← back to timeline

EXP-REVORD-N5

informative

Reversed Ordering N=5 — Does First-Mover Advantage Persist in C2+ Data?

2026-03-21 L3 level paper3 1 runs $0.01
new
premortem
mock
real
metrics
analyze
review
krit
done

In plain language

We reversed who speaks first in a 5-agent conversation. Cycle 1's hierarchy perfectly inverted (pure position artifact), but in later cycles Alpha stayed on top regardless of position. The "analytical" persona creates more exportable language — it's cognitive style, not turn order.

What we found

UNDERPOWERED (N=1): No predictions scored

Predictions we made before running

0/4 confirmed

Technical details

Research hypothesis

POSITION INVARIANCE TEST: Does agent ordering affect C2+ hierarchy metrics?

CONTEXT — BURN-IN ARTIFACT: We discovered that Cycle 1 in round-robin multi-agent experiments is a UNIVERSAL measurement artifact. The first speaker has maximum export (TR=1.0), the last has zero. TR_range in Cycle 1 correlates r=0.999 with a positional null model.

After burn-in correction (C2+ only), Kruskal-Wallis across densities: p=0.983 (FLAT).

BUT: in exp-full-n5, Alpha (first speaker) retains HIGHEST TR even in C2+ data: Alpha C2+ TR = 0.6452 (highest) Delta C2+ TR = 0.3971 (lowest) C2+ TR_range = 0.2481

CRITICAL QUESTION: Is Alpha's C2+ advantage a RESIDUAL first-mover effect, or is it due to Alpha's persona (analytical) being inherently more "exporting"?

EXPERIMENT: Run identical config but with REVERSED agent ordering. Original order: Alpha, Beta, Gamma, Delta, Epsilon Reversed order: Epsilon, Delta, Gamma, Beta, Alpha

DISCRIMINANT: IF position-driven (first-mover residual): → Epsilon (now first) should become highest TR in C2+ → Alpha (now last) should become lowest TR in C2+ → C2+ TR_range should be similar (~0.16-0.25) but hierarchy REVERSES

IF persona-driven (intrinsic to cognitive style): → Alpha should STILL be highest TR regardless of position → Ordering doesn't change C2+ hierarchy

IF truly transient (C1-only artifact): → C2+ metrics should be similar to original, order-invariant → No systematic relationship between position and C2+ TR

GROUNDING: This is a classic confound resolution experiment. Position and identity are confounded in single-order designs. Counter-balancing (reversing order) is the standard solution.

Experimental setup

Type: simple

Condition Parameters
FULL_N5_LIVE topology: FULL_N5, interaction: LIVE, note: Complete graph with REVERSED agent ordering. Every agent sees all 4 others.

Factors: topology (FULL_N5)

Parameters

n_agents
5
n_rounds
50
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).