Skip to content

Repository files navigation

LibreEvolve

LibreEvolve

Evolve code. Show the evidence.

LibreEvolve is an open-source workbench for inspectable code-evolution experiments. Engineering preview: bounded Python bin-packing optimization with Codex OAuth.

It starts from a deliberately simple Python bin-packing heuristic, searches a bounded local space of changes, and saves every candidate workspace and evaluation so you can read the evidence yourself — before a fresh export check re-runs the candidate you chose.

The product is an experiment you can examine, not a promise that every search improves code. No improvement is a valid result. A successful named check is not proof of universal correctness, optimality, security, or performance.

What makes it worth inspecting

Bounded by design Local call, evaluation and runtime limits keep a run small enough to follow. They are not token or dollar caps, and subscription usage may stay unknown.
Saved, not summarized Candidate workspaces, evaluations, archive events and LLM calls land in a run directory you can read.
Honest by construction Run status, validity, quality, retention and usage stay separate, so a finished run is never mistaken for a win.
Checked again before use Reports and exports re-run the selected candidate in fresh local subprocesses instead of trusting old metadata.

What is available today

  • A source install for the public engineering-preview checkout.
  • Local initialization and an offline task/evaluator preflight that makes no provider request.
  • One documented model lane: Codex OAuth with gpt-5.6-luna and high reasoning. A local login does not prove account-specific model access.
  • Bounded runs whose saved artifacts can be inspected with their status, validity, quality, retention and usage kept distinct.
  • A local HTML report and an export that freshly checks the selected candidate on named training and held-out cases.

An approved inspectable example viewer is not included in this branch. The seed-only flow in the quickstart is an onboarding check, not an optimization result or live-model demonstration.

Research features, additional providers, plugins, papers and experimental tools are out of scope for this checkout; they are not capabilities of this public engineering preview.

How a run moves

  1. Seed. Initialize the bundled bin-packing task and inspect its intentionally simple heuristic.
  2. Change. Propose and evaluate bounded local changes — optionally through the Codex OAuth lane, after you authorize your own subscription use.
  3. Inspect. Read the saved run. Completion, validity, training quality, retention and recorded usage are separate dimensions, so a finished run is not the same as a retained win.
  4. Verify. Freeze your selection, then let a fresh export check it on named training and held-out cases and record its identity.

Run the preview through the canonical quickstart.

Start with the canonical quickstart

The complete clone, source-install, offline, seed-only and live authorization path is maintained in the engineering-preview quickstart. It is the command source for the other docs; use a fresh destination whenever the CLI says an output already exists.

Before authorizing subscription use, read safety and scope and how to read a result. Candidate execution has host access and is not a security sandbox. Local call and runtime bounds are not dollar or hard-token limits; subscription charges and quota may remain unknown.

Support boundary

The package metadata declares Python >=3.11. The documented reference route uses Python 3.12 on Linux/WSL. Windows installation and offline checks are documented, while native Windows Codex execution and external usability remain unvalidated. See the support matrix in the quickstart for the evidence boundary.

Read, contribute and release

Contributions should preserve the engineering-preview scope, keep training and held-out evaluation separate, and include focused offline tests. The project does not advertise a public package-registry install; install from a checked-out source tree as described in the quickstart.

Project links

MIT license.

About

Open-source workbench for inspectable code-evolution experiments on bounded Python bin-packing.

Topics

Resources

Code of conduct

Contributing

Security policy

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages