KEYSTONE-ABLATION

✓ success

L1.5 Keystone Ablation — Does Trophic Hierarchy Survive Agent Removal?

2026-03-09 L3 level paper3 45 runs

new

premortem

mock

real

metrics

analyze

review

krit

done

In plain language

Three agents form a hierarchy: one sets the vocabulary, the others adopt it. We removed the leader — and expected hierarchy to collapse: after all, two agents without history (EXP-DYAD) are always flat. But no: the remaining two preserved hierarchy (TR diff 0.49 instead of 0.06), because shared memory carries an "organizational imprint." This is a signal of autopoiesis: the system structure is encoded in the text (substrate), not only in the number of agents.

What we found

Socialized dyad maintains hierarchy (TR diff 0.489 vs fresh-dyad 0.060, t(14)=6.68, p<0.0001, d=1.72). Memory overrides topology.

Predictions we made before running

✓ Socialized N=2 TR diff 0.489 >> fresh N=2 baseline 0.060
✗ No disruption gradient — hierarchy reorganizes rather than collapses
✓ Control stable across 60 rounds
✗ Role inheritance marginal — trend in predicted direction but underpowered

2/4 confirmed · 2 refuted

▶ Technical details

Research hypothesis

If the multi-agent trophic hierarchy is autopoietic (actively self- maintaining), then removing the keystone agent (highest trophic ratio) will NOT permanently destroy the hierarchy: the remaining agents will reorganize and maintain trophic differentiation, even as a dyad (N=2).

If the hierarchy is merely a passive byproduct of persona directives + N=3 topology, then keystone removal collapses it permanently — the remaining N=2 system reverts to the undifferentiated state observed in EXP-DYAD (TR diff 0.060).

KEY DISTINCTION: This tests whether organizational structure is encoded in the SHARED MEMORY (autopoietic — memory maintains role pressure even after agent removal) or only in the AGENT DIRECTIVES + TOPOLOGY (not autopoietic — structure disappears with the topology that created it).

PRIOR EVIDENCE: EXP-ASYM: behavioral directives create stable trophic roles (d=1.2) EXP-SLC: three-level gradient confirmed (p=0.002), interpreter-dependent EXP-DYAD: fresh N=2 → NO hierarchy (TR diff 0.060) EXP-QUAD: N=4 COLLAPSES hierarchy (TR range 0.059) → N=3 is Goldilocks topology. Removing 1 agent → N=2. → The question: can "socialized" N=2 maintain what fresh N=2 cannot?

AUTOPOIESIS FRAMING (adapted from EXP-11 pond-level design): Varela (1974): does the system "continuously regenerate the network of processes that produced" its components? Rosen (1991): is the "repair function" internal to the system? Beer (2004): which perturbations are recoverable? Our test: remove keystone → does hierarchy self-repair?

Experimental setup

Type: ablation

Condition	Parameters
K1	ablation: keystone, persona: PERSONA, interaction: LIVE
R	ablation: random, persona: PERSONA, interaction: LIVE
C	ablation: none, persona: PERSONA, interaction: LIVE

Factors: ablation (keystone, random, none)

Parameters

n_agents: 3
n_rounds: 60
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).