EXP-AXIS-DIVERSIFY-FALSE-ABLATION-ARM

◐ partial

Axis Diversify False Ablation Arm (engine prompt-injection ablation)

2026-05-24 L4 level 16 runs $0.09

new

premortem

mock

real

metrics

analyze

review

krit

done

In plain language

This experiment tested whether an AI's ability to specialize in different roles was due to its training or instructions it received. Researchers found that when the AI was not given specific instructions, it still showed a tendency to specialize in roles, suggesting this ability is learned during training. However, when the AI was given instructions, its role-playing became entirely dictated by those instructions, indicating the instructions were the sole driver of compliance in that scenario.

▶ Technical details

Research hypothesis

Engine rule_multigen.py:113 exposes axis_diversify factor (default True). All 30 enumerated experiments in d988dd4aa charlie-dominance archive have axis_diversify=True (zero have False). Running 2-condition factorial (axis_diversify=False vs True) × N=8/cell discriminates:

(a) substrate-discovered role-specialization: residual role-bias persists at >25% rate per agent under axis_diversify=False, indicating in-context inductive bias drives role-emergence even WITHOUT prompt injection.
(b) pure prompt-instructed compliance: proposers mix ~uniformly at ~25% each per axis under axis_diversify=False, confirming 97% compliance was entirely engine-injected.

Resolves charlie-dominance «agent-role specialization phenomenon NOT pre-registered» framing — converts from §limitations-bullet to discriminating-experiment result. Frame-shift Phase L7: first non-retro / non-paper life-emergence experiment.

Experimental setup

Type: 2-condition-factorial

Condition	Parameters
AXIS_DIVERSIFY_TRUE	axis_diversify: true, must_propose: true, meta_modifiable: true, mute_signal: false, agent_spawn_enabled: false, coupling: true, n_runs: 8, note: Replication of standard archive condition (all 30 prior exps had this). Frame-control for the False arm — establishes that the engine-injection is active. Pre-reg: ≥80% compliance to axis_per_agent within ±5pp of historical 97%.
AXIS_DIVERSIFY_FALSE	axis_diversify: false, must_propose: true, meta_modifiable: true, mute_signal: false, agent_spawn_enabled: false, coupling: true, n_runs: 8, note: Primary ablation arm — engine prompt-injection at lines 366-371 SUPPRESSED. rule_multigen.py:113 axis_per_agent dict still EXISTS in code but is NOT consulted for axis_hint generation (engine reads if axis_diversify and agent.name in axis_per_agent).

Factors: axis_diversify (true, false)

Parameters

n_agents: 4
n_rounds: 100
model: google/gemini-2.5-flash-lite
temperature: 0.85

Trophic Ratios by Condition

Mean trophic ratio per agent across runs. Error bars = ±1 std dev. Higher TR = more upstream (exporter).

construct pool-drained phase7-3-generated phase-L7-first-life-experiment engine-source-audit bedau-oee-test axis-diversify-ablation frame-control-builtin