EXP-CA-RULE-MULTIGEN-POC

… pending

CA + multi-gen rule heredity hybrid substrate pilot (paper6 step 3)

2026-05-04 L6 level paper6 2 runs $0.03

new

premortem

mock

real

metrics

analyze

review

krit

done

In plain language

This experiment tested if a system that learns rules could still work when one part of its "brain" was a simple grid-based simulation instead of a complex language model. Researchers observed how the system proposed new rules and whether it focused on features specific to the grid simulation or only on the types of features it learned from the language model. The results are still being analyzed.

▶ Technical details

Research hypothesis

Step 2 established that coupling-shaped developmental program produces canalized rule lineages (depth ~7) with substrate-internal rule heredity. Question: does this generalise when one substrate fragment is NON-LLM? Conway's Life CA dynamics are not in the LLM training distribution. Coupling agents to a CA grid (observe patches, optionally mark cells) while running multi-gen rule proposals tests whether the coupling principle survives the substrate change. Pilot — N=2 runs, 3 generations, 30 rounds/gen. Plumbing + first signal, not strong claims. Pre-registered observations to make: - Q1: do agents propose on CA-specific axes (patch_size, mark_allowed, cross_patch_view) or only on the LLM-only axes from step 2?

Q2: does the signal_type chain (progress→combined→concept) replicate, or does CA presence shift it?
Q3: do agents reference CA observations (gliders, oscillators, density) in their rule reasoning?
Q4: does grid density evolve non-trivially or stagnate?

Experimental setup

Type: simple

Condition	Parameters
COUPLED	coupling: ALL, n_runs: 2

Factors: coupling (all)

Parameters

n_agents: 4
n_rounds: 90
model: google/gemini-2.5-flash-lite
temperature: 0.85

autopoiesis paper6-step3 hybrid-substrate ca-coupling multi-generation pilot