Log 016 — Foundation Models (v0.8)

Date: TBD
Version: v0.8
Pipeline: Classical + foundation model ensemble

This log will be written during the v0.8 implementation. Outline below.

Planned Sections

A. Implementation Record

• Claude Vision API integration (automated scene description)
• Grounded SAM 2 segmentation (text-prompted net mask)
• Depth Anything V3 (monocular depth for occlusion)
• CountFormer / CounTR (few-shot hole counting)
• NEFI graph extraction (Euler characteristic counting)
• Classical pipeline preservation (the existing voter)

B. Mask Comparison

• Classical hole-density mask vs. SAM 2 mask
• Overlap analysis, disagreement regions
• Which mask is more accurate?

C. Count Comparison

• Classical ensemble count
• CountFormer density count
• NEFI topological count
• Meta-ensemble reconciliation

D. Depth-Aware Classification

• Depth map visualization
• Flesh-occlusion via depth vs. brightness
• How many holes are reclassified?

E. Results

• Morgan's v0.8 answer
• Per-method breakdown
• Comparison to v0.7 (696)
• Confidence assessment

F. What Morgan Learned

• Did foundation models improve the count?
• Did they change what Morgan counts?
• What remains unsolvable without new data?

This scaffold will be replaced with full content during implementation.