N-MATCHED CONTROL: Prompt-based forgetting at N=5.
PROBLEM: exp-forget (N=3, prompt decay): TR=0.015 — hierarchy COLLAPSES exp-memory-truncation (N=5, structural truncation window=20): TR=0.298 — hierarchy SURVIVES
Two confounds: (a) N=3 vs N=5 — different group size, different hierarchy baselines (b) Prompt-based "attend to last 2-3" vs actual context truncation (window=20)
THIS EXPERIMENT: prompt-based forgetting at N=5. Same "attend only to last 2-3" instruction as exp-forget, but with N=5 agents (same as exp-memory-truncation).
If TR collapses (< 0.15): prompt-based decay kills hierarchy regardless of N. → exp-memory-truncation's TR=0.298 is due to STRUCTURAL truncation being less effective than prompt-based instruction (agents still see full text). If TR survives (> 0.25): N=5 is sufficient to maintain hierarchy even with decay. → exp-forget's collapse was an N=3 artifact (too few agents). If intermediate (0.15-0.25): partial effect, needs characterization.
BASELINES:
- exp-pent (N=5, full memory, 60 rounds): TR~0.5
- exp-forget (N=3, prompt decay, 48 rounds): TR=0.015
- exp-memory-truncation (N=5, window=20, 60 rounds): TR=0.298
- exp-memory-truncation-r1 (N=5, window=20): TR=0.298±0.079 (pooled N=3)
This creates a clean 2×2: N=3 full memory (exp-asym): TR=0.444 N=3 prompt decay (exp-forget): TR=0.015 N=5 full memory (exp-pent): TR~0.5 N=5 prompt decay (THIS): TR=?
EXPLORATION TIER (N=1). Goal: fill the 2×2 cell.