← back to timeline

EXP-HOMEOSTASIS-R1

negative

Homeostasis Replication R1 (L4)

2026-03-26 L4 level paper3 2 runs $0.01
new
premortem
mock
real
metrics
analyze
review
krit
done

In plain language

Researchers studied whether a group's social structure, like a hierarchy and how tasks are divided, would stay the same even if the individual agents within the group suddenly had their roles and instructions changed. They found that the group's organizational pattern did not resist the change; instead, the structure immediately reorganized to match the new instructions given to the agents. This suggests the group acted more like a collection of individuals following orders rather than a self-maintaining system.

Predictions we made before running

2/5 confirmed · 3 refuted

Technical details

Research hypothesis

HOMEOSTASIS HYPOTHESIS (L4 — organism-level self-maintenance)

L3 experiments show stable hierarchy in N=5 groups (TR_range ~0.5, exp-pent). exp-persona-swap (L2) showed position > persona for individual role assignment. exp-self-repair (L4) showed seamless hierarchy recovery after keystone removal.

KEY QUESTION: if we SWAP agent personas mid-conversation while preserving shared memory, does the organizational pattern (hierarchy structure, role distribution) resist the perturbation?

If the group is an "organism" (L4 homeostasis), organizational patterns should be encoded in shared memory and resist persona changes — the hierarchy pattern should partially persist despite agents receiving new cognitive instructions.

If the group is merely "tissue" (L3), hierarchy follows persona instructions directly — swapping personas should immediately reorganize hierarchy to match new persona assignments.

This extends the self-repair finding: self-repair = recovery after LOSS; homeostasis = resistance to CHANGE. Both are organism-level phenomena.

PROTOCOL:

  • Phase 1: N=5 agents interact for 20 rounds (establish hierarchy + roles)
  • Perturbation: rotate all personas by 2 positions (A gets C's persona, etc.)
  • Phase 2: N=5 agents continue for 20 more rounds with swapped personas
  • Control: no perturbation (natural drift baseline)

COMPARISON:

  • PERTURBED vs CONTROL: does TR_range drop after perturbation?
  • Post-swap role persistence: does agent with highest pre-swap TR maintain high TR despite receiving a different persona?
  • If pre-swap hierarchy correlates with post-swap hierarchy (r > 0.5) → homeostasis signal. If r ≈ 0 → no homeostasis.

REPLICATION R1 of pilot (ρ=0.80). Same protocol, seed set S2, rotation 1. Goal: characterization (N=3-5).

Experimental setup

Type: two_phase_perturbation

Condition Parameters
CONTROL interaction: LIVE, persona: PERSONA, perturbation: none
PERTURBED interaction: LIVE, persona: PERSONA, perturbation: rotate

Factors: perturbation (rotate, none)

Parameters

model
gemini-2.5-flash-lite
n_agents
5
n_runs_per_condition
1
scheduler
round_robin
temperature
0.9

Trophic Ratios by Condition

Mean trophic ratio per agent across runs. Error bars = ±1 std dev. Higher TR = more upstream (exporter).