Lab

One directory, one command

A lab is a directory that contains two things: a task description and the ./lab command. An agent, whether Claude Code, Codex, a Mutome worker or a person, touches the worlds only through that command.

./lab worlds                          # the list of worlds
./lab experiment w3 0120010 5         # run world w3 from 0120010 for 5 steps
./lab check w3 0 12 8                 # does block 12 on background 0 return within 8 steps?
./lab propose '{"kind":"conservation","world":"w3","w":1,"f":[0,1,2]}'
./lab status                          # budget used and accepted results

Everything is metered

• Every proposal costs 1, accepted or not.
• Every experiment costs the number of cell updates it asks for: number of cells times number of steps.
• When either budget runs out, the lab refuses further work.

Metering experiments is essential. Without it, brute-force search looks free and wins every benchmark.

The answers stay outside

The hidden answers used to score the run are never in the lab directory. The agent’s transcripts are then audited for any access outside the lab. That audit is a marker scan plus a review of the tool calls, not an operating-system sandbox.

Two modes

• Whitebox: each world’s table is given. It is like knowing the laws of physics and looking for their consequences.
• Blackbox: the rules are hidden. The agent learns only through metered experiments, as in experimental science.

What the first trial gave

In blackbox mode, a Claude Sonnet agent found every law and more structures than a scripted reference, with 150 to 225 times fewer cell updates and no false claim.

These worlds are still too easy: one well-built experiment reveals their entire rule. See A hidden world rebuilt. The guide Create a lab gives the full commands.