Slean.
Exact worlds, exactly checked claims, and labs with hidden answers.
Slean (Science + Lean) is Mutome’s verification and measurement foundation. Its worlds are defined in Lean. A claim about them counts only when it is checked exactly: laws and structures by the Lean kernel, mechanisms by exact comparison. Its labs score an agent against answers it cannot see.
Runs today on cellular-automaton worlds.
See an example
A claim counts only when it is checked exactly.
Slean defines its worlds in Lean, accepts a claim only when it is checked exactly, and measures explorers against answers they cannot see.
Worlds you can state exactly.
Slean stands for Science + Lean. Lean is a language in which a definition is exact and a proof is checked by a small trusted program, the kernel. Slean uses it to define whole worlds: today, cellular automata, grids of cells that change by a fixed local rule.
Because the rule is exact, a claim such as “this quantity is conserved” is either provable or false. There is no grey zone to argue about.
A proof or a counterexample, checked exactly.
Slean accepts a claim only when it is checked exactly: a law or a structure needs a proof the Lean kernel checks, and a proposed mechanism must match the world’s rule by exact comparison. A refutation comes with a counterexample, which is checked too.
Every accepted result joins a library of verified results. Later work can build on it without redoing it or taking it on trust.
Labs with hidden answers.
In a Slean lab, an agent studies a world whose laws it does not know. It can act only through a ./lab command that meters every experiment. The answers stay hidden until scoring.
Its score is compared with a scripted reference that gets exactly the same budget. This is how Mutome measures its own progress.
What we have measured.
Six measurements, including the ones that do not flatter us. The details and the data are published on slean.org.
Read the measurementsWorlds with hidden rules: the agent edges past the reference.
On generated worlds whose rules were hidden (200 proposals, 200,000 cell updates), a Claude Sonnet agent found every hidden conservation law and slightly more structures than a scripted reference: 70 of 70 against 68 on one seed, 78 of 87 against 75 on another, plus 2 beyond the reference’s bounds. It used 150 to 225 times fewer experiments and made no false claim. But each of these worlds can be identified with a single experiment, so this is not open-ended discovery.
Lessons that changed almost nothing.
Measured by Mutome’s engine on held-out labs, adding lessons from earlier labs changed the results by 0 and by +1 discovery. That battery was saturated: it could no longer show improvement.
Harder worlds: the raw agent falls behind.
On “compressible” worlds, where a 1,024-entry rule hides a short law and the experiment budget is tight, a first run left the raw agent far behind a scripted reference: 3 of 48 discoveries against 12. It never tried to infer the short law.
With lessons, the agent improves on unseen worlds.
On three held-out seeds, the raw agent found 4, 6 and 5 verified discoveries; the same agent with lessons kept from earlier labs found 13, 13 and 17. We repeated the whole comparison three times: lessons won in 8 pairs out of 9, by 6 discoveries per lab on average, and lost one pair by 6. This is the first change to Mutome’s harness that improved results on worlds it had never seen. A general “method” instruction was erratic (10, 4 and 36).
Compact laws: lessons help induction.
The agent can now submit a short formula that computes a world’s whole rule; the Lean kernel checks it on all 1,024 neighbourhoods. With lessons, the agent found 8, 12 and 8 of the 12 hidden laws per lab; without them, 2, 6 and 5. One run per arm and seed.
A smarter script draws level.
A scripted reference that tries the textbook families of rules (linear, totalistic) found 9 laws and 47 to 54 discoveries per lab. With lessons the agent is level with it (50, 59 and 46); without them it falls behind on two seeds out of three. The next worlds must use families outside that list.
A checked claim still has a scope.
A kernel check proves a statement about a defined world. It does not show that a similar law holds in nature, and the worlds measured so far are still too simple to speak of open-ended discovery.
Learn it by playing
slean.org explains these ideas in plain language, with four mini-games in which you look for the laws of a world yourself.
Try the mini-games