Every current review, in one file a month
On the first of each month Pith writes every current review (the most recent one of each paper) to one gzipped file with a line of JSON for each paper. The list of files is public.
The files themselves are not publicly readable yet: a download link taken from the list answers 403. To get a file, write to hello@pith.science and name the month you want.
What the list of files holds
curl -s https://pith.science/api/v1/dumps/manifest.json
# {"dumps": [{"key": "dumps/connected-record/...", "size": ..., "last_modified": "...", "download": "https://..."},
# ...
# {"key": "dumps/2026-11/verdicts.jsonl.gz", "size": ..., "last_modified": "...", "download": "https://..."}]}
The list has at most 1,000 keys. The monthly files are the keys of the form dumps/<YYYY-MM>/verdicts.jsonl.gz; on 13 Nov 2026 there were three, for September, October and November 2026.
Every other key is a part of the connected-record export, under dumps/connected-record/, and those keys come first in the list. So the first key is not the newest monthly file: pick the newest key that ends in verdicts.jsonl.gz.
When the list cannot be read it answers 503 with an empty list and a sentence under "error".
What a line of the file holds
One JSON object for each paper.
-
arxiv_id, title, primary_cat, cross_cats
which paper, and its arXiv categories
-
submitted_at, latest_version, source, source_label, source_url
where and when the paper appeared
-
verdict_id, verdict, confidence
the current review and its result
-
novelty_score, correctness_risk, formal_verification, parameter_count
the review's scores
-
one_line_summary, lead, discovery_kind
the review's short texts
-
reviewed_at, model_set
when the review ran and which models wrote it
-
upvotes, downvotes, pith_score, inbound_count
readers' votes and citation counts when the file was written; upvotes counts votes cast by people
When it is written, and what you may do with it
- Written on the first of each month at 04:00 UTC. Each month's file is everything, not what changed, so the newest is the only one you need.
- A review's text and scores are written by a machine at Pith: reuse them and say they come from pith.science. A paper's title stays under the terms of where it came from (arXiv and others).
- The file does not say which reviews were written from the abstract alone. The API does (the basis field of /api/v1/verdicts/{arxiv_id}), and a review of that kind is a first read, not a full review.
- Signed reviews by people are written to dumps/<YYYY-MM>/signed-reviews.jsonl.gz when that job is run; on 13 Nov 2026 the list had no such file.