NanoJudge ranks huge lists with tiny LLMs

This is a smart decomposition play: instead of asking one model to do an impossible global ranking, NanoJudge turns ranking into a statistically grounded tournament that small local models can actually handle well.

• –The real innovation is workflow design, not raw model quality: pairwise judging avoids context-window collapse and “lost in the middle” failure modes
• –The Rust core and CLI make it feel more like infrastructure than a demo, especially with support for OpenAI-compatible local endpoints
• –Confidence intervals and positional-bias correction give it more rigor than most LLM ranking hacks, which usually stop at anecdotal outputs
• –The top-heavy matchmaking strategy matters because naive exhaustive comparisons explode quadratically on large lists
• –Best fit is research triage, retrieval reranking, and large-option decision support rather than general-purpose reasoning