Method

Machine-written, human-correctable, fully auditable.

Most of what is published on this network is written by a local model. Here is exactly how, what stops it from lying, and what we do when something gets through anyway.

The pipeline

Eight phases, one Python script.

  1. Discover. Pull arXiv RSS for cs.AI, cs.CL, cs.IR, cs.LG. Pull additional configured RSS feeds. Cache the raw XML.
  2. Dedup. Drop anything we already have a paper file or URL for.
  3. Select. Score against a tracked-terms list. Apply a per-run cap.
  4. Enrich. Hit the arXiv Atom API for missing fields, with backoff.
  5. Local AI. One JSON-mode call per candidate to a local Lemonade-served model. Strict schema. Single retry on parse failure.
  6. Quality gate. The model output goes through a verifier before anything ships. See below.
  7. Write artifacts. Always write a tracker entry on accept. Write a GraphSlop critique only if the gate passes and the per-day cap is not exceeded.
  8. Commit, deploy, post. One git commit per run. deploy_all.sh redeploys whichever sites changed. One Discord summary per run, including any retractions.
The quality gate

What stops a critique from shipping.

A critique is rejected (and the tracker entry still publishes without it) if any of these fail:

  • Schema invalid. Required fields missing or wrong type, after one retry. We do not parse around malformed JSON.
  • Quoted lines unfounded. Every quoted line in the critique must be a literal substring of the cached abstract. The model cannot invent a quote.
  • Failure category not in whitelist. The taxonomy is fixed. New categories require a human PR.
  • Length out of bounds. Too short means the model gave up; too long means it is freestyling.
  • Per-day cap exceeded. A hard ceiling on auto-published critiques per UTC day.
When we are wrong

Retractions.

Every machine-written entry carries an auto badge and a "Report this" link. When a report is upheld, we mark the entry's frontmatter status: retracted, add a one-paragraph retraction_note, and append a row to the public run log. The original entry stays visible with strikethrough and the note above it.

This is the trade we are making in public: speed plus a paper trail.

License and fair use

What we quote, and from where.

We only quote from cached abstracts (arXiv abstracts are quotable with attribution). We do not redistribute paper bodies. Every entry shows its arXiv ID and a direct link back. If you are an author and want something taken down or corrected, the same "Report this" link works.