← back to timeline

EXP-STAR-DECAY

artifact

Star Topology × Memory Decay — 2×2 Missing Cell

2026-03-14 L3 level paper3 1 runs $0.00
new
premortem
mock
real
metrics
analyze
review
krit
done

In plain language

Three AI agents communicate in a text space. One is the "hub": sees both others. Two are "spokes": see only the hub. Previously we learned: the hub becomes a word "absorber" (not a generator). What if we add forgetting? The hub absorbs — but forgets. Spokes generate — and remember. Who is the source now?

What we found

UNDERPOWERED (N=1): No predictions scored

Predictions we made before running

0/5 confirmed

Technical details

Research hypothesis

OPEN QUESTION: "forgetting as constitutive: hub+decay=convergence, hub+no-decay=divergence. mechanism?"

2×2 FACTORIAL — THREE CELLS FILLED, ONE MISSING: | | No Decay | Partial Decay | |--------------|---------------|-------------------| | Complete | TR=0.444 (N=9)| TR=0.261 (N=1) | | Star | TR=0.141 (N=1)| ??? ← THIS EXPT |

KNOWN MECHANISMS:

  • Star/no-decay: hub=SINK (Alpha TR=0.409, lowest). High in-degree → absorbs vocabulary → imports > exports. VP=0.243 (highest convergence).
  • Complete/partial-decay: intermediate hierarchy (TR=0.261). Decay weakens path-dependent positional advantage. VP=0.245.
  • Complete/no-decay: strong hierarchy (TR=0.444). Full memory → accumulated advantage compounds. VP=0.075.

COMPETING HYPOTHESES FOR THE MISSING CELL: (A) ADDITIVE: Both effects compound. Star suppresses hierarchy (0.444→0.141), decay suppresses hierarchy (0.444→0.261). Combined: TR ≈ 0.05-0.08 (near flat). Hub still SINK but weaker. (B) INTERACTION: Decay REVERSES hub-as-sink. Hub absorbs → but forgets → can't accumulate. Spokes generate independently AND remember → become sinks instead. Hub TR INCREASES (toward source). Hierarchy may increase vs star-alone. (C) FLOOR: Both suppressions hit the same mechanism (accumulated advantage). Once topology already suppressed it (TR=0.141), decay has NO additional effect. TR ≈ 0.141 ± noise.

DISCRIMINANT: The interaction term (does decay change WHO is source/sink?) is more informative than the main effect (does TR go up or down?).

BASELINE DATA: Complete/no-decay (exp-asym PERSONA_LIVE, N=9): TR=0.444±0.15, VP=0.075 Complete/partial-decay (exp-forget-partial, N=1): TR=0.261, VP=0.245 Star/no-decay (exp-topo-star, N=1): TR=0.141, VP=0.243, Alpha TR=0.409 (SINK) Cycle/no-decay (exp-topo-cycle, N=1): TR=0.071, VP=0.167 Complete/full-decay (exp-forget, N=1): TR=0.015, VP=high

Experimental setup

Type: factorial_2x2_missing_cell

Condition Parameters
STAR_DECAY topology: STAR, memory: PARTIAL_DECAY, interaction: LIVE, note: Hub (Alpha) reads Beta+Gamma + partial decay. Spokes read only Alpha + partial decay.

Factors: topology (STAR) × memory (PARTIAL_DECAY)

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).