← back to timeline

EXP-DILUTION-N7

informative

Dilution Test — Is N=7 Hierarchy Collapse from Rounds-per-Agent or Genuine Regime Change?

2026-03-10 L2 level paper3 4 runs $0.01
new
premortem
mock
real
metrics
analyze
review
krit
done

In plain language

With 7 agents, hierarchy collapsed (TR=0.30 vs 0.53 with 5 agents). But maybe it is not about the number of agents, but that each got too few rounds? We give each of the 7 agents 14 rounds instead of 7 — if the issue is insufficient rounds, hierarchy should recover.

What we found

UNDERPOWERED (N=1): No predictions scored

Predictions we made before running

0/4 confirmed

Technical details

Research hypothesis

QUESTION: The N=7 hierarchy collapse (TR range = 0.297) has two competing explanations:

(A) DILUTION: N=7 round-robin gives each agent only ~7 rounds (49/7). At N=5, each agent gets ~10 rounds (48/5). Insufficient per-agent data compresses trophic differences. If we give N=7 agents 14 rounds each (matching N=5 density), hierarchy should recover to ~0.50.

(B) GENUINE REGIME CHANGE: N=7 creates a "vocabulary soup" where too many speakers erode unique vocabulary fractions. More rounds won't help because the mechanism is agent COUNT, not per-agent exposure. TR stays ~0.30 regardless of rounds.

Prior evidence:

  • exp-hept: N=7, 49 rounds (7 rds/agent), TR = 0.297
  • exp-may-sigma: N=7, 28 rounds (4 rds/agent), TR = 0.418 (single run!)
  • N=5 (exp-pent): 48 rounds (9.6 rds/agent), TR = 0.528

NOTE: exp-may-sigma's TR=0.418 at FEWER rounds/agent (4) contradicts simple dilution theory. But N=1 run, unreplicated.

This experiment: N=7, 98 rounds (14 rounds/agent). If dilution is the cause, 14 rds/agent should exceed N=5's 9.6 rds/agent → TR ≥ 0.50.

Experimental setup

Type: simple

Condition Parameters
DILUTION_N7 persona: PERSONA, interaction: LIVE, n_agents: 7, n_rounds: 98, note: Same 7 agents as exp-hept. 98 rounds = 14 rounds/agent. Compare to exp-hept (49 rounds, 7 rds/agent).

Factors: n_agents (7) × rounds_per_agent (14)

Parameters

n_agents
7
n_rounds
98
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).

Series (4 experiments)