← back to timeline

EXP-STAR-HUB-SWAP

artifact

Star Hub Swap — Is Hub-as-Sink Structural or Persona-Dependent?

2026-03-21 L3 level paper3 9 runs $0.04
new
premortem
mock
real
metrics
analyze
review
krit
done

In plain language

Three AI agents in a star network: rotating which personality holds the hub position across 9 runs (3 per condition). Hub is word SOURCE in 7/9 runs, not sink — reversing the N=1 finding. Grounding persona dominates as hub (3/3 source). Hub position amplifies persona tendencies; structure × personality interaction confirmed (F=7.61, p=0.023, eta²=0.72). Neither topology nor persona alone predicts hierarchy direction.

What we found

UNDERPOWERED (N=3): No predictions scored

Predictions we made before running

0/5 confirmed

Technical details

Research hypothesis

OPEN QUESTION: "degree distribution operationalization: need SPECIFIC threshold — density < X% → F₀_excess > Y AND degree heterogeneity > Z → hierarchy. without this, mechanism = unfalsifiable."

TENSION: Two findings create a paradox: 1. exp-topo-star (N=3): Alpha (hub, analytical) = SINK (TR=0.409, LOWEST) 2. exp-revord-n5 (complete, N=5): Alpha (analytical) = SOURCE (TR highest, r=0.70 across orderings)

Alpha is SOURCE in complete topology but SINK in star topology. Is hub-as-sink STRUCTURAL (any persona becomes sink when hub)? Or is it PERSONA × TOPOLOGY interaction (Alpha specifically absorbs when hub)?

EXPERIMENT: Three conditions — rotate which persona is hub: (A) ALPHA_HUB (baseline = exp-topo-star replication) (B) BETA_HUB (associative persona as hub) (C) GAMMA_HUB (grounding persona as hub)

DISCRIMINANT: IF all hubs become sink → STRUCTURAL: hub position causes sink regardless of persona. IF only Alpha-hub is sink → PERSONA: analytical style absorbs, topology is context. IF mixed → INTERACTION: specific persona × position combinations matter.

BASELINE DATA: Star N=3, Alpha-hub (exp-topo-star, N=1): hub TR=0.409 (SINK), TR_range=0.141 Complete N=3 (exp-asym, N=9): TR_range=0.444 Complete N=5 reversed (exp-revord-n5, N=1): Alpha TR highest, persona-driven

GROUNDING: Classic within-subjects confound resolution. Position and persona were perfectly confounded in exp-topo-star (Alpha=always hub). Rotating the hub role across personas separates the effects.

Experimental setup

Type: factorial

Condition Parameters
ALPHA_HUB_LIVE hub_agent: ALPHA_HUB, persona: PERSONA, interaction: LIVE, note: Replication of exp-topo-star. Alpha=hub, Beta+Gamma=spokes.
BETA_HUB_LIVE hub_agent: BETA_HUB, persona: PERSONA, interaction: LIVE, note: Beta=hub (reads Alpha+Gamma). Alpha+Gamma=spokes (read only Beta).
GAMMA_HUB_LIVE hub_agent: GAMMA_HUB, persona: PERSONA, interaction: LIVE, note: Gamma=hub (reads Alpha+Beta). Alpha+Beta=spokes (read only Gamma).

Factors: hub_agent (ALPHA_HUB, BETA_HUB, GAMMA_HUB) × persona (PERSONA) × interaction (LIVE)

Parameters

n_agents
3
n_rounds
48
n_runs_per_condition
3
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).