Literature tracking

Three literature firehoses, one deduplicated stream, ranked against what the user actually works on.

scripts/track.py does the deterministic half — query, normalise, deduplicate. The judgement half is not its job: deciding what is worth the user's attention, and saying why. Do not try to script that. A model reading titles and abstracts beats a keyword filter at it, and can explain its reasoning afterwards.

Workflow

1. Establish the research profile

Ranking is impossible without topics. Take them from the conversation, from project memory, or ask. A usable profile has: