Tools for transparent, reproducible and verifiable science.

lightconeresearch.org · documentation

Computational research is full of methodological choices that quietly shape the results: how a dataset is filtered, how a model is parameterized, what cutoff defines an outlier. These choices are often defensible, often debatable, and almost always invisible by the time the paper is written.

We build tools that surface those choices, make them explicit, and let researchers explore the alternatives systematically — all the way from a research question down to materialized, integrity-checked results. Most of that machinery is driven by AI agents now, so a researcher can stay focused on the science instead of the plumbing.

A specification language and validator for analyses. One astra.yaml declares your inputs, outputs, methodological decisions, and the prior insights from the literature that justify each option. Includes the astra CLI, machine-verified evidence quotes, and the schema everything else builds on.

The lc command line, plus a Claude Code plugin (skills, agents, hooks) that turns an astra.yaml into materialized results. Recipes run through Snakemake on a Dask cluster — local laptop, SLURM allocation, or external scheduler. Every output gets a content-addressed manifest so lc verify can prove what you have on disk is what was actually computed.

uv tool install lightcone-cli
lc init my-analysis && cd my-analysis
claude
# then inside Claude Code start scoping a research question with
/lc-new

The user guide walks you from there to your first reproducible result.

• One spec, one truth. astra.yaml is the single source of truth. Every input, output, recipe, and decision lives there.
• Decisions, not defaults. Every meaningful methodological choice is a named decision with options, a default, a rationale, and (where possible) literature evidence behind it.
• Provenance you can prove. Every materialized output ships with a sidecar manifest recording the recipe, the container image, the decisions, the input chain, and content hashes. Tampering and drift are detectable, not hopeful.
• Agents do the typing, you do the science. The Claude Code plugin (/lc-new, /lc-migrate, /lc-feedback) handles the implementation grind so researchers can stay on the questions that matter.

• Found a bug? – two ways to report
  • Write an issue report using our provided templates.
  • Inside any Claude Code session, run /lc-feedback <what went wrong> and we get a clean issue with the context auto-attached.
• Looking for the deeper docs? Maintainer documentation lives inside the main docs or in (docs/).
• Want to contribute?
  We welcome contributions across the stack — schema design, the agent surface, the execution engine, the docs, the rough edges you hit when you try to use any of this for real work. Each repo has a docs/contributing/ directory with development setup and conventions. Smaller fixes are welcome as PRs without prior discussion; for larger changes, opening an issue first is appreciated.

Our code ships under BSD 3-Clause unless a repository says otherwise. We aim to keep things permissively licensed so they can be used freely in academic and commercial research alike.