docs: Jina v5 paper analysis + vision sensor plan

[AdaWorldAPI]

Summary

• Jina v5 paper analysis (arXiv:2602.15547) — GOR, CoSENT, LoRA adapters, Matryoshka
• Vision sensor plan — ViT-Huge-14 for medical imaging + CLIP cross-modal

Jina v5 Paper Findings

• GOR regularizer: makes embeddings robust to binary quantization → our i8 tables lose LESS info on v5 than v3
• CoSENT loss: directly optimizes Spearman ρ (= our calibration metric). 5 lines in candle.
• 4 LoRA adapters (retrieval/STS/cluster/classify) = our ThinkingPresets (Analytical/Balanced/Creative/Focused)
• Matryoshka: 256D minimum for CLAM codebook building (below = quality drops fast)
• Listwise Reranker v3: reads ALL candidates together, not pairwise — our reranker_relevance() needs rework
• Cronbach α as meta-debugger: where Reranker and Engine disagree on ranking = where our encoding is wrong

Vision Sensor Plan

• FP32 ground truth: Kijai/WanVideo_comfy ViT-Huge-14 (2.53 GB safetensors)
• BF16 production: DeepBeepMeep/Wan2.1 combined CLIP (2.39 GB)
• Medical pipeline: DICOM → ViT patches → codebook → SPO → NARS
• Cross-modal: text ↔ image in same CLIP embedding space
• Three tools: candle (text+vision), ort (reranker), rten (medical ViT)

Calibration Predictions

Jina v5 (GOR-trained): i8 Spearman ρ > 0.95 (quantization-robust)
Jina v3 (no GOR):      i8 Spearman ρ < 0.90 (not trained for quantization)
CoSENT fine-tune in candle if ρ < 0.998 (directly optimizes rank order)

https://claude.ai/code/session_019RzHP8tpJu55ESTxhfUy1A