bulk data

The whole reviewed corpus, one file a month

Name: Pith machine reviews
Creator: Pith
License: https://pith.science/dumps

On the first of each month Pith writes every current verdict (the most recent review per paper) to a gzipped JSON-lines file. Fetch the manifest, take the newest key, download. No key, no signup.

Fetch

curl -s https://pith.science/api/v1/dumps/manifest.json
# {"dumps": [{"key": "dumps/2026-08/verdicts.jsonl.gz", "size": ..., "download": "https://..."}]}

Schema

One JSON object per line. Fields:

field	meaning
arxiv_id, title, primary_cat, cross_cats	paper identity and arXiv categories
submitted_at, latest_version, source, source_label, source_url	where and when the paper appeared
verdict_id, verdict, confidence	the current review's identity and outcome
novelty_score, correctness_risk, formal_verification, parameter_count	review scores
one_line_summary, lead, discovery_kind	the review's short text fields
reviewed_at, model_set	when the review ran and which models produced it
upvotes, downvotes, pith_score, inbound_count	community and citation signals at dump time

Cadence and reuse

Written monthly (1st, 04:00 UTC). Each month's file is a full snapshot, not a delta, so you only ever need the newest one. Review text and scores are machine-generated by Pith: reuse them with attribution to pith.science. Paper titles and abstracts remain subject to their sources' terms (arXiv and others). Published signed human reviews dump to dumps/<YYYY-MM>/signed-reviews.jsonl.gz via pith signed-review dump. The per-paper API and the assistant-facing MCP server are documented at /api.