Roadmap¶
Lazarus set out to prove one thing: that an agent can take dead research code — a bare GitHub URL, nothing else — and bring it back as a callable, verified component. As of v0.4 that's proven and measured:
- It works, and it generalizes. 40+ repositories revived across biology and a dozen other fields — from materials science and plasma physics to hydrology and retrosynthesis — at ~92% on both peer-reviewed and unreviewed code.
- The problem is real and quantified. A controlled study (peer-reviewed vs. unreviewed research software) shows unreviewed code is packaged half as often (42% vs. 95%) and fails to install today far more (61% vs. 37%) — yet is just as recoverable. The reviewed/unreviewed gap is packaging discipline, not recoverability.
- The revivals are public. A registry of 25 tools (23 pullable from GHCR), each with a verified contract, plus fixes sent upstream as pull requests.
That's the foundation. This document is where it goes next.
It's tiered on purpose. Tier 1 is the near-term direction — what we'd build next. Tier 2 is what Lazarus becomes if it grows beyond a research project into standing infrastructure. Nothing here is locked; it's an invitation. If one of these is what pulls you in, open an issue or a PR.
Tier 1 — The next chapter¶
Three directions, best pursued as a pair: one flagship demonstration for reach, and one durable investment that compounds.
Flagship · Real science from resurrected bricks¶
The compose/contract layer already lets revived tools snap together — but the vision (revival as
a supply chain for new work, not a museum) is under-demonstrated. Build a genuine multi-tool
pipeline entirely from tools that were dead a month ago — e.g. structural biology:
pocket detection (fpocket) → docking (DiffDock / EquiBind) → interface classification
(PRODIGY-CRYSTAL) → scoring — and produce a real result.
Why it matters: it's the most quotable proof there is, and it lands hardest with the scientists
who feel the reproducibility pain. Good first steps: pick a question answerable by 3–4 registry
bricks; wire the contracts into one compose pipeline; write up the result and the fact that every
component was unrunnable a week prior.
Durable · The reproducibility observatory¶
Turn the one-off benchmark into a continuous service: decay-check sweeping a large corpus (all
of JOSS, bioRxiv-linked repos), auto-reviving the dead ones, the registry growing to hundreds. A
public dashboard becomes a live map of what science is rotting and what's been brought back.
Why it matters: it makes the reproducibility crisis visible and actionable, every revival is a
public good, and it's an endless source of data for follow-on studies. Good first steps: schedule
decay-check over a seeded corpus; persist verdicts; a minimal dashboard over the results.
Durable · A benchmark for reviving dead research code¶
The instrument already exists — seeded frames, an agent-free baseline, verified outcomes. Package it as a public leaderboard: SWE-bench for resurrecting dead scientific software.
Why it matters: it recruits the whole AI-agent community to the problem, makes Lazarus the reference implementation, and is a clean second paper. Good first steps: freeze a held-out repo set + a scoring harness (install → run → verify); publish a submission format and a baseline.
Tier 2 — If Lazarus becomes more than a research project¶
The tracks you invest in once one of the above builds momentum — turning a capability into a tool people use, and a harder one.
Adoption · Meet researchers where they are¶
Ship the dashboard ("search a repo, watch it revive live"), a GitHub App / one-click "revive this
paper's code," and integrations (Zenodo, or JOSS running decay-check on submission). Turns a
capability into a daily tool — which brings users, feedback, hard real-world cases, and the
give-back loop at scale.
The technical frontier · Harder revivals¶
Today's bar is install + run + verify. Push it to full training reproduction (not just inference), data- and credential-gated tools, multi-node / large-scale jobs, and new languages — there are mountains of dead scientific Fortran, MATLAB, and Julia nobody has touched. Each one widens the set of science Lazarus can bring back.
Get involved¶
- Throw a dead repo at it. The best test cases come from the wild — open an issue with a URL
that won't run, or try
lazarus resurrect <url>yourself. - Adopt a revival. Browse the registry,
lazarus pull <name>, and tell us where it breaks. - Pick up a direction above. Any of these is a good first contribution — say hi in an issue and we'll help scope it.
This roadmap is a living document; it will change as the work and the community do.