OpenJev

Context → questions and candidate answers → probabilities → your application's next step.

OpenJev is a local decision interface and a collection of controlled learning experiments. Score choices with stable IDs, try text decisions in your browser, or explore recorded chess, Doom and robot policies.

Text demo · Chess replay · API · Setup · Experiment archive

Decision interface

Supply English context, your questions, and candidate IDs with descriptions. OpenJev returns each chosen ID and the relative probability of every supplied answer. Your application decides what to do next; the API itself executes no external action.

The local scorer uses a pinned Qwen model, maps candidates to single-token labels, and scores those labels without generating an explanation. Questions are independent. Stable IDs constrain the response format, not whether the choice is correct.

Use POST /api/decide with the X-OpenJev: 1 header, or openjev decide request.json. The complete example and API reference cover request limits, errors, model metadata and execution boundaries.

Run locally

Apple Silicon macOS, Python 3.11-3.13:

git clone https://github.com/kw2828/OpenJev.git
cd OpenJev
uv sync --frozen --extra language
uv run --extra language openjev setup
uv run --extra language openjev serve

Open http://127.0.0.1:8000. Setup downloads the pinned Qwen3-4B weights; inference then uses the local cache. No paid API key is needed.

The free browser demo runs Qwen3-0.6B through WebGPU. Its scores are not interchangeable with the local 4B model. Browser requirements · Linux, Docker and hosting.

Chess

Four small policies score legal moves across twelve fits and 288 games. Exact next-board differences raise ordinary move agreement from 31.75% to 33.15%, but cost more computation and fail the engine-loss and game-score criteria. The GIF is the first scheduled game, not a selected win. No learned world model or Elo rating is established.

Results, all checkpoints and controls · ChessBench transfer · Connectome comparison · Graph-quality comparison.

Doom

13 kills in 19.26 game seconds, from the first fit and first evaluation seed. JEPA supplies training rewards; a small policy plays this recording. The full nine-method comparison did not establish a reliable JEPA advantage. This is one illustrative episode, not an inference-speed benchmark.

Results and limits · Recording receipt · Other Doom controllers.

Robot reaching

The stronger memory comparison failed its rule: 24/25 checks passed. Persistent GRU lowers control cost by 4.44% / 2.94% versus a separately trained two-observation controller; the rule requires at least 3% on both sensing-gap panels. Supplied-physics controllers still perform better.

The GIF is from the earlier, weaker-control study. The stronger comparison retains all nine models, five references and complete costs. Neither establishes a new architecture or biological-wiring advantage.

Latest completed research

Spatial features did not beat the simpler scalar controls: 15 fits, five model families, 52,800 updates. The 225-parameter statistics model has the lowest exposed-validation mean squared error in every paired seed. The spatial model's mean error is 4.93% higher, with 3.64x the fitting time. These are prediction diagnostics; the separate autonomous evaluation has completed and awaits its audit.

All models and limits · Protocol · Raw evidence and checkpoints.

Independent replay checked all 100,470 final predictions. All fifteen models also passed deployment checks, agreeing on 780 action choices across shared public and boundary states. The autonomous comparison completed all 1,152 searches; its saved-output audit is running, so the performance verdict remains pending. The earlier 504-search input-scaling comparison remains failed on all 30 conditions. No novel architecture advantage is established.

Earlier dialogue, RockSample, compact-memory and pretrained-policy comparisons are collected in the experiment archive, including failed attempts, frozen criteria and independent audits. The separate shared-prefix scoring experiment measured 2.08x speedup on four-question synthetic workloads; single-question caching was slower. This does not reproduce a proprietary RLCD system.

Limits and provenance

Candidate probabilities are uncalibrated and relative to the supplied choices. Rewording or changing the choices can change the scores.

These demos use different models and protocols. Replays illustrate preserved episodes; they do not replace complete comparisons. The research has not established a novel architecture, a biological-wiring advantage or a new RL algorithm.

OpenJev is independent of TypeSafe Jev, zhihz/openjev and openjev.com. It does not reproduce proprietary Jev architecture or weights.

The root MIT license applies except where component notices specify other terms, including GPL-3.0-only chess research files. Third-party data and model licenses apply separately.