Paul’s learnings

Visual Decisions.
Pixels to Probabilities.

Test multimodal System One models on images. Upload screenshots, receipts, or robotics feeds, edit typed question rubrics, and inspect decision distributions.

Multimodal System One: Traditional vision-language models decode multi-token descriptive prose character-by-character (taking 1.8–4.8 s and costing $0.35–$2.06 per 1k). Visual decision models project image tokens directly into bounded option logits in a single forward pass—achieving 0.26–0.71 s latency and $0.0074–$0.063 per 1,000 decisions on specialized Jev-class vision deciders.

Decision Engine & Gate Policy

Notice: UI and schema validation simulation. Connect a local vision server (e.g. llama.cpp / vLLM / Kev serve) or multimodal API endpoint below to run actual model weights.

Threshold policy: Choice gates on confidence; Noul accepts positive when p ≥ threshold and negative when p ≤ (1 − threshold); Score checks confidence across level mass.

Visual Input & Questions

Drag & drop an image here, click to browse, or paste directly with Ctrl+V
Visual decision input preview

Decision & Distribution

Ready. Select a preset or upload an image and click "Run vision decision".

JevImageBench v0.1.5: 50 Vision Decision Systems Ranked

Independent benchmark data from benchmarkheaven.com/image-jev-bench evaluating 50 vision models on bounded multimodal decision tasks across speed, cost, intelligence, and calibration.

View use-case catalogue · Read the decision models report · Run text deciders on-device