EXP-ASO

✓ success

Symmetric Observation Experiment

2026-03-08 L3 level paper3 5 runs $0.03

new

premortem

mock

real

metrics

analyze

review

krit

done

In plain language

Three identical AI agents communicate in a shared text field. Does a hierarchy emerge — who influences whom? Words do propagate between agents, but "who leads" is simply determined by who speaks first. Without differences between agents — no genuine structure.

What we found

VP real, structure = first-mover artifact

VP_excess is real — identical agents develop directional vocabulary propagation from interaction alone (p=0.015). But structural metrics (S, diversity) show no significant differences. The upstream position is a first-mover artifact (confirmed in v3 seed randomization test: 3/3 runs, forced first-speaker becomes upstream). Structure that looked like emergence is actually an artifact of speaking order.

Predictions we made before running

3/5 confirmed · 2 refuted

▶ Technical details

Research hypothesis

Identical agents in shared text space develop directional vocabulary propagation (VP) from interaction alone. VP_excess(LIVE) > 0 indicates that shared context creates asymmetric influence even without persona differences. Controls test whether structure requires live interaction.

Experimental setup

Type: multi_condition

Condition	Parameters
ISOLATED	label: Isolated (private memory per agent), n_runs: 3
LIVE	label: Live interaction (shared memory), n_runs: 3
REPLAY	label: Replay (fixed sequence, no live generation), n_runs: 3
SHUFFLED	label: Shuffled agent names between rounds, n_runs: 3
TEMPORAL_SHUFFLE	label: Temporal shuffle (randomized round order), n_runs: 3

Parameters

model: gemini-2.5-flash-lite
n_agents: 3
n_rounds: 50
n_runs_per_condition: 3
scheduler: round_robin
temperature: 0.9

Trophic Ratios by Condition

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

agent symmetric vocabulary-propagation first-mover