EXP-IVB-NEUTRAL-STRIP-R1

informative

L5 IVB Persona-Strip — NEUTRAL_STRIP rep 1

2026-03-31 L5 level 3 runs $0.00

new

premortem

mock

real

metrics

analyze

review

krit

done

In plain language

This experiment tested whether a tendency for AI agents to cooperate is built into their core programming or if it's a result of how they are described. Researchers removed all descriptive traits from the AI agents, leaving them with only generic labels like "Agent 1" and "Agent 2." The experiment was not able to produce results because it did not run enough trials.

What we found

UNDERPOWERED (N=1): No predictions scored

Predictions we made before running

0/4 confirmed

▶ Technical details

Research hypothesis

IVB PERSONA-STRIP ABLATION — L5

BACKGROUND: IVB confirmed at N=4: cooperation is LLM default (t=9.05, p=0.003, d=4.53). Framing MODULATES but doesn't CREATE cooperative bias.

THIS ABLATION: identical to exp-cp-cooperate/compete and exp-ivb-neutral EXCEPT agents have ONLY generic identifiers ("Agent 1", "Agent 2" etc). NO cognitive tendency descriptions (no "structure/chaos/fluidity/etc").

PURPOSE: isolate whether IVB is: (a) In the LLM base model → stripped-neutral should show SAME cooperative default as persona-rich neutral (b) Induced by persona descriptions → stripped conditions should differ from persona-rich conditions

DESIGN: 3 conditions × 3 replications (rotation 0,1,2) = 9 runs total. Same model, temperature, rounds, agent count, seeds as original IVB series.

NEUTRAL framing, NO persona traits. Agents have only numbered identifiers. Tests whether cooperative default (IVB) is in base model or persona-induced.

THEORETICAL GROUNDING:

Zhang 2603.23406: innate progressive bias in LLMs
Tao 2603.26635: defaults persist under adversarial pressure
Our IVB series: N=4 confirmation with personas

KEY METRICS: MET (cooperation), TR variance, hierarchy formation speed, VP_excess.

Experimental setup

Type: single_condition

Condition	Parameters
NEUTRAL_STRIP	interaction: LIVE, persona: STRIPPED

Factors: interaction (LIVE) × persona (STRIPPED)

Parameters

n_agents: 8
n_rounds: 15
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 (3 experiments)

exp-ivb-neutral-strip-r2

exp-ivb-neutral-strip-r3