← back to timeline

EXP-CLOSURE

success

Behavioral Closure Test — Does Structural Hierarchy Persist Across Radically Different Seed Content?

2026-03-16 L2 level paper3 3 runs $0.01
new
premortem
mock
real
metrics
analyze
review
krit
done

In plain language

We've shown that differently-minded AI agents form stable hierarchies when communicating — but always with the same abstract poetic text. What if we give them mundane everyday topics instead? If the same hierarchy appears, it means the structure comes from the agents' interaction, not from the content.

What we found

UNDERPOWERED (N=3): No predictions scored

Predictions we made before running

0/4 confirmed

Technical details

Research hypothesis

OPEN QUESTION: "agent-level theoretical grounding: what is 'food set' at agent level? what threshold for behavioral closure?"

Gershenson E/S/C framework predicts that genuine emergence produces process-level invariants: the SAME structural patterns arise regardless of the specific content being processed. This is analogous to RAF closure in chemistry: the reaction network is self-sustaining regardless of which specific molecules are present.

ALL experiments so far use abstract/poetic seed vocabulary ("crystalline vapor", "unnamed frequencies", "membrane of intent"). If the hierarchy (TR range ~0.444), role differentiation, and information asymmetry we observe are genuine emergent properties of the multi-agent system, they should appear with CONCRETE/EVERYDAY seeds too.

If structural metrics are seed-INVARIANT → behavioral closure EXISTS at agent level. The "food set" = any coherent text input; the "closure" = structural patterns (hierarchy, role differentiation) are preserved.

If structural metrics are seed-DEPENDENT → hierarchy is content-driven, not process-driven. The food set matters, and we don't have closure.

BASELINE: exp-asym PERSONA_LIVE (N=9+, abstract seeds S1-S3): TR_range: 0.444 ± 0.15 VP_z: ~5.67 Role persistence: significant (interaction effect p=0.005)

This is a CHEAP PILOT (1 run × 3 seed sets = 3 runs × 16 rounds × 3 agents = 144 API calls).

Experimental setup

Type: simple

Condition Parameters
CONCRETE_LIVE seed_type: CONCRETE, interaction: LIVE, note: Same persona setup as exp-asym PERSONA_LIVE, different seed content domain

Factors: seed_type (CONCRETE)

Parameters

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