Every law on this site was written down before it was measured, then checked blind. This page is how the checking is done.
BEFORE ANY RESULT IS SEEN
Before any result is seen.
Preregistration.
Before outcome data is unsealed, the claim is written down: what it predicts, where, and the line it must cross. The specification is frozen and hashed. Several preregistered questions have failed and stayed failed. That is the discipline working, and the failures are recorded beside the findings.
Blind specimen gates.
A reading is not allowed to read anything until it has passed a test on specimens it has never seen: a set of sentences judged by people first, then by the reading, and compared against a bar set before the test. Some readings were instead ratified in open court: the laboratory’s principal ruled on named examples in a written session that stays on the record. Every reading on the Anatomy did one or the other before it was run on the record, and each says which.
Sealed holdouts, read once.
A held-out set is a body of pieces kept aside and unseen until the claim about them is fixed. It is opened once, for one verdict. It is not re-mined, and a failed verdict is not re-run until it passes.
THE CHECKS, BY LAW
The register of checks.
The record of each check, moved here from the laws themselves.
The reading that finds which idea is the drum passed a blind specimen test, its answers checked against people’s on sentences it had never seen, before it was allowed to read anything; then run on every piece on the record.
PASSEDThe receipt reading passed a blind specimen test at each of its three rungs, a named person, an institution, a filing with a date; the ledger was then run on every piece on the record.
PASSEDThe hope and fear readings were ratified in open court, ruled on named examples in a session that stays on the record; the funding test passed its blind gate; measured on every piece on the record.
PASSEDThe shape of the close was written down before it was measured, read on the whole record, and sealed. The reading that finds the ask was written down before it was measured, then read on every piece on the record.
SEALEDMeasured on the strongest pieces on the record. A measurement, not yet a sealed law; when it is sealed, this line will say so.
MEASUREDACROSS FORMS
Across forms.
Invariance.
A law of the reader must hold wherever a mind meets a sentence. Readings built on long-form pieces are pointed at other forms with the predictions written down first. What travels is a law of the reader; what stays is a law of the form; the site says which.
AGAINST OURSELVES
Against ourselves.
Adversarial review.
Every derivation is reviewed by a model that took no part in it, with the epistemics questioned first.
The struck claim.
When a published claim fails a later test, the page says so plainly, and the claim stays where it stood.
Machines label; people judge; the market grades.
Language models may label what can be pointed to in the words, with the evidence quoted: this sentence is in the future tense, this one names a source. They never judge what a text will do. People adjudicate the reference sets the machines are tested against. Only the market grades what sells. No language model sits inside the engine.
The engine, as a fact.
The engine that produces a reading is deterministic. The same text gives the same reading, byte for byte, today and next year. Its model is sealed and versioned, and a reading names the version and the seal that produced it, so a reading can be checked against the engine that made it.
What is not published.
The instruments: which frames find which moves. The calibration: bars, weights, norms. The specimen record itself. The laws are published. The instruments are not.
REFERENCES
McCloskey, D. N., and Klamer, A. (1995). “One Quarter of GDP Is Persuasion.” American Economic Review, 85(2).
Antioch, G. (2013). “Persuasion is now 30 per cent of US GDP.” Economic Round-up, Australian Treasury.