Public ledger

Verdict lookup

Every verdict is reviewed by a human editor before publication. Unreviewed claims stay a structural read.

Index

• 2401.00001 Paged KV-cache for 13B-70B serving without quality loss
• 2402.00002 INT4 post-training quantization for self-hosted 70B decode
• 2403.00003 A legal-document reranker for 32k contexts
• 2404.00004 Competing lab announces cheaper decode for 70B
• 2405.00005 Survey of recommender systems in 2025
• 2508.00006 Yesterday's preprint on speculative decoding variants
• 2406.00007 Flash-style attention rewrite for SGLang
• 2107.00009 Online softmax as a streaming reduction for attention
• 2205.00010 IO-aware attention for long sequences on a single GPU