Jev Projects Added This Week

Open bilingual mmBERT-based 144.3M decision model returning Choice, Noul, and 2–10 Score answers with probabilities through a separate local PyTorch runtime.

Open 4B and 35B-A3B video decision models that share a video prefix across questions and directly score temporal-grounding intervals and multiple-choice answers without text generation.

Local SQLite BM25 retrieval pipeline where Jev answers batched Noul evidence-relevance questions to rerank passages before a separate model generates an answer; selected text is sent to hosted providers.

Open clients, evaluation tools, and runnable text/image/video decision demos for the hosted System One-compatible Trio-Spark service; model weights and training code are closed.

Pilot research monitor where Jev uses typed Noul and Score judgments to keep, review, or drop new papers and repositories against open research questions, while another model writes daily notes.

Batch and streaming log-analysis application where Jev or Laya answers configurable typed questions about events, with validated numeric results and optional Python actions.

Coding-agent plugins where Jev or a compatible System One model judges shell risks, task complexity, context retention, and tool exposure, with capabilities varying by host agent.

Python client for typed Choice, Score, and Noul decisions across Jev, OpenRouter, OpenAI Decisions, and local models, with confidence-based fallback and a System One-compatible server.

VisiData plugin that sends each text-column row to Jev or another OpenRouter decision model and adds typed labels, scores, probabilities, and confidence as columns.

Experimental pre-action gate that applies deterministic rules before asking Jev or a local compatible model whether an unresolved tool action should proceed or require confirmation.

Reproducible guardrail benchmark that compares Jev and other decision systems using typed Noul checks and a fixed 0.5 blocking threshold, with public scoring contracts and author-run results.