Autoresearch Pro-1: Fidelity optimization findings (M1/M2/M5/M6)

[realityinspector]

Fidelity Cluster Dry-Run Results

Ran 30-iteration dry-run sweeps over the fidelity cluster (M6 tensor compression: compression_method, n_components) on two templates.

board_meeting — Pareto frontier (3 configs)

Run ID	Quality	Cost	CR	Config
dry_9fb185ee2d04	0.8111	$0.1082	0.6104	svd, n_components=5
dry_a318e5a5ae8a	0.8122	$0.1100	0.6330	pca, n_components=5
dry_f52c6a0cc910	0.8862	$0.1266	0.7755	nmf, n_components=2

• Best quality: nmf with n_components=2 (q=0.8862, CR=0.7755)
• Best efficiency: svd with n_components=5 (eff=7.4979)

mars_mission_portal — Pareto frontier (7 configs)

Run ID	Quality	Cost	CR	Config
dry_3831b0561afd	0.6864	$0.1180	0.4240	pca, n_components=5
dry_7b83fe9ef3a7	0.7590	$0.1368	0.5619	pca, n_components=3
dry_c33cce171400	0.8272	$0.1379	0.6608	nmf, n_components=5
dry_e2aef531ddd3	0.8564	$0.1522	0.7465	nmf, n_components=2
dry_ddde2f812ae0	0.8590	$0.1542	0.7301	pca, n_components=9
dry_e61f0290b751	0.8742	$0.1585	0.7625	nmf, n_components=6
dry_f085ecc74357	0.8874	$0.1687	0.7727	nmf, n_components=10

• Best quality: nmf with n_components=10 (q=0.8874, CR=0.7727)
• Best efficiency: nmf with n_components=5 (eff=5.9961)

Cross-template observations

1. NMF consistently dominates both PCA and SVD for quality composite and causal resolution across both templates.
2. Low n_components (2) with NMF yields highest quality on board_meeting; mars_mission_portal benefits from higher components (6-10), likely due to greater scenario complexity.
3. SVD at n_components=5 is the most cost-efficient config for board_meeting (7.50 quality/$), while NMF at n_components=5 leads efficiency for mars_mission_portal (5.99 quality/$).
4. The fidelity cluster search space is only 2 dimensions — these results are high-confidence within the dry-run synthetic metric model.

Recommended next steps

• Run live (non-dry-run) validation of top NMF configs on both templates
• Expand fidelity cluster params (add compression ratio, reconstruction error threshold)
• Cross-cluster sweep: combine fidelity + temporal for joint optimization