claim screener · the step before fact-checking
Where to aim your skepticism.
PAS reads a claim for one thing: can it be tested at all? It sorts the claim into one of three categories and shows its reasoning. It won't tell you whether a claim is true — it tells you whether truth is even the right question yet, and where to point your attention next.
→ a triage aid, not a verdict machine
→ judges epistemic structure, not correctness
→ shows the reasoning in every case
the discrimination
The category isn't a grade. It's a reading of what kind of claim you're holding — and therefore what would, or could never, settle it.
Empirical and testable — some observation could show it wrong. Worth engaging on the evidence.
example
“Grounding bare feet to the earth reduces inflammation.”
Proceed ≠ true. It may be false — but it's checkable.
Framed as about the world, but nothing could disconfirm it — self-sealing, or its key terms resist measurement.
example
“Subtle energy fields govern the body's return to health.”
No observation would count against it.
Not an empirical claim at all — a value, a definition, an aesthetic judgment, or a settled piece of math.
example
“A scientist should always disclose their funding.”
Settled by argument, not evidence.
try it
The verdict is quiet; the reasoning is loud. That's the whole design — you're meant to read the argument, not obey the label.
Accuracy is measured per model. The framework is the product — it runs against any endpoint, cloud or local. But a model you point it at is only as reliable as it's been tested to be; weaker models screen measurably worse. When you choose your own, PAS shows you the reliability map, not a bare dropdown.
why trust it
The core framework was tested on 101 claims it didn't author, blind-adjudicated by three independent human reviewers from different fields. Here's what that found, including the part most tools would bury.
101
outside-sourced claims in the gold set — none written by the tool's builders
0.874
Fleiss κ across three independent human reviewers — strong agreement before the tool was scored
~90%
accuracy on the full set; ~97% on the 94 claims the panel considered settled
◆ the 7% we don't hide
On roughly 1 in 14 claims — 7 of the 101 — the framework has no settled answer, and neither did the independent human-and-AI panels. These sit on genuine seams: numerology edging toward cause, the metaphysical-vs-empirical boundary, claims too vague to place. PAS flags them “on a boundary” rather than forcing a call. A tool that pretends to be certain here is lying to you; mapping the contested band is the honest alternative.
The two specialist audit modules — one for mathematical derivations, one for statistical inference — are validated by cross-model consensus rather than human adjudication, a deliberately weaker tier of evidence that's kept distinct from the human-anchored core. Accuracy and honesty over polish: the number you see is the real one, even when a nicer one was available.
what it is not
Not a verdict machine
It screens whether a claim can be tested — never whether it's true. “Proceed” means worth checking, not correct.
Not a fact-checker
It does no lookup and gathers no evidence. It's the step before fact-checking — deciding whether fact-checking is even the right move.
Only sees what you paste
A claim can read as untestable because its source defined the terms elsewhere. PAS flags those “context-limited” — paste the definitions and re-screen.
Model-dependent, by nature
It's a prompt run on an LLM — non-deterministic, and measured between 70% and 94% depending on the model underneath. The framework is stable; the substrate isn't.
access
The same screener serves a curious reader, a researcher drowning in papers, and a team that needs claim-triage built into a pipeline — but those are different jobs, so they get different products. Reasoning stays visible in all three; what changes is scale and workflow.
For anyone meeting a questionable claim and wanting to think straight about it.
Its job is reach and trust — the way most people first meet PAS. A strange-sounding claim that's testable still passes; that's the teaching, not a bug.
For people who evaluate unfamiliar papers and claims for a living — and want help aiming, not a machine deciding for them.
The job-to-be-done: “I have too much to read — tell me which claims deserve my limited verification time.” Subscription, roughly $20–50/mo (still in willingness-to-pay testing), with heavier runs like specialist audits as metered add-ons. The early revenue engine.
For teams and platforms that need claim-triage as infrastructure, under their own controls.
Decision-support infrastructure — PAS maps where scrutiny is warranted; it never “screens the deal” or rules a claim valid. High value, but deliberately after professional use is proven.
Reasoning is never the thing behind the paywall — legibility is the whole thesis, so every tier shows its work. What the paid tiers add is scale and workflow: whole-paper mode, model control, cross-checking, team infrastructure. The paid and enterprise tiers are in design and sequenced after a hosted backend; this is the shape, stated honestly ahead of the build.
your data
A screening tool is only useful if it keeps a record of what you've sorted. PAS keeps that record for you — and is deliberate, and public, about where it lives and which models are ever allowed to touch it.
Your log — yours to keep
Every claim you screen, its verdict, the reasoning, and your own notes are saved to your log, and export to Markdown or JSON whenever you want. You own your inputs and your outputs; nothing about the record is held hostage to a subscription.
On the local-model path, the record can stay entirely on your machine — nothing leaves your device at all.
The shared corpus — only if you opt in
Screenings can feed a shared corpus that makes PAS better over time — measuring real-world accuracy, growing the validation sets, mapping the contested band from real use rather than one study.
That only ever happens with your explicit consent. No claim joins the corpus unless you opt in. Harvesting your work quietly would betray the one thing this tool is built on.
◆ the model we use
PAS runs on Anthropic's Claude, on a tier where your inputs and outputs are not used to train their models. That's a deliberate choice, and it's the reason it's the default.
If you ever point PAS at a different model, that model's data policy is between you and its provider — the AI landscape shifts constantly, and understanding a given provider's terms is a call we leave in your hands rather than pretend to track for you.
One thing worth knowing, because it trips up almost everyone: a consumer subscription — ChatGPT Plus, a paid Gemini plan — does not by itself stop a provider training on your chats. That's usually a separate setting, or a business/API tier. We mention it only so the choice you make is an informed one.