controls and baselines foil

tries to explain the result away

builds the dullest explanation it can, from sequencing depth to the target’s own baseline, and tests whether the result still stands.

see the whole lab

call it: a result needs the boring explanation it has to beat, or a score might only be tracking depth.
it can: read and search the code, run commands, write and edit files, and edit notebooks.model · same as your session
hands off: running the analysisbench

the pre-registrationseal

the verdictassay

foil comes with the lab how a study runs

claude plugin install agentmash-lab@agentmash

first time on this machine? add the marketplace before it: claude plugin marketplace add agentmash-dev/agentmash-crew

foil at work

what it does, how, what you get back, who it hands off to and when to call it, in 23 seconds.

each one does one job. open one to see when to call it.