jev-evolve

September 21, 2026 · View on GitHub

Self-improving agent experiments where every decision is a typed TypeSafe Jev question: evolve a policy from the agent’s own mistakes and report how much measured gain is selection noise versus real improvement.

At a glance	Details
Source	Source
Maintainer	novaleolin. Independently curated; this page is not an upstream submission or endorsement.
Format	Python package jev-evolve on PyPI (`pip install jev-evolve`); depends on `evalfloor`.
Requirements	Python ≥ 3.9. Live agent runs need TypeSafe (or documented backends). Optional `local` extras for torch/transformers. Offline unit tests need no API key.
License	MIT. TypeSafe/provider usage has separate costs.
Disclosure	AI-assisted catalog review; no affiliation. Listing is not an endorsement. Offline pytest run; live evolution loops and billed Jev not run.

When to use

Use it when studying policy improvement with typed decisions and honest selection-bias reporting. Prefer simpler harnesses (jev-harness, daf-jev) when you only need gates/evals without evolution. Distinct from open-weight lookalikes that imitate Jev’s interface without TypeSafe.

How it works

Upstream API centers on Policy, typed choice/noul builders, run_policy, and measurement helpers (permutation_sensitivity, Marginalized, confusions, point_accuracy, overconfident, cost). Agents return tool results or {jev_evolve.STOP: True} to finish. Typed Jev answers drive control flow instead of free-form generated text; code owns execution and stop conditions. README documents comparative latency notes for typesafe/jev-1.13 versus other backends (upstream measurements, not re-run here).

Get started

pip install jev-evolve
python -c 'import jev_evolve; print(jev_evolve.__doc__[:80] if jev_evolve.__doc__ else "ok")'

Pinned review checkout:

git clone https://github.com/novaleolin/jev-evolve.git
cd jev-evolve
git checkout 3a937d42f2e0a7fa19795ad0881f347a20a9defa
pip install -e '.[dev]'
pytest -q

Follow upstream Quickstart for a live policy loop (incurs provider charges).

Examples and demos

Upstream README Quickstart / API sections and bilingual docs (README.zh-CN.md).
tests/test_jev_evolve.py.
This listing ran pytest -q: 30 passed. No live TypeSafe evolution run.

Limits and data handling

Live runs send episode state and questions to the configured decision backend. Treat README benchmark tables as upstream reports. Optional local-model extras are separate from TypeSafe Jev.

Review and maintenance

Reviewed on 2026-09-21 at commit 3a937d4: MIT; AI-assisted source review of README, LICENSE, package layout and tests; pytest 30 pass. No live TypeSafe call.