fix: PDF notes wrap inside table; hide trend chart until 2+ inspections (no empty box)

fix: comparison report sentence built from specialist counts (no more mite contradiction)

ui: fill_width=True so layout uses full width (capped at 1600px via css)

ui: queen tile shows count (1/0) not YES; widen layout to 1600px

off-brand: custom tab bar, themed inputs/labels/tables, custom scrollbar, strip Gradio tells

feat: inspection dashboard (verdict + stat tiles), colony trend charts, polished A/B kill-shot cards

polish: hero header + detection legend + footer (Off-Brand); document Modal training pipeline + published artifacts

remove cyan queen candidate boxes - only confident green QUEEN box now

feat: ranked top-3 queen candidates with probability shown - honest AI, user verifies

feat: 5 sample photos with verified queen detection + disclaimer about operating conditions

feat: VLM cascade for queen verification (YOLO finds candidates, Qwen-VL confirms via grid comparison)

feat: lower queen threshold to 0.05, show top-confidence per class in narrative

fix: distinguish empty YOLO detections from missing weights in narrative

chore: nudge build hook

polish: remove emojis and em dashes from sources and docs

polish: seed sample hives, polished CSS theme, rich About + README, dataset live on Hub

fix: per-class confidence thresholds, queen-only label, mites clearly marked as VLM estimate

refactor: YOLO runs on main container (not GPU worker), Qwen-only in @gpu

fix: run YOLO on main container, not GPU worker (file access + log visibility)

feat: PDF weekly report generator + download button in Hives tab

feat: real A/B comparison logic (Apiarist vs Qwen-alone)

feat: YOLO + Qwen pipeline with graceful fallback when weights absent

feat: A/B Compare tab shell, ready for YOLO weights

fix: refresh Inspect tab hive dropdown when hives are added/refreshed/analyzed

feat: SQLite hive registry + inspection history tab, persistent across sessions

swap to Qwen2.5-VL-3B (sweet spot: better vision than SmolVLM, fits ZeroGPU)

revert: Qwen2.5-VL-7B crashes ZeroGPU container, use SmolVLM-Instruct 2.25B (known good)

upgrade SmolVLM-500M -> Qwen2.5-VL-7B for real vision quality on ZeroGPU

wire @spaces.GPU decorator for ZeroGPU; auto cuda/cpu device handling

swap moondream2 (broken custom code) -> SmolVLM-500M (standard transformers API)

wire moondream2 (placeholder VLM) + add iNaturalist scraper + data sources doc

scaffold: dummy analyze pipeline + custom-themed Gradio UI