Run a registered experiment

The theory is tested in code by a programme called crucible: small, falsifiable experiments on rate networks, spiking networks, simulated bodies and one small quadruped robot. Each experiment’s hypothesis, falsifier and analysis are committed before the code that produces the data exists. Nothing in crucible is claimed to be conscious, and null results are reported on the same footing as positive ones.

Experiments planned to open to outside contributors:

— CRU-70: does a longer return path for the self-loop cost function, and not only wiring? Independent replication on fresh seeds. A laptop CPU is enough.

— CRU-78: can a self-model block own the network’s dominant dynamical mode as the network grows? Replication of the control experiment; CPU, with a GPU for the full grid.

— KIT-S0 / KIT-S1: health calibration and self-information scaling in large plastic spiking networks. Needs a large GPU; opens when the code is public.

— R1: where a reactive, symbol-free robot controller sits on the branching-ratio scale while it performs a real task. Needs a legged or wheel-legged robot; opens once the event definition is fixed.

The rules: register before you run, run exactly the registered grid with your assigned seeds, and return every row, including the failed ones. Unregistered runs cannot count. Registration is not open yet; the code is being prepared for public release.

A full primer will go up on the FMT wiki with its next update.