Lazarus

Lazarus¶

Turn dead research code into a callable pipeline component — and give the revival back.

The wall¶

Computational science has a reproducibility problem. A huge fraction of published methods are open, cited, and unrunnable within a few years: the repo is stale, wired to a stack that no longer resolves, and the real capability is buried in scripts with no API. The exact method you need exists — but getting it to run costs days you don't have, so it gets abandoned.

What Lazarus does¶

Lazarus is an agent that revives dead research code, lets you compose the revivals into pipelines, and gives the fixes back to the community.

Revive — point it at a bare GitHub URL. A web-enabled Scout reads the repo + paper and writes its own goal and a falsifiable sanity check; the agent then runs a build → run → read-traceback → repair loop in a sandbox and emits a fixed integration contract (importable module, CLI, pinned container, smoke test).
Compose — every revival emits the same contract, so a revived tool is a composable brick. Wire bricks from any domain/language/era into a pipeline with a little YAML.
Give back — the fixes (rotted URLs, broken paths, a 15-year-old undefined-behaviour bug) become maintainer-ready pull requests with CI, so the method can't silently rot again.

And a registry of everything already revived — 13 tools today — so you can lazarus pull one instead of re-reviving it. The flip side ships too: lazarus decay-check answers "does a repo still run today?" agent-free — as a CLI and a GitHub Action.

Measured — most of these repos are dead, and Lazarus revived them all¶

On a principled, seeded random sample of 20 Bioinformatics tools (2018–2021):

85% don't install or run on their own today (95% CI 64–95%)
Lazarus revived 100% of the dead ones — 17/17 (95% CI 82–100%)
5 reproduced the original paper's reported metric

Nothing is cherry-picked — the frame, seed, per-repo outcomes, and the runnable harness are in benchmark/.

Six of them, up close¶

Repo	Era / stack	Result
MaSIF-site	Py3.6 · TF 1.12 · MSMS/APBS	interaction sites, ROC-AUC 0.9137
ScanNet	Py3.6 · TF 1.14 · Keras	binding sites, ROC-AUC 0.9233
dMaSIF	torch cu111 · PyKeOps · GPU	binding sites, ROC-AUC 0.8390
fpocket	2010 C on modern GCC	3 druggable pockets
Basset	2016 Lua Torch7 · genomics	from a URL → reproduced the paper, AUROC 0.894 vs 0.895
DiffDock	2023 PyTorch diffusion · ESM-2 · GPU	blind docking → reproduced ~40% top-1 (<2Å)

Each was revived from its own dead environment using only general heuristics — no repo-specific notes — and emits a package that passes its own smoke test standalone. The hard-won details are in the hard problems it solved.

Prefer to just try it? The dashboard lets you search a repo, watch a recorded resurrection replay, and browse the registry.

Install¶

pip install lazarus-bio

The base install (dependency pinner, contract/compose tooling) needs nothing but Python ≥ 3.9. The autonomous loop + Scout need the agent extra and Docker:

pip install "lazarus-bio[agent]"

Next: the Quickstart, or how it works.