KEYORDER-AUDIT-RULE-PROPOSAL
✓ successkeyorder-audit-rule-proposal
In plain language
Researchers tested if a language model would propose rules in a specific order when it wasn't given hints about which order was best. They found that the model did indeed propose rules in the order they were listed in its instructions, confirming their prediction.
▶ Technical details
Research hypothesis
Candidate rule rule-space-key-order-as-second-engine-pathway-audit (1st anchor: L7 exp-axis-diversify-false-ablation-arm, ρ=0.800 on rule_multigen engine) predicts that under axis_diversify=False, LLM proposals follow schema-listing-order of RULE_SPACE keys instead of uniform-mix or substrate-emergent bias.
This is the 2nd-anchor test on the BASE rule_proposal engine (different engine class, single-gen lifecycle vs multi-gen, different prompt construction). Same RULE_SPACE = {broadcast_to, signal_type, coupling_strength, voting_threshold} listed in that fixed dict-insertion order.
Pre-registered:
- ANCHOR-CONFIRMED: Spearman ρ ≥ 0.6 between LLM axis-proposal distribution (across all 8 runs aggregated) and the schema-listing rank (broadcast_to=1, signal_type=2, coupling_strength=3, voting_threshold=4) → candidate rule promoted to standing.
- ANCHOR-REFUTED: ρ < 0.6 → rule stays candidate, document why (engine-specific to rule_multigen? sample size? different schema-order influence?).
must_propose=True is set to ensure ≥40 propose events for statistics; axis_diversify=False is the critical factor (no axis_hint injection, LLM sees raw rule_space_str).
Experimental setup
Type: simple
| Condition | Parameters |
|---|---|
| KEYORDER_AUDIT | coupling: true, must_propose: true, axis_diversify: false, rule_stickiness: false, n_runs: 8 |
Factors: coupling (true)
Parameters
- n_agents
- 4
- n_rounds
- 40
- model
- google/gemini-2.5-flash-lite
- temperature
- 0.85