Skip to content
ProofSentry Illustrative Walkthrough

Watch the guard work.

Two drafts: one written by an AI assistant, one by a researcher. Before either reaches a reader, ProofSentry — our first product — checks every scientific citation against the evidence. Pick a case and press Run.

AI Assistant Draft · Not Yet Delivered

    How It Works

    A checkpoint between writing and delivery.

    01

    Isolate

    Every factual claim in a draft, human or AI, and the source it leans on.

    02

    Trace

    Each claim, to the exact passage that supports or contradicts it.

    03

    Corroborate

    When a source is locked, the open papers that cite it testify to what it says.

    04

    Decide

    Supported claims pass with their evidence. The rest are stopped before delivery.

    The Benchmark

    A budget model, within 1.5 points of the ceiling.

    169,543 citation pairs sampled from PubMed Central, 2020–2025. The ceiling is a frontier model handed the full text of every cited paper — access real deployments rarely have. ProofSentry reaches nearly the same score with a budget model, and no such access.

    Miscitation Detection · F1 Score
    GPT-4oOn Its Own14.87
    GPT-5.5 xhighOn Its Own39.75
    GPT-5.5 xhigh+ Web Search81.80
    GPT-5.5 xhigh+ Full Text of Every Cited Paper98.00
    DeepSeek V4 FlashBudget Model + ProofSentry96.50
    GPT-4o+ ProofSentry95.54

    Ceiling: full text of every cited paper

    ProofSentry is where we start. The guard is where we're going.