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.

iteration#009best resultBPB 0.984-1.4%frontier held
Graph search expands the space.

Mutations branch from prior candidates, sweeping protocol and parameter variants before the run chooses which route deserves verification.

Mutome turns an uncertain research candidate into a measured run.

A candidate can be a protocol, patch, proof strategy, parameter setting, simulation setup, or experiment variant. Mutome records what was tried, how it was evaluated, which artifacts were produced, and why the candidate was kept, hardened, or discarded.

Discovery stays broad: models, search heuristics, and researchers can propose many plausible routes across the search space.

Verification stays narrow: a candidate advances only when fixed tests, benchmark scripts, solvers, proof checkers, or reviewers reproduce the claim, leaving a command, metric or check, artifacts, and an evidence trail another researcher can inspect.

Randomness in discovery.Determinism inverification.

Three steps. Many loops.

Mutome runs repeatable research loops where candidates are proposed, evaluated against fixed criteria, and folded back into the next search frontier with enough context to inspect or replay the run.

01

Propose

Generate candidate protocols, patches, proof routes, and parameter settings for the same research question.

Propose research plate illustration
  • protocol and architecture mutations
  • parameter, prompt, and data variations
  • baseline and parent candidate recorded
  • rationale and inputs captured
02

Verify

Evaluate each candidate with predefined commands, metrics, tests, or verifier outputs.

Verify research plate illustration
  • failed checks and noisy lifts rejected
  • replay command stored with the run
  • metric direction declared in advance
  • BPB, Spearman, or domain metrics attached
  • promotion tied to a specific gate
03

Iterate

Promote what improved, learn from what failed, and start the next search from a cleaner frontier.

Iterate research plate illustration
  • kept improvements become new parents
  • failed routes become negative evidence
  • lineage stays replayable across loops
  • artifacts and decisions remain attached

Every result carries the evidence gate it actually reached.

00
Candidate recorded research plate illustration
Candidate recorded

the proposed change or research route is written down with its objective and parent context

01
Replay reproduced research plate illustration
Replay reproduced

fixed tests, benchmark scripts, or replay commands reproduce the measured result

02
Independent tool checked research plate illustration
Independent tool checked

a solver, verifier, simulator, or external benchmark checks the narrow claim where applicable

03
Formal artifact checked research plate illustration
Formal artifact checked

machine-checkable proof files or formal certificates support the result where the domain allows it

04
Researcher reviewed research plate illustration
Researcher reviewed

a researcher reviews attribution, limitations, reproduction steps, and the final decision

A candidate is ready only when its evidence can be replayed, compared, and challenged.