I built this to stop losing signal in the arXiv firehose. It's a single Bun script — ~450 lines — that fetches papers, scores them for LLM and agent relevance using a weighted signal table, groups by topic tag, and writes timestamped Markdown digests you actually keep.

No framework, no setup, no cloud, no API keys. I run it as part of my own daily research loop. This is the exact file. Yours to read, run, and modify.

Requires: Bun >= 1.0