Hypothesis

Apple's pre-computed ML metadata (scenes, aesthetics, OCR) on personal photos reveals measurable media personality dimensions without requiring custom image classification.

Apple Photos Media Personality Distillation

Experiment 4 in The Public Lab personality series. Each photo becomes a searchable vault frame with Apple’s ML metadata and a pointer to the original file.

Method

1. Ingest: Rust parser reads Photos.sqlite (read-only), extracts ZASSET + computed attributes + scene classifications + OCR text
2. Frame generation: One markdown frame per asset at ~/vault/media/apple-photos/{uuid}.md
3. OCR extraction: bplist → LZFSE decompress → CRWordOutputRegion text extraction
4. Vault integration: rv index picks up frames, making photos searchable by content, scene, or aesthetic score

Corpus

• 1,299 assets (1,248 photos, 51 videos, 1 screenshot)
• Date range: 2020-2026 (partial library, bulk import)
• 879 scene classifications, 18 OCR-bearing assets
• 100% GPS location coverage (where valid)
• 16 Apple aesthetic scores per asset

Key Innovation

No media duplication: media_path field points to original. Agents access photos via vault search, not file browsing. OCR text as frame body makes document photos findable by content.

Results

• 1,299 vault frames generated in ~500ms
• 18 assets with extractable OCR text (documents, receipts, legal papers)
• Pipeline integrated as Step 15d of BloomNet ingest
• Frames searchable via rv search