EXP-CA-RULE-MULTIGEN-AXIS-CONTROL
… pendingStep-3 axis-redirection confound resolver — LLM-aligned per-agent preferences with CA available
In plain language
This experiment tested whether a previous finding of erased "attractors" (stable behaviors) was due to how agents were guided to choose their actions. Researchers changed the guidance to align with what a language model preferred for each agent, while still allowing the agents to interact with their environment. The results are still pending to see if the original stable behaviors return.
▶ Technical details
Research hypothesis
Step-3 evidence stack (N=16) showed apparent step-2 attractor erasure: coupling_strength=strong attractor went 10/10 → 0/16, voting=majority went 2/10 → 14/14. Mnsky-research audit (2026-05-04) found this is partially axis-redirection artifact: per-agent axis preference (axis_diversify) was extended for step-3 with CA-specific preferred axes (Alpha→patch_size, Charlie→mark_allowed, Delta→cross_patch_view), so 3/4 step-2 LLM-axes received zero proposals in step-3. Only Bravo→signal_type pairing was shared (62%→18% acceptance, robust). This control: same hybrid CA+LLM substrate, but LLM-aligned per-agent axis preferences (Alpha→broadcast_to, Bravo→signal_type, Charlie→coupling_strength, Delta→voting_threshold). CA axes still in rule space, agents still observe CA grid. Tests whether step-2 attractors return when agents are nudged toward LLM axes despite CA presence. Pre-registered: - SUBSTRATE-COMPETITION hypothesis: step-2 attractors stay erased (≤30% return). coupling=strong rate ≤30% even when Charlie nudged to coupling axis. Substrate truly competes for proposal bandwidth.
- AXIS-REDIRECTION hypothesis: step-2 attractors return (≥50%). When Charlie proposes coupling and Delta proposes voting, the step-2 universal canalization re-emerges. Step-3 finding was methodological artifact of axis-prefs.
- PARTIAL: 30-50% return — both effects present.
Discriminating: shared (agent, axis) pairings — Bravo→signal_type signal_type acceptance rate. If matches step-3 (~18%), CA-effect on rejections is real. If matches step-2 (~62%), even shared pairing was confounded.
Experimental setup
Type: simple
| Condition | Parameters |
|---|---|
| AXIS_LLM_ALIGNED | coupling: ALL, axis_per_agent: [object Object], n_runs: 6 |
Factors: coupling (all)
Parameters
- n_agents
- 4
- n_rounds
- 100
- model
- google/gemini-2.5-flash-lite
- temperature
- 0.85