L3/L4 BOUNDARY PROBE: Does hierarchy survive memory truncation?
CONTEXT:
- exp-forget (N=3, full decay via prompt): TR=0.015 — hierarchy COLLAPSES
- exp-forget-partial (N=3, partial decay): TR=0.261 — intermediate
- exp-asym PERSONA_LIVE (N=3, no decay): TR=0.444 — strong hierarchy
- exp-pent (N=5, no decay, 60 rounds): TR~0.5 — strong hierarchy
BUT: exp-forget used prompt-based "attend only to last 2-3 entries." This is a COGNITIVE constraint, not a STRUCTURAL one. The shared text still contained all history — agents were TOLD to ignore it.
NEW APPROACH: actual memory truncation. Agents see ONLY the last 20 entries of shared text (sliding window). Earlier text is deleted from their context. N=5 (stronger hierarchy baseline).
QUESTION: if hierarchy at N=5 requires cumulative positional advantage (L3 tissue), truncation should destroy it. If hierarchy is encoded in CURRENT interaction patterns (L4 organism-level property), it should survive truncation — the group "remembers" its structure through ongoing dynamics, not through text history.
KEY COMPARISON:
- exp-pent baseline (N=5, full memory, 60 rounds): TR~0.5
- THIS experiment (N=5, window=20, 60 rounds): TR=? If TR > 0.35 → hierarchy survives truncation (L4 signal) If TR < 0.15 → hierarchy requires full memory (L3 only) If 0.15 < TR < 0.35 → partial, needs characterization
WHY THIS MATTERS: Biological organisms maintain identity despite molecular turnover. If agent groups maintain hierarchy despite "molecular turnover" (memory truncation), this is a genuine organism-level property — the organization is in the DYNAMICS, not the SUBSTRATE.
EXPLORATION TIER (N=1). Goal: does the phenomenon exist?