← back to timeline

EXP-COMMUNITY-MERGE

informative

Community Merge — First L5 (Community-Level) Experiment

2026-03-25 L5 level paper3 4 runs
new
premortem
mock
real
metrics
analyze
review
krit
done

In plain language

When two independently-formed agent groups merge, hierarchy amplifies beyond either pre-formed group. Agents prefer in-group vocabulary (64%) but form a single interleaved hierarchy. Abstract vocabulary propagates 3.5x more than concrete vocabulary across groups. First L5 (community-level) evidence.

What we found

UNDERPOWERED (N=1): No predictions scored

Predictions we made before running

0/4 confirmed

Technical details

Research hypothesis

COMMUNITY-LEVEL IDENTITY PERSISTENCE HYPOTHESIS

L3 shows stable hierarchies in N=3-5 groups (TR_range ~0.3-0.5). L4 shows self-repair — groups maintain structure after perturbation.

KEY QUESTION: Can two stable groups interact without collapsing into a single group? This is the FIRST L5 experiment. Level 5 = community: interacting organisms, meta-organization (Szathmáry & Maynard Smith 1995).

PROTOCOL (three phases):

  • Phase A: Group_1 (Alpha, Beta, Gamma) runs 10 rounds → establish hierarchy
  • Phase B: Group_2 (Delta, Epsilon, Zeta) runs 10 rounds → establish hierarchy (Phase A and B are INDEPENDENT — separate shared memories, separate seeds)
  • Phase C: Merge all 6 agents into single shared memory for 10 rounds (seed = concatenation of final states from both groups)

CRITICAL INNOVATION: pre-formation. Standard multi_agent starts from scratch. This experiment starts two groups SEPARATELY, then merges them. The merge creates a situation where agents have "cultural memory" from their origin group but must now interact with outsiders.

WHY N=3 per group (not N=4 per ROADMAP): N=3 is the minimal group that forms hierarchy (exp-asym: TR 0.444). N=3+3=6 total, which is within the scaling curve (exp-n6: TR ~0.38). Using N=3 keeps the pilot small and comparable to existing baselines.

WHAT WE MEASURE:

  • Do agents preferentially interact with (cite/respond to) original group members?
  • Does the merged group form a NEW hierarchy or maintain two parallel ones?
  • Is merged TR_range > max(group_A_TR, group_B_TR)? (more complex = L5 property)
  • Does vocabulary from group A propagate into group B and vice versa?

THEORETICAL GROUNDING:

  • Szathmáry & Maynard Smith (1995): major evolutionary transitions involve formerly independent replicators forming higher-level units
  • Mathis (2024): composition barrier — combining functional units may NOT produce higher-level functional units without new mechanisms
  • Campedelli: social emergence requires interaction between pre-formed groups

THIS IS EXPLORATION TIER (N=1). Goal: does the phenomenon exist at all?

Experimental setup

Type: three_phase_merge

Condition Parameters
COMMUNITY_MERGE persona: PERSONA, interaction: LIVE, merge: full

Factors: n_agents_per_group (3) × n_rounds_phase_a (10) × n_rounds_phase_b (10) × n_rounds_phase_c (10) × merge_strategy (concatenate_memories)

Parameters

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 (4 experiments)