI) real	602	0.022
mean P(AI) AI	180	0.976
balanced accuracy		0.9859

Gold 164×25 is a Vermeer exam, not a 4.4 public claim.

Intended use

Video pair scores. Citation of the homework table. Not a v2 overwrite.

Limitations

Two-spine serve. Homework is Layer A, not a posterior over every generator. Stamp-cap is not a temporal model.

License

Apache License 2.0 for DotCheck heads in this repo.

Citation

Prefer Muybridge@4.4 and https://dotcheck.ai/docs.