Jev Architecture Research

An evidence-first archive of experiments that study TypeSafe's Jev as a black-box decision system. It is not a leaderboard and the README states it makes no claim to Jev's weights, training data, source code or internal traces — only behavioural evidence and caveats.

The strongest behavioural attribution combines shared state/option representation, dynamic semantic option encoding, global semantic competition, typed choice/score/noul heads, a ranking-like decision layer and batch-aware inference. Hidden provider batching, top-k reduction, reranking, the backbone and the training objective remain unresolved.

It is described as a controlled cluster experiment whose outcome favoured global joint semantic scoring over strong cluster-first filtering. The README lists it as the strongest experiment in the archive while leaving hidden provider batching, top-k reduction and reranking as open questions.

No. It studies behaviour from the outside and marks each conclusion observed, inferred or speculative. It preserves API limits as limits rather than converting them into model claims, and it does not use provider output to train a surrogate model.

A master report with the full timeline, an evidence matrix, a hypothesis tracker, competing architecture attributions, a failure surface covering negation, contradiction, collision and language failures, latency and calibration analyses, open questions, a final meta-analysis, a machine-readable knowledge base and experiment receipts.

The original scripts and curated result files live under research/experiments/. Reproduction requires an applicable Jev API account and an environment variable for the key; secrets are never committed, and the README advises checking the receipts before running a costly or rate-limited probe.

Experiment code and scripts are MIT licensed, while the research notes and curated data are CC BY 4.0 under the repository's LICENSE and DATA-LICENSE.md files. The README also includes a BibTeX citation for the 2026 archive.