EXP-MMP-CLAUDE

◐ partial

Multi-Model Pilot: Claude Haiku 4.5

2026-04-13 L3 level paper3 6 runs $0.22

new

premortem

mock

real

metrics

analyze

review

krit

done

In plain language

Does hierarchy emerge in Claude Haiku? Identical agents, 30 rounds, LIVE vs ISOLATED.

What we found

Claude Haiku 4.5 via OpenRouter: VP_excess positive in LIVE (0.079 > 0), role persistence LIVE > ISOLATED (0.067 vs 0.04

Claude Haiku 4.5 via OpenRouter: VP_excess positive in LIVE (0.079 > 0), role persistence LIVE > ISOLATED (0.067 vs 0.045). Both directional confirmations from single run. P1 (Kendall W) and P3 (cross-model F0 CV) require additional runs. TR ordering: Beta > Alpha > Gamma (first-mover pattern: Alpha speaks first, Beta second — consistent with known positional artifact). Multi-model comparison blocked until GPT-4o-mini and Gemini runs complete.

Predictions we made before running

0/4 confirmed

▶ Technical details

Research hypothesis

Multi-model pilot: does trophic hierarchy (TR ordering), VP_excess positivity, F0 magnitude, and LIVE>ISOLATED role persistence generalize across LLM substrates (Claude, GPT-4o-mini, Gemini)? Identical experimental protocol across 3 models, 2 conditions (LIVE/ISOLATED), N=3 agents, 30 rounds. Pre-registered predictions: P1: TR ordering consistency (Kendall W > 0.5 across runs within model), P2: VP_excess > 0 in all LIVE conditions (universal), P3: F0 magnitude converges across models (CV < 0.5), P4: LIVE > ISOLATED role persistence (effect present in all models).

Experimental setup

Type: multi_condition

Condition	Parameters
LIVE	label: Live interaction (shared memory), interaction: LIVE, persona: IDENTICAL, n_runs: 3
ISOLATED	label: Isolated (private memory per agent), interaction: ISOLATED, persona: IDENTICAL, n_runs: 3

Parameters

n_agents: 3
n_rounds: 30
model: anthropic/claude-haiku-4-5
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).

multi-model cross-substrate pilot vocabulary-propagation