Propose
Explore different ways to answer the same research question.
- Try a new formula, experiment, or change to the code
- Explain what changed and why it might help
- Save the starting point so each idea can be compared fairly
An autonomous, self-improving lab for frontier science and engineering.
Mutome tackles open-ended problems in mathematics, physics, engineering, and technical R&D by running many candidate research routes in parallel. AI models and search heuristics explore broad spaces inspired by evolutionary algorithms, while every promising protocol, code change, proof route, or parameter setting becomes a replayable run with metrics, artifacts, and recorded decisions.
Each idea becomes a bounded, measured attempt.
An attempt can be a formula, a code change, a protocol, or a statement written in Lean. Its question, evaluator, and budget are fixed before it starts. Afterwards, Mutome keeps what was tried, how it was judged, what it produced, and why it was kept, discarded, or left inconclusive.
Discovery stays broad: language models and search heuristics, inspired by evolutionary algorithms, propose many possible routes.
Verification stays narrow: an attempt advances only when a frozen evaluator, a reference given the same budget, or the Lean kernel confirms it.
Randomnessin discovery.
Determinismin verification.
Three parts of one lab.
Slean checks claims exactly and measures Mutome itself. Lab runs the campaigns. Benchmark compares each candidate with a reference.
SleanExact worlds in Lean, exactly checked claims, labs with hidden answers.Cellular automata- LabThe private console where bounded campaigns are planned, run, and approved.Private, nothing published
BenchmarkCompare a candidate with a reference at equal budget, under a frozen evaluator.Oscillators and Slean labs
Two kinds of research.
Fundamental research proves or refutes a law. Engineering improves a frozen metric. Both start on small worlds, on purpose.
Three steps. Many loops.
Mutome proposes ideas, tests them, and uses the results to decide what to try next. Each attempt is saved so others can understand what was done and repeat the test.
Verify
Test each idea to find out whether it actually works.
- Decide what counts as an improvement before running the test
- Compare the result with the starting point and a simpler approach
- Keep the test instructions and results so the check can be repeated
Iterate
Use what the tests reveal to choose what to try next.
- Build on ideas that passed the checks
- Keep failed attempts so the next search can learn from them
- Save the steps and reasons behind each decision
See what has actually been checked.
A promising idea, a repeated result, and a checked proof are different things. Each published result will say which of these checks it has passed.
Idea recorded
The idea and its purpose have been saved. This tells us what to test, not whether it works.
Result repeated
Running the saved test again produces the measured result. The instructions are kept so the check can be repeated.
Checked with a separate tool
A separate test, simulation, or verification tool checks the specific claim. Its conclusion applies to what it tested.
Proof checked by software
For a claim stated in Lean, the Lean kernel checks the proof against the stated definitions and assumptions.
Reviewed by a researcher
A researcher examines the evidence, sources, and limitations to assess what the result supports and what still needs testing.