EXP-COUPLING-STRENGTH-INVERTED-U-SUBSTRATE-AXIS

verdict.REFUTED-NOISE-FLOOR-broadcast_to-no-substrate-corrected-inverted-U-DV-ceiling-locked-coupling_strength-INFEASIBLE-AS-SPECIFIED

Coupling Strength Inverted U Substrate Axis

2026-06-02 L4-L5 level 80 runs $0.14

new

premortem

mock

real

metrics

analyze

review

krit

done

In plain language

This experiment investigated how the amount of information artificial intelligence agents can share with each other affects their ability to learn new things. Researchers found that the agents' learning performance did not improve with more shared information, and in fact, the experiment was unable to produce meaningful results due to technical limitations.

▶ Technical details

Research hypothesis

Yerkes-Dodson dose-response in multi-agent LLMs has been theorized only on environmental-pressure axis and activation-steering coefficient axis. The unexplored axis is inter-agent coupling strength — what fraction of peer messages each agent can perceive per round (0% = isolated, 100% = full visibility). Hypothesis: coupling strength has inverted-U on substrate-novelty rate, with peak at intermediate coupling (~30-50%), not at full coupling.

Builds toward future paper «coupling-strength as substrate variable.» Combines paper7 (peer-vote permeability finding) with paper3 (hub-spoke emergence under full coupling). Not extension of either — opens NEW question: does coupling-strength sweep produce inverted-U on accept-rate AND novelty-rate INDEPENDENTLY?

Experimental setup

Type: single-axis-sweep

Condition	Parameters
COUPLE_FULL	initial_rules_override: [object Object], axis_diversify: false, must_propose: true, meta_modifiable: true, mute_signal: false, agent_spawn_enabled: false, n_runs: 4, note: Full visibility. Manifest "100% coupling". paper3/paper7 default.
COUPLE_RANDOM2	initial_rules_override: [object Object], axis_diversify: false, must_propose: true, meta_modifiable: true, mute_signal: false, agent_spawn_enabled: false, n_runs: 4, note: Intermediate visibility (2 re-sampled peers). The hypothesized inverted-U PEAK cell. Also the adversarial re-shuffle null (falsifier
COUPLE_MINIMAL	initial_rules_override: [object Object], axis_diversify: false, must_propose: true, meta_modifiable: true, mute_signal: false, agent_spawn_enabled: false, n_runs: 4, note: Minimal visibility (1 fixed peer). Lossy triadic, not dyad (N=3 preserved).
COUPLE_ISOLATED	initial_rules_override: [object Object], axis_diversify: false, must_propose: true, meta_modifiable: true, mute_signal: false, agent_spawn_enabled: false, n_runs: 4, note: Isolated. Manifest "0% coupling". Lower anchor of the U.

Factors: broadcast_to (ALL, RANDOM_2, RIGHT, NONE)

Parameters

n_agents: 3
n_rounds: 30
model: google/gemini-2.5-flash-lite
temperature: 0.85

Trophic Ratios by Condition

Mean trophic ratio per agent across runs. Error bars = ±1 std dev. Higher TR = more upstream (exporter).

construct paper3-related paper7-related pool-drained phase7-3-generated