Autoresearch Pro-4: Entity simulation findings (M9-M16)

[realityinspector]

Entity Cluster Autoresearch Results (Pro-4)

Dry-run sweep of the entity mechanism cluster (M9-M16: Animism + Circadian parameters) across two templates, 30 iterations each.

Parameters Explored (7 dimensions)

• animism.level (0-6)
• animism.entity_generation.animal_probability (0.0-0.5)
• animism.entity_generation.building_probability (0.0-0.5)
• animism.entity_generation.abstract_probability (0.0-0.2)
• circadian.energy_multipliers.night_penalty (1.0-3.0)
• circadian.energy_multipliers.fatigue_accumulation (0.1-1.0)
• circadian.energy_multipliers.base_fatigue_threshold (12-20)

---

Template: kami_shrine (seed 500)

• 30 iterations, 10 kept (6 new bests, 4 Pareto)
• Pareto frontier: 5 configs
• Best quality: q=0.8926 (dry_f704270102a1), CR=0.7851, $0.16
• Best efficiency: eff=6.2425 (dry_9fbe98dc9da7), q=0.7956, $0.13

Best config (kami_shrine):

• animism.level=2, animal_prob=0.25, building_prob=0.27, abstract_prob=0.19
• night_penalty=2.22, fatigue_accumulation=0.40, base_fatigue_threshold=17

Template: hospital_crisis (seed 600)

• 30 iterations, 9 kept (4 new bests, 5 Pareto)
• Pareto frontier: 7 configs (wider frontier, more cost-quality tradeoff options)
• Best quality: q=0.8872 (dry_f4200bf5a19b), CR=0.7955, $0.15
• Best efficiency: eff=6.9420 (dry_ae0866375676), q=0.7874, $0.11

Best config (hospital_crisis):

• animism.level=3, animal_prob=0.20, building_prob=0.01, abstract_prob=0.17
• night_penalty=2.19, fatigue_accumulation=0.56, base_fatigue_threshold=19

---

Cross-Template Observations

1. Animism level sweet spot: 2-3. Both templates peaked with moderate animism levels rather than extremes (0 or 6). High animism levels increase entity diversity but may fragment narrative coherence.

2. Abstract entity probability matters. Both best configs use ~0.17-0.19 abstract probability (near the parameter ceiling of 0.2), suggesting abstract entity generation improves quality metrics.

3. Building probability diverges by template. kami_shrine benefits from higher building_prob (0.27) — likely because shrine/temple settings have strong architectural entities. hospital_crisis prefers near-zero (0.01) — the scenario is character/action-driven.

4. Circadian night penalty ~2.2x is optimal. Both templates converge on moderate night penalties. Too low (1.0) reduces temporal structure; too high (3.0) over-constrains nighttime actions.

5. Base fatigue threshold: 17-19h. Higher thresholds allow more action before fatigue degrades performance, improving quality for both templates.

6. hospital_crisis has a wider Pareto frontier (7 vs 5 points), meaning more viable cost-quality tradeoffs — likely because the crisis scenario has more entity interaction paths.

Files

• autoresearch/results/dry_run_entity_20260316_083243.jsonl (kami_shrine, 30 runs)
• autoresearch/results/pareto_entity_20260316_083243.json (kami_shrine frontier)
• autoresearch/results/dry_run_entity_20260316_083248.jsonl (hospital_crisis, 30 runs)
• autoresearch/results/pareto_entity_20260316_083248.json (hospital_crisis frontier)

Next Steps

• Validate top configs with live runs (real OpenRouter calls)
• Cross-cluster interaction: test entity params with optimized temporal/fidelity configs from Pro-2/Pro-3
• Investigate whether animism.level=2-3 holds across all 11 templates