Pith. sign in

CONTRACT 186,609 reviews published 2,905,484 papers cataloged 11,530,880 citation rows

Every paper gets an honest, public read.

pith.science · about · updated 2026-08-04

Read the review methodology

Pith machine-reviews the open literature, six structured passes per paper, for about three cents a paper, and publishes its measured error rate. Journals sell prestige and Scholar sells retrieval; Pith sells the truth-status of the literature, with receipts.

reviews published
186,609 live · as of 2026-08-04
papers cataloged
2,905,484 live · as of 2026-08-04
citation rows in the graph
11,530,880 live · as of 2026-08-04
cost per review, current model
$0.031 live · average over 35,758 reviews, trailing 30 days
planted-defect catch rate
5 of 6 measured 2026-07-31 · blind adjudication, frozen corpus
review model
DeepSeek V4 Flash high effort, screening clause · since 2026-07-31

The catch rate comes from a planted-defect audit: six papers seeded with known errors, reviews judged blind by an independent model against a rubric frozen before unblinding. On the six clean papers in the same audit, the current configuration was the quietest of the three models measured. The audit method and per-model scores are in the methodology.

  1. Every paper gets an honest, public read

    Careful review cost fell from roughly $1,000 of expert labor to three cents; Pith makes coverage universal.

    1. Six-pass structured machine review
    2. On the record, never anonymous or private
    3. Applied to the whole literature, not a chosen few

  2. A queryable claim map of science

    Reviews are structure, not prose: claims, assumptions, falsifiers, and citations accumulate into a living map.

    1. The load-bearing claim and weakest assumption per paper
    2. Circularity and citation-integrity checks
    3. A named falsifier for every central claim
    4. Who asserts what, resting on what, contradicted by what

  3. The citation graph with meaning

    Beyond who cites whom: which citation supplies the method, which contests the result, which the claim depends on.

    1. Canonical works with resolution audit trails
    2. Role and polarity labels on every citation
    3. Supporting citations: who depends on whom

  4. Review as a re-runnable computation

    Machine review upgrades every model generation; the scrutiny layer over the whole literature compound-improves.

    1. Versioned, re-runnable review pipeline
    2. Whole-corpus upgrades for car-cost each generation
    3. Human review is a snapshot of its year; this is not

  5. Credibility earned with receipts

    An AI review is worth its published error rate and nothing more; trust is measured, never asserted.

    1. Planted-defect corpora with measured catch rates
    2. Blind adjudication by independent judges
    3. Published cost and quality numbers

  6. Built for the AI-authored literature

    Paper volume will soon detach from human authorship; only machine review scales with what is coming.

    1. Review capacity that scales 10-100x with volume
    2. A trust layer for the new literature
    3. Neutral infrastructure, open to every field

  7. Open and exportable by design

    Centralized today, export-shaped always: signed records and public dumps make federation a deployment choice.

    1. Structured, minable records for every paper
    2. Signed integrity records and attestations
    3. Decentralization later without a rewrite

  8. What we do not claim yet

    Credibility at scale takes time, distribution is unsolved, and the paid service exists to fund validation.

    1. Published error rates over marketing
    2. Paid peer review as a validation signal
    3. No editorial stake in any paper's outcome

The product is the page. Three examples, unedited:

Pith is neutral infrastructure for scientific truth. It has no editorial stake in any paper's outcome.