← back to timeline

EXP-ASYM-DECAY

informative

Asymmetric Decay — Isolating Hub vs Spoke Forgetting Mechanism

2026-03-15 L2 level paper3 2 runs $0.01
new
premortem
mock
real
metrics
analyze
review
krit
done

In plain language

Three AI agents communicate: one "hub" sees the other two, and the two see only the hub. Previously we discovered: if ALL agents forget old entries, the hub shifts from a word "absorber" to a word "generator." But WHY? Because the hub forgets and is forced to create new terms? Or because the others remember and accumulate an advantage? We apply forgetting ONLY to the hub or ONLY to the others — and see which one reverses the polarity.

What we found

UNDERPOWERED (N=1): No predictions scored

Predictions we made before running

0/5 confirmed

Technical details

Research hypothesis

MECHANISM ISOLATION: The polarity reversal in exp-star-decay (hub SINK→SOURCE when all agents forget) has two competing mechanistic explanations:

(A) HUB-FORGETTING HYPOTHESIS: Hub forgetting prevents vocabulary accumulation → forces synthesis from dual input → generates novel terms → becomes SOURCE. Prediction: hub-only decay → reversal occurs.

(B) SPOKE-MEMORY HYPOTHESIS: Spokes remembering creates accumulation advantage → spokes become long-term repositories → become sources regardless of hub. Prediction: spoke-only decay → reversal occurs.

(C) INTERACTION HYPOTHESIS: Both mechanisms are necessary. Hub must forget AND spokes must remember for the polarity to flip. Prediction: neither asymmetric condition alone produces full reversal.

BASELINE DATA: exp-topo-star (no decay): Alpha=0.409 SINK, Beta=0.543, Gamma=0.550, TR_range=0.141 exp-star-decay (symmetric): Alpha=0.572 SOURCE, Beta=0.497, Gamma=0.420, TR_range=0.152

This is a MECHANISM ISOLATION experiment: same topology, same agents, but asymmetric memory manipulation to test WHICH agent's forgetting drives the effect.

COST: 2 conditions × 48 rounds = 96 API calls ≈ $0.01

Experimental setup

Type: two_condition

Condition Parameters
HUB_DECAY persona: PERSONA, interaction: LIVE, decay_target: HUB, note: Only Alpha (hub) gets decay instruction. Beta and Gamma have full memory.
SPOKE_DECAY persona: PERSONA, interaction: LIVE, decay_target: SPOKE, note: Only Beta and Gamma (spokes) get decay instruction. Alpha has full memory.

Factors: persona (PERSONA) × interaction (LIVE) × decay_target (HUB, SPOKE)

Parameters

n_agents
3
n_rounds
48
n_runs_per_condition
1
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).