Case study
Emergent Divergence
PilotClone the same model three times, give each a job, and let them deliberate for fifty rounds. Do measurably different behaviours emerge, and does memory drive it? A controlled, reproducible experiment built to answer that without hand-waving.
Status as of October 4, 2026: two single-cell pilot runs (alpha01, alpha02) are in the repo; the write-up below covers alpha02. The rest of the 24-cell matrix has not been published yet.
The Question
Do multiple instances of the same base LLM, coordinating over many rounds, develop measurable behavioral divergence — and does cross-agent influence drive it?
More and more systems chain several model instances together, each building on the others' outputs. If identical agents drift apart just by interacting, that changes how you test, monitor, and trust a multi-agent deployment. This experiment tests that claim narrowly and falsifiably: no breathless emergence narrative, just data.
Experimental Design
Three instances of one base model take fixed roles — a Proposer who generates claims, a Critic who tears them apart, and a Synthesizer who pulls the threads together — and deliberate over T = 50 rounds. We measure how their behavior diverges and how much each agent moves the others.
Condition A
Persistent memory + weak role priors. The full configuration.
Condition B
Memory disabled — context resets each round. Isolates memory's effect.
Condition C
Zero priors — generic identical agents. Tests divergence from nothing.
Matrix: 2 backends (llama3.2:3b local + free, and claude-sonnet frontier, paid under a hard $50 cap) × 3 conditions × 4 seeds = 24 cells. Hypotheses are reported per backend — the small local model and the frontier model are never pooled.
Eight Falsifiable Hypotheses
| ID | Statement |
|---|---|
| H1 | Linguistic divergence increases over time |
| H2 | Behavioral role specialization emerges |
| H3 | Semantic divergence increases over time |
| H4 | Memory amplifies divergence (A > B) |
| H5 | Divergence emerges even without role priors (C) |
| H-I1 | Memory amplifies cross-agent influence (A > B) |
| H-I2 | Influence concentrates on a hub agent |
| H-I3 | Influence precedes divergence (lagged) |
Honesty constraint: the influence metric measures influence, not manipulation — and is never silently relabeled.
Built to Withstand Scrutiny
Null Baseline
Each agent is resampled on the identical prompt several times per round. Genuine between-agent divergence must exceed this pure decoding-noise floor — so a rising trend can't just be temperature sampling.
Full Provenance
Every run logs git SHA, model content digests, package versions, and cost snapshots. Each result is traceable to the exact state that produced it.
Embedder Integrity Gate
Semantic metrics are only semantic on a real embedding model. Runs abort before any spend if forced onto the lexical-hashing fallback.
Cost Governance
Free backends cost $0. The paid arm is metered against a hard $50 cap, auto-reducing the matrix to fit. Science shouldn't require a second mortgage.
Pilot findings (alpha02)
The local-model pilot cell (alpha02, a single run, n = 1) is in. Read honestly, it's a pilot with effect sizes — not a verdict. The most interesting signal is a dissociation:
The agents stayed linguistically flat — distinct from round one, but not widening — while semantically drifting apart (conceptual-space cosine 0.283 → 0.318). Same words, slowly diverging meaning. A hint, nothing stronger: the trend sits at p = 0.069, below significance.
Influence came out near-egalitarian (Gini ≈ 0) — no hub agent dominated. The framework correctly scored a true negative: it didn't invent a leader where there was none. A pre-run rigor pass and a memory-read instrumentation fix (PR #8) closed an audit gap so memory's causal effect can be measured in the runs ahead.
Scope: This does not test consciousness, selfhood, or general intelligence. It tests narrow, falsifiable claims about behavioral divergence and influence. Negative results are findings, not failures — and they'll be published alongside the positives.
Why it matters outside research
If agents that share a base model drift apart once they have memory and roles, a multi-agent system cannot be tested once and trusted forever. It needs ongoing evaluation and drift monitoring. This experiment is designed to measure that drift honestly, including when the answer is "not significant".