Experiments

Archive Organization

The experiment codebase is now organized around the distinction this appendix needs: study records versus implementation packages. A study is the canonical scientific unit — title, status, seeds, result location, and key finding. A package is runnable code that several studies may share.

Canonical index: experiments/registry.json

experiments/studies/ — numbered records like 022_intrinsic_predictive_gradient and 027_nonlinear_mlp_prediction_head.
experiments/packages/ — runnable code packages: ca_affect, llm_affect, c_computational, and inevitability_empirical.
— validates that every study record exists and every referenced code file is present.

Read each section as an experiment in an evolving chain, not as an isolated result. ask what a cellular automaton can do without agency. asks what changes once action-observation loops exist. isolate the decomposability problem in prediction architecture. test whether social targets, direct selection, contrastive losses, or communication solve it. Most do not. The nulls are load-bearing.