← back to timeline

EXP-SEED-SENSITIVITY

artifact

Seed Sensitivity at N=6 — Attractor Structure vs Position Confound

2026-03-10 L1 level paper3 16 runs $0.08
new
premortem
mock
real
metrics
analyze
review
krit
done

In plain language

With six agents, hierarchy is unstable — sometimes strong, sometimes weak. Is this due to speaking order (whoever speaks last gets suppressed) or does the seed text determine which "regime" the system falls into? 16 runs with different texts and orders separate these two explanations.

What we found

No predictions scored

Predictions we made before running

0/6 confirmed

Technical details

Research hypothesis

COMPETING EXPLANATIONS FOR N=6 VARIANCE:

N=6 shows high TR range variance across 4 runs (SD=0.158, range 0.311-0.672). Two explanations:

H1 (POSITION CONFOUND): Variance is driven by speaking position effects. Different rotations place different agents in "suppressed" positions, changing who has low TR and therefore changing TR range. The r=-0.90 correlation between bottom-agent TR and overall TR range in exp-n6 supports this.

H2 (ATTRACTOR STRUCTURE): Different initial conditions (seeds) pull the system toward genuinely different dynamical attractors. Some seeds lead to high differentiation, others to low differentiation, independent of rotation. This would be a real bistability signal.

DISCRIMINATING TEST: 4 seeds × 4 rotations factorial design. Two-way ANOVA on TR range:

  • If F(seed) >> F(rotation): H2 supported — attractor structure exists
  • If F(rotation) >> F(seed): H1 confirmed — position confound dominates
  • If both significant: both mechanisms contribute
  • If neither significant: variance is pure noise (no systematic source)

ARCHITECTURAL VALIDITY: This test works in a memoryless-between-runs system because we're testing whether different INITIAL CONDITIONS (within a run) lead to different dynamics. No between-run state persistence required. (Verified per candidate improvement: unlike hysteresis test, which was killed.)

Experimental setup

Type: simple

Condition Parameters
SEED_SENS persona: PERSONA, interaction: LIVE, n_agents: 6

Factors: n_agents (6) × seed_variation (S1, S2, S3, S4) × rotation (0, 1, 2, 3)

Parameters

n_agents
6
n_rounds
48
n_runs_per_condition
16
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).