← back to timeline

EXP-SEED-ALT

artifact

Seed Content Effect — Does Fitness Landscape Depend on Prompt Content?

2026-03-23 L3 level paper3 1 runs $0.00
new
premortem
mock
real
metrics
analyze
review
krit
done

In plain language

All our experiments used abstract, poetic seed texts. What if we use concrete, scientific seeds instead? Same three AI agents, same rules — but different starting material. If the hierarchy changes, our findings depend on WHAT agents discuss, not just HOW they're connected.

What we found

UNDERPOWERED (N=1): No predictions scored

Predictions we made before running

0/5 confirmed

Technical details

Research hypothesis

OPEN QUESTION: "prompt-dependent fitness landscapes: does prompt cycling create frequency-dependent selection? if yes → diversity maintenance mechanism."

PRIOR DATA: All 86 experiments use abstract/poetic seed texts (S1-S3). Zero data on whether hierarchy, role assignment, or convergence patterns depend on seed content. If they do → the fitness landscape IS prompt-dependent, and our findings are seed-content-contingent. If they don't → hierarchy is a robust structural property independent of content domain.

THIS EXPERIMENT: Replace abstract seeds with CONCRETE/TECHNICAL seeds (biology, physics, everyday objects). Same topology (complete N=3), same personas, same params. Compare TR hierarchy, VP, role persistence to exp-asym PERSONA_LIVE baseline.

CRITICAL TEST: Does Gamma (grounding persona: "notices what's real, specific, observable") become a STRONGER source with concrete seeds? In exp-asym with abstract seeds, Gamma's analytical grounding had no natural substrate. With concrete seeds, Gamma's cognitive style matches the content domain → persona×content interaction.

BASELINE DATA: exp-asym PERSONA_LIVE (N=9, abstract seeds): TR_range: 0.444 ± 0.15 VP_z: 5.67 Role persistence: significant (p=0.005) Gamma role: receiver (low-TR end)

GROUNDING: Kauffman fitness landscape theory. Different substrates = different fitness landscapes = different attractors. If LLM agents' vocabulary hierarchy depends on seed content, then seed = environmental parameter that shapes the selection landscape for word usage.

Experimental setup

Type: simple

Condition Parameters
CONCRETE_LIVE seed_type: CONCRETE, interaction: LIVE, note: Same PERSONA_LIVE as exp-asym but with concrete/scientific seeds.

Factors: seed_type (CONCRETE)

Parameters

n_agents
3
n_rounds
16
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).