amico papers digest — intelligent daily arXiv picks to Slack, ranked against the corpus

[aarontrowbridge]

Problem — The #papers channel's daily arXiv digest comes from an unidentified personal automation posting as a user token: generic picks ('3 of 39 new'), no relation to the lab's actual research profile, no dedup against what the lab has already read. The corpus fold (#405) just made the lab's demonstrated interests machine-readable (91 notes, tags/systems/relevance); nothing consumes it.

Approach — amico papers digest: pull the arXiv RSS (quant-ph default, configurable feeds), rank each new item against a profile DERIVED from the corpus fold (tag/system term frequencies, weighted by recency) plus a small overrides file, skip identities already in the corpus or previously posted (state file), and post the top N to Slack as the Amico bot — each pick with its match explanation. Deterministic and auditable: the intelligence is the corpus-derived profile, not a model (an LLM summarizer layer can come later, Prova-scored).

Scope — in: RSS-subset parser (zero-dep), scorer, state, formatter, --post, launchd schedule on the server · out: LLM summaries, per-study subscriptions (slice 3 of the literature spec), reactions feedback.

Acceptance Criteria

• digest ranks new arXiv items against the corpus-derived profile; every pick carries an explanation (matched terms)
• identities already in the corpus or the posted-state are skipped — reruns never repost
• --dry-run prints; --post sends via the server-side Slack token; the bot is in #papers
• a launchd agent on the mini runs it daily; one real digest lands in #papers today
• deterministic: same inputs → same digest (test-pinned)

Source

Literature-plane spec spec-20260817-140000 · builds on #405 (stacked branch, retarget after #407).