the problem
Most claims land in a pile marked “maybe.”
The claims you meet all day aren't neatly true or false. They sit in a vast middle where you honestly can't tell — and you have no system for sorting them.
You can't fact-check everything; there isn't time, and most claims never get checked at all. So they accumulate — half-believed, half-doubted, unresolved. A researcher's own working estimate is that something like half of what crosses their desk lands in exactly this pile.
The honest reader's problem was never a shortage of information. It's the absence of a first move: a fast, repeatable way to decide which claims are even worth the effort of checking — and which are a waste of it. PAS is that first move. It doesn't empty the maybe pile by telling you what's true. It sorts the pile by telling you what's checkable.
The bottleneck was never access to claims. It's knowing which ones deserve your attention.
the core move
Before “is it true?”, ask: could it be shown false?
A claim that nothing could ever disprove isn't strong. It's empty. That one reversal is the whole engine of PAS.
The idea has a century of philosophy of science behind it — Popper's falsifiability — sharpened here into a working test. A claim earns the name empirical only if some observation, some measurement, some outcome could in principle go against it. If nothing could ever count as evidence against a claim, then nothing counts as evidence for it either. It's language that feels factual without committing to anything about the world.
So the first question PAS asks isn't “is this right?” It's “is truth even the right question for this claim yet?” Getting that reframing in your bones is most of the value.
One consequence trips everyone up at first: “Proceed” does not mean true. It means the claim is the kind that evidence can settle — go check it. A false-but-testable claim passes the screen cleanly, because being wrong in a checkable way is exactly the condition science is built to correct. The screen is about the shape of the claim, never its correctness.
the discrimination
Three categories, one reading of what kind of claim you hold.
The category isn't a grade. It reflects what would — or could never — settle the claim, and therefore what to do with it.
Proceed
Empirical and testable. Some observation could contradict it, so evidence can do its work. It may turn out false — that's fine; falsity is checkable. This is the category where fact-checking belongs.
Failed empirical
Framed as a claim about the world, but built so nothing could disconfirm it — any outcome confirms it, or its key terms can't be measured. No evidence moves it, because it absorbs everything. The claim is broken upstream of the data.
Out of scope
Not an empirical claim at all — a value (“should”), a definition, a matter of taste, or a settled proof. This isn't a failure; evidence is simply the wrong instrument. The error is arguing it as though it were a question of fact.
Notice the three aren't ranked best-to-worst. They're different kinds of thing. The work PAS does is telling them apart quickly and consistently — and showing you the reasoning, so you can disagree.
the important catch
The claims that do damage are wearing a lab coat.
Nonsense that announces itself is easy to dismiss. The dangerous claim is the one that borrows the furniture of science while committing to nothing.
It arrives with numbers, a mechanism, technical vocabulary — and underneath, nothing that could ever be wrong. “Subtle energy fields govern health.” A prediction that was quietly fitted to the data after the fact. A theory that explains every possible outcome equally well, and so forbids none of them. These slip past casual inspection precisely because they sound rigorous.
Failed empirical is the category PAS is tuned hardest to catch, for two reasons. It's where skepticism is needed most and spent least — the costume does its job on tired readers. And waving an untestable claim through as if it were testable is the single costliest mistake the tool can make: it's how dressed-up emptiness gets treated as live science. So the framework leans careful here, and tells you plainly when a claim only looks like it's saying something.
There's a humane flip side. A claim can read as untestable only because its source defined the terms elsewhere — a study's “significant,” a named protocol. PAS flags those as context-limited rather than condemning them: paste the definitions, screen again, and they often pass. The point is accuracy, not suspicion.
the payoff
It doesn't think for you. It tells you where to think.
Each category is a pointer. It tells you your next move — and, just as useful, which moves would be wasted.
On a Proceed →
Go find the evidence. This is where studies, data, and fact-checking actually earn their keep. The question is live; answer it.
On a Failed empirical →
Stop hunting for evidence — no study rescues a claim that can't be tested. Ask instead for a definition, or for the single condition under which the claim would be false. If none exists, you have your answer.
On an Out of scope →
Recognize the argument you're actually in. You're debating values or definitions, not facts — so bring reasons, not measurements. Demanding data here is a category error.
The quiet win is that you stop fighting the wrong fight: combing for data on a claim no data could settle, or demanding proof of what was always a value judgment. Aiming skepticism is most of the skill. PAS just makes the aim fast.
the honesty
A tool that admits where it can't decide is worth more.
PAS shows its reasoning every time, in every tier, on purpose. The label is quiet; the argument is loud.
That's a design stance, not a courtesy. You're meant to be able to read the reasoning, check it, and overrule it. A screen you can't inspect is just an authority asking for trust — which is the exact posture the tool exists to push back on.
And on roughly 1 in 14 claims, the honest output is: this sits on a genuine boundary. These are seams where careful, independent judges disagree — numerology edging toward cause, the line between the metaphysical and the empirical, claims too vague to place. The framework declines to force a call, and says so. Most tools hide their uncertainty to look authoritative; PAS maps it, because a mapped doubt is useful and a hidden one is a trap.
A guide you can argue with, not an oracle you have to trust.
the next layer
Passing the screen isn't the same as being sound.
The screen asks whether a claim can be tested. A claim can clear that bar and still rest on a broken calculation or a misread statistic — so there's a second, deeper pass.
Triage is the first move, not the last. Once a claim proceeds, the two places confident-looking work most often hides a flaw are its math and its statistics — and those failures are invisible to the screen, because structurally the claim is fine. A constant that was quietly fitted to a known answer and then presented as a prediction. A real, measured correlation reported as if it proved a cause. Both pass the screen cleanly; both are wrong one level down.
PAS has two specialist audits for exactly this — one for derivations, one for statistical inference — that pressure the reasoning inside a claim that already passed. They're a deeper tool for a later question, which is why they sit in the higher tiers rather than the first screen.
One honesty note, because it's the project's whole posture: these audits are validated to a weaker standard than the core three-category screen — tested across models rather than by independent humans — and PAS says so plainly rather than letting them borrow the core's credibility. See the specialist audits →
the moment
When fluent text is free, sorting claims is the scarce skill.
Persuasive language can now be generated at scale — polished, confident, and indifferent to whether it commits to anything real.
In that world the bottleneck shifts. It isn't producing claims, and it isn't even looking them up. It's the fast structural judgment that decides which claims are worth a lookup at all — and which are smooth surfaces over nothing. The more fluent the text gets, the more that judgment is worth, and the rarer it becomes.
Here's the durable part: because the judgment is about structure rather than content, it doesn't depend on any one model knowing the answer. It travels across models, across topics, across the next wave of tools. As inference moves onto everyone's own hardware and the model underneath becomes a commodity, a validated way of reading claims is the part that holds its value.
PAS grew out of Greg Glassman's “46 premises” — a working set of rules for telling a real claim from one only dressed as one — refined and tested into the framework described here.
Now try it on a claim of your own.
Pick something from your own maybe pile — a headline, a study's abstract, a line from a podcast — and watch it get sorted, reasoning and all.
Screen a claim →