visual-jev

# Visual Jev

Open, local, image-native typed decisions using a Qwen3-VL multimodal backbone, isolated packed question branches, a learned pointer head, language-backbone LoRA, and held-out temperature calibration.

The production path accepts one or more images, optional structured context, and independent `choice`, `score`, and `noul` questions. It returns numerical distributions without generating answer text. A candidate-token Qwen path remains available only as an explicit uncalibrated baseline.

## Run

```bash
uv sync --extra dev
npm --prefix web ci
npm --prefix web run build
uv run visual-jev --registry models --model visual-jev-public-mps-bf16-2026-03-v3
```

A promoted checkpoint must exist at `models/visual-jev-public-mps-bf16-2026-03-v3`. Missing, experimental, incompatible, or checksum-mismatched weights fail closed; the server never substitutes baseline or random weights.

See [the operations guide](docs/OPERATIONS.md) for the immutable dataset contract, controlled head-only and candidate baselines, LoRA training, untouched calibration/test evaluation, architecture probes, honest generative benchmark, exact-input comparison against pinned OpenJev, evidence-based direction gate, promotion, registry installation, external-backend adoption, and cold-start release workflow. The canonical typed request is [`examples/canonical-request.json`](examples/canonical-request.json).

## Verify

```bash
uv run ruff check src tests
uv run pytest
npm --prefix web run typecheck
npm --prefix web run build
```