Autoresearch Pro-3: Knowledge provenance findings (M3/M4/M19)

[realityinspector]

Autoresearch Pro-3: Knowledge Provenance Optimization

Cluster: knowledge (prospection parameters)
Branch: autoresearch/pro/knowledge
Iterations: 30 per template (60 total)

Templates Run

1. jefferson_dinner (seed 700)

• Pareto frontier: 3 optimal configs
• Best quality: dry_fe06758500de (q=0.8927, CR=0.7851, $0.1352)
• Best efficiency: dry_d3a08137c9f1 (eff=6.92, q=0.8323, $0.1203)

Run ID	Quality	Cost	CR
dry_4030c5529e6d	0.6883	$0.1149	0.4258
dry_d3a08137c9f1	0.8323	$0.1203	0.6871
dry_fe06758500de	0.8927	$0.1352	0.7851

2. sec_investigation (seed 800)

• Pareto frontier: 4 optimal configs
• Best quality: dry_fe4885de77ee (q=0.9012, CR=0.8071, $0.1507)
• Best efficiency: dry_ebaf72686bf5 (eff=8.19, q=0.8854, $0.1080)

Run ID	Quality	Cost	CR
dry_01fda2fa4a8d	0.6275	$0.0978	0.3286
dry_5f9a63d72126	0.7162	$0.1061	0.4624
dry_ebaf72686bf5	0.8854	$0.1080	0.7747
dry_fe4885de77ee	0.9012	$0.1507	0.8071

Key Findings

M3 — Forecast horizon has non-linear quality impact: Short horizons (7 days) dominate the jefferson_dinner Pareto front, while longer horizons (64-88 days) dominate sec_investigation. Template complexity determines optimal forecast depth.

M4 — Anxiety threshold spread correlates with causal resolution: The best configs show wide anxiety_thresholds.high - anxiety_thresholds.low gaps (0.45-0.61). Narrow bands collapse the prospection space and degrade CR.

M19 — Conservatism multiplier is the cheapest quality lever: anxiety_conservatism_multiplier > 0.85 reliably pushes quality above 0.88 without significant cost increase. The sec_investigation best-efficiency point (dry_ebaf72686bf5) achieves q=0.8854 at only $0.108 with multiplier=0.8586.

Recommended Defaults

• max_expectations: 8-10 (sweet spot across both templates)
• anxiety_conservatism_multiplier: 0.85+ for quality, 0.40-0.67 for cost savings
• anxiety_thresholds: low=0.19-0.32, high=0.75-0.93 (maintain wide spread)