← back to timeline

EXP-TOPO-CYCLE

artifact

Cycle Topology vs Complete — Does Communication Structure Affect Hierarchy?

2026-03-14 L3 level paper3 1 runs $0.00
new
premortem
mock
real
metrics
analyze
review
krit
done

In plain language

Three AI agents communicate in a text space. Normally everyone sees everyone, and hierarchy emerges — someone becomes a word "source," someone a "receiver." But what if each sees only one other? Alpha reads only Gamma, Beta — only Alpha, Gamma — only Beta. A circle. Will hierarchy emerge without direct access to everyone?

What we found

UNDERPOWERED (N=1): No predictions scored

Predictions we made before running

0/5 confirmed

Technical details

Research hypothesis

OPEN QUESTION: "cyclicity confound: sparse but highly cyclic networks may still prevent hierarchy. need to pre-screen BOTH density AND cycle fraction."

CONTEXT: All prior experiments use COMPLETE topology — every agent reads every other agent's contributions. Hierarchy (trophic structure) emerges reliably (exp-asym N=9, TR_range=0.444). But is hierarchy an artifact of complete visibility, or does it form even in restricted topologies?

DIRECTED CYCLE: A reads only C, B reads only A, C reads only B. Each agent has in-degree=1, out-degree=1 (uniform, no hubs). Information propagates transitively: A→B→C→A.

HYPOTHESIS: Cycle topology REDUCES hierarchy (TR_range) relative to complete topology because: 1. No agent can accumulate DIRECT positional advantage over all others 2. Influence is mediated — A affects C only through B 3. Uniform degree distribution prevents hub-formation

BUT hierarchy may persist partially because transitive influence creates temporal advantage: early vocabulary innovations propagate around the cycle, and the originator is always 1 step ahead.

BASELINE DATA: Complete topology (exp-asym PERSONA_LIVE, N=9): TR_range = 0.444 ± 0.15 Full decay complete (exp-forget, N=1): TR_range = 0.015

This is the FIRST of 4 designed topologies (cycle/star/chain/complete) for the TC analysis program. Cycle first because it directly tests cyclicity confound.

Experimental setup

Type: simple

Condition Parameters
CYCLE_LIVE topology: CYCLE, interaction: LIVE, note: Directed cycle — each agent reads only predecessor

Factors: topology (CYCLE)

Parameters

n_agents
3
n_rounds
48
n_runs_per_condition
1
model
gemini-2.5-flash-lite
temperature
0.9
scheduler
round_robin

Trophic Ratios by Condition

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