specialist audits pro
The screen decides whether a claim can be tested. But a claim can be perfectly testable and still rest on a broken derivation or a misread statistic. Two specialist audits pressure the reasoning inside a claim that already passed — the two places where confident-looking work most often hides a flaw.
1 · the screen
Is this claim testable at all?
if it proceeds
It's checkable — now check it
2 · the audit
Does the reasoning inside it hold?
A different, lower tier of evidence — stated up front. The three-category core was validated against an outside-sourced gold set, blind-adjudicated by three independent humans (κ = 0.874). These two audit modules were not human-adjudicated. They were validated by cross-model consensus — six models, each flaw tested in both directions, one weakness found and fixed per module and confirmed on a reference model. That's real, repeatable evidence, but it is weaker than the core's, and PAS keeps the two tiers distinct rather than letting the modules borrow the core's numbers. How the core was validated →
Audit one · derivations
When a claim leans on a calculation or a derived constant, the audit checks whether the math actually predicts the number — or whether the number came first and the math was arranged around it afterward.
worked example
“Our model derives the fine-structure-like constant 1/137 from first principles, confirming the theory.”
what it caughtThe “derivation” reaches a target that was already known before the model existed. Two of its constants aren't fixed by the theory — they're tuned until the output lands on 137. A derivation with enough free parameters can hit any target, so arriving at the famous one confirms nothing.
why the screen missed itStructurally the claim is fine — it's testable, and a different derivation could contradict it. The flaw isn't in whether it can be checked; it's in the check itself being rigged. That's exactly the layer the screen doesn't reach and the audit does.
what to ask forHow many parameters were free to move, and were they fixed before the target was known? A real derivation predicts the number from constraints set in advance; a fitted one needs the answer first.
Status: built, V2. Validated across six models in both directions (does it flag real retrodictions, and does it leave sound derivations alone); one boundary weakness found, fixed, and confirmed on the reference model. No human adjudication of its cases — a weaker tier than the core, kept distinct.
Audit two · inference
When a claim rests on data, the audit checks the inference — whether the numbers support the conclusion drawn, or whether a confound, a fishing expedition, or a base-rate slip is doing the real work.
worked example
“In our eight-week pilot, the grounded group's inflammation fell significantly — earthing works.”
what it caughtThe grounded group also changed their sleep and routine during the trial, and there was no sham-grounded arm. The CRP drop is real, but the design can't separate earth contact from everything that moved alongside it. “Fell significantly” supports something happened, not earthing did it.
why the screen missed itThe claim is structurally testable and its terms are measurable — the screen has no quarrel with it. The failure is one level down, in the leap from a correlation to a cause. That leap is invisible to triage and central to the inference audit.
what to ask forA sham-controlled design — grounded vs. convincingly-not-grounded, same routine in both arms — measuring the same marker. Until the confound is isolated, the result stands; its stated cause doesn't.
Status: built, V2. Validated across six models in both directions (does it flag real inference errors, and does it leave sound inferences alone); one boundary weakness found, fixed, and confirmed on the reference model. No human adjudication of its cases — a weaker tier than the core, kept distinct. This is the same confound that paper mode surfaces as a paper's crux — the audit is what pressures that crux.
Design mockup — the worked examples above are authored, on invented claims, to show each audit's shape. Both modules are built and cross-model validated; wiring them into the live tool, like model-choice and paper mode, waits on the hosted backend and follows the project's discipline: the specialist audits earn their place in the public product only once the core screener proves useful in real hands. The evidence tiers shown here are stated as they actually stand — the core's human validation is the stronger claim; these modules are the weaker, honestly labelled.