EXP-COUPLING-NULL-MODEL

… pending

Coupling null model — real vs random state-block volume control

2026-05-01 L6 level paper5 6 runs $0.04

new

premortem

mock

real

metrics

analyze

review

krit

done

In plain language

This experiment tested whether agents learning together performed better because they were genuinely interacting or simply because they had more information to process. Researchers compared agents learning in a real, connected environment to agents learning with randomly shuffled information, keeping the amount of information the same. The results are still being analyzed to determine if the observed improvement was due to real interaction or just increased data.

▶ Technical details

Research hypothesis

Paper5 full experiment: COUPLED 0.721 vs UNCOUPLED 0.392 (d≈8.7). Adversarial check: is the 2x progress effect due to genuine coupling SIGNAL (real neighbor states), or just increased context VOLUME per round? COUPLED-RANDOM: agents receive state-block with SHUFFLED/RANDOM states (same volume, no real signal). If RANDOM ≈ REAL-COUPLED → effect is volume. If RANDOM ≈ UNCOUPLED → effect is genuine signal. Pre-registered: COUPLED-REAL > COUPLED-RANDOM ≈ UNCOUPLED. (genuine coupling, not volume artifact)

Experimental setup

Type: factorial

Condition	Parameters
LEVIN_REAL	coupling: true, coupling_mode: real, goals_mode: diverse, n_runs: 2
LEVIN_RANDOM	coupling: true, coupling_mode: random, goals_mode: diverse, n_runs: 2
LEVIN_NONE	coupling: false, coupling_mode: none, goals_mode: diverse, n_runs: 2

Factors: coupling_mode (real, random, none)

Parameters

n_agents: 4
n_rounds: 40
model: google/gemini-2.5-flash-lite
temperature: 0.85

levin-bioelectric null-model volume-confound-check retraction-prevention paper5-adversarial