Block 6: Vision, Curation & AI-Assisted Workflow Lessons

19 lessons from biased voting blocks, vision tagging, CLIP classification, image deduplication, interactive selection tooling, and AI-assisted development patterns.

#	Lesson	Core Teaching
039	Mock Tagger for Vision Pipeline Testing	Hash-based mock tagger provides deterministic, self-scaling attribute generation for CI without GPU
040	Controlled Vocabulary as Schema Contract	A single YAML vocabulary enforces attribute consistency across VLM prompt, parser, DB, config, and analysis
041	Utility Function for Synthetic Voting Bias	Additive utility (base appeal + preference x match + randomness x noise) gives orthogonal control over bias strength
042	Lift as the Primary Bias Detection Metric	Lift (block rate / global rate) is scale-invariant and base-rate-invariant for bias detection
043	PII Sanitization in Static Exports	Defense in depth — query exclusion, recursive field stripping, and post-export content scanning
044	Acceptance Tests as Executable Specifications	Assert structural properties for statistical outputs and exact values for deterministic outputs
045	Embedding Deduplication for Image Curation	CLIP cosine similarity + connected components reduced 12,217 images to 2,163 unique representatives
046	Lazy Imports for Deployment Compatibility	Move heavy dependencies (numpy, torch) to function scope so lightweight deployments don't break
047	CLIP Zero-Shot as a Database Column Factory	One CLIP model + descriptive prompts = arbitrarily many structured database columns, no training needed
048	Greedy Max-Min Diversity Selection	Iteratively pick the most distant item from selected set — O(n*k), near-optimal, 15 lines of code
049	Drag-and-Drop as Simplest Viable Interaction	Co-visible source pool + target slots with HTML5 DnD beats modals, search, and multi-step wizards
050	Connected Components for Transitive Dedup	scipy sparse graph components correctly group transitive chains that pair-based merging splits
051	Sigmoid Calibration for Domain-Specific CLIP	CLIP logits are domain-specific (16-32 for space photos); sigmoid center and scale must be empirically tuned
052	Incremental Feature Extraction	Delete-and-rewrite only new attribute codes; label_source and attribute_code columns enable surgical updates
053	Audit-First Design	A 15-minute written audit of existing code prevents hours of reimplementation and identifies extension points
054	Phased Autonomous Execution	Numbered phases with commit checkpoints turn multi-session projects into a queue of self-contained tasks
055	Session Continuity via Artifacts	CLAUDE.md + startup docs + plan files eliminate ramp-up overhead across session boundaries
056	Environment Self-Interference	AI assistants fight their own orphan processes — explicit cleanup before each launch is mandatory
057	Test-Gated Commits at Scale	Never commit red tests; run incrementally after each change, not in batch at the end