Many hard problems do not fail because nobody can imagine a next step. They fail because the space of possible next steps is too large, too noisy, and too expensive to evaluate with discipline. A new protocol might help. A code patch might improve a benchmark. A proof route might unlock a lemma. A simulation setting might reveal a more stable regime. Most of these candidates will be wrong, weak, or not worth keeping. A few may move the frontier.
Mutome exists to make that search systematic. It runs broad exploration across candidate protocols, code changes, proof strategies, parameter settings, simulation setups, and experiment variants. Models, scripts, solvers, search heuristics, and researchers can all propose the next move. The system records where a candidate came from, what changed, how it was evaluated, and what artifacts were produced.
Why Mutome.
The name compresses several nearby ideas: mutation as a search operator, atom or atome as a small unit of matter, autonomous as self-directed execution, and the -ome suffix from systems such as genome or proteome: a whole map, not a single sample.
Mutations explore. Atoms keep the unit small. The -ome suffix suggests a complete record. Autonomy points to research loops that can run, measure, and improve without losing evidence.
A visible proof point.
Andrej Karpathy's autoresearch is a useful reference point: an agent edits a compact nanochat training setup, runs short experiments, checks validation loss, keeps improvements, reverts failures, and repeats. In a March 2026 update, Karpathy described roughly 700 autonomous attempts producing about 20 additive changes and reducing nanochat's "Time to GPT-2" benchmark from 2.02 to 1.80 hours.
Mutome takes the same core lesson seriously but points it at a broader research problem: many candidates can be generated by models, scripts, solvers, search heuristics, or humans, but only candidates that leave reproducible evidence should move the frontier.
Discovery should be broad.
Open-ended research needs variance. It needs many plausible routes, including routes that look strange before they are tested. Mutome treats exploration as a graph: candidates branch from prior candidates, mutations expand the search frontier, and failed attempts remain useful because they narrow the space. This is inspired by evolutionary search, but aimed at research work where the object is not only a score. The object is a candidate with lineage, context, metrics, logs, and enough structure to be inspected later.
Verification should be narrow.
A candidate is not promoted because it sounds convincing. It advances only when it passes the evidence gate declared for the run: a replay command, benchmark script, solver output, proof checker, simulator, reviewer decision, or domain metric. The metric direction is explicit. The baseline and parent are recorded. The artifacts are attached. The decision is written down as keep, harden, or discard.
This separation is the core principle behind Mutome: randomness in discovery, determinism in verification. The system can search widely, but the path from a claim to an accepted result must be replayable. It is not enough to say that a result improved. Mutome should be able to show what changed, how the result was measured, whether it beat the parent, and what evidence supports the decision.
Why this matters.
Scientific and technical progress increasingly depends on exploring spaces too large for a single researcher to enumerate by hand. AI systems can help generate candidates, but generation alone is not a research lab. A lab needs memory, constraints, baselines, artifacts, negative evidence, and reproducible decisions. Mutome is designed to be that workbench: a place where autonomous exploration can create many possibilities while the research standard remains strict.
The first version of Mutome focuses on the loop itself: propose, verify, score, and iterate. Over time, that loop becomes a record of the frontier. What was tried? What failed? What improved? What can be replayed? What deserves a harder check? These questions are the difference between a promising trace and a result another researcher can challenge.
Mutome is for hard problems in mathematics, physics, engineering, and technical R&D where exploration can be broad, but acceptance must remain precise. The goal is not to replace scientific judgment. The goal is to give that judgment a better machine: one that searches wider, remembers more, and promotes only what leaves evidence behind.