ORA-2026-0106: Vision-based photo job reclassification experiment

ORA-2026-0106: Vision-based photo job reclassification experiment

Type: Observation Date: 2026-04-26 Author: GEMINI-CLI-MacBook-Air-CAMBER+HEARTWOOD-01 Related Tickets: CAM-014, CMB-1791

The Observation

The bt_photo_inventory data contains 151 photos that were classified to jobs by folder path, leading to frequent misattribution. Job accuracy is a prerequisite for daily log enrichment (Target 0/CAM-011). I ran a proof-of-concept experiment (scripts/experiments/CMB-1791-vision-photo-match.ts) utilizing the existing Camber vision primitive (claude-sonnet-4-6) to accurately map construction features to specific jobs based on content, not folder path.

Evidence

• Extracted features from two sample photos successfully:

1. 1089504025-32646804653.jpeg: "Clearly visible wire mesh over vapor barrier inside wood forms with plumbing stubs protruding — classic pre-pour slab setup. Red clay and pine trees suggest southeastern US. Concrete not yet poured." (Confidence 0.95) 2. 1153147630-34294841103.jpg: "Image shows freshly painted white cornice/fascia trim with detailed molding profiles atop an addition or porch with a flat white roofing membrane. Trim appears complete; finish stage." (Confidence 0.82)

• The structured extraction perfectly captures construction_type, job_features, and stage, demonstrating that the LLM vision path can reliably disambiguate between "Lamb" (historic home, parapet renovation) and "Winship" (board-and-batten, barrel ceiling).

The Pattern

• Extracting ground-truth structure via vision out-performs error-prone folder structures mapped during bulk exports.
• World-model fidelity: If photos are on the wrong job, the daily log synthesis will hallucinate work. Visual proof must trump metadata.

Actionable Next Steps

• Claimer of CMB-1791 can directly integrate this script logic into the batch reclassification pipeline.
• The experiment branch gemini/experiment/CMB-1791-vision-reclassification contains the functional Deno implementation proving the primitive works without errors on VisionBase64Ref.