Every finding was handed to a skeptic told to falsify it. What survived that is what got printed. The same discipline you can read for free, learn in person, or commission for your own corpus.
A measurement rig that prints only its wins is not a rig — it is a brochure. The discipline below is the part that earns the figures the right to be believed.
Six movements · M1 through M6
Movement one
Not reviewed for polish — adversarially falsified. A finding that wanted to be true was made to defend itself against a reading that wanted it false.
Stated the claim, fit the curve, found the supporting cases. The natural motion of anyone reporting their own work.
Re-matched baselines, hunted the confound, asked what the flattering version was hiding. Given equal authority to the first.
In 5 of 8 dimensions the baselines had been chosen where the effect looked best. Re-matched honestly, five of the eight headline wins shrank or vanished. The survivors are the three that did not.
Movement two
The verification that matters is the kind that overturns your own preferred reading. Each of these moved a number the wrong way for the party who reported it — and was kept anyway.
Claude’s own first pass counted eight moments of defensiveness in its replies. The skeptic, told to look harder, found one hundred and seventy-six — the model had under-reported its own defensiveness twenty-fold, in its own favour.
An early word-count silently swallowed the seams between concatenated documents. Restoring them moved the total up, not down — a correction that made the human side’s share of the page smaller, and was kept regardless.
A dead hypothesis tried to return as “Claude learned to say more with less.” It was register moving — more operational text, less exposition — wearing the costume of skill. Killed; see The Data, Figure 08.
An internal self-analysis quietly weighted Claude’s contribution upward. The skeptic caught the thumb on the scale and the figure was corrected into the substrate plate, Figure 00 — two co-equal claims, neither subordinate.
Across eight dimensions the baselines had been chosen where the effect looked best. Re-matched honestly, five of the eight headline wins shrank or vanished. Only three survived the rematch.
Movement three
It began as one kill: titles were being auto-generated and then cited as if authored. Generalised, it became a floor that the whole rig respects.
For three of every four pivots the body simply does not record who turned the surface into the principle. Reporting a confident author for those would be invention — so the rig reports undecidable, and the Pedro-attribution model that once guessed for them is retired. → The Data, Figure 09
Movement four
Some citations point outside the measured corpus. Their weight is unknown — which is not the same as zero, and not the same as decisive.
“We measured nothing there, therefore there is nothing.” The off-instrument edge is not evidence of absence; it is the absence of evidence, and is labelled as exactly that.
“The real effect is out there, off the instrument, and that’s why it didn’t show.” A rescue that can never be tested is not allowed to revive a dead hypothesis. → The Data, U1
Movement five
Of the correspondence, 503 letters are open and 258 are held. The gate is an absence in a lesser reader’s copy, never a curtain drawn over present text.
A one-year timer protects what is still being filed; the body opens when the timer does.
A handful of letters are simply personal, and stay closed for no reason grander than that.
Four moments are evidence about the instrument, not the subject — handled as sensitive, labels only, never bodies.
The raw inventory — UUIDs, the patent timer, the flag table — is set only into the owner’s copy. It is not hidden here; it was never printed into yours.
Movement six · the creed
Every figure attributes to a dyad, never to one hero. The human is admitted; the machine is named; no adjective is added to either.
The 90.4 % word-share and the near-parity contribution are two instruments, printed at one weight. Neither is allowed to be the headline that buries the other.
A null falsifies the hopeful story; it never proves the opposite. n = 1 can kill a claim — it cannot crown one.
One correspondence, n = 1. It can falsify “the AI atrophies the human” — and it has. It cannot establish “the AI enriches the human.” The structure’s whole job is to make that distinction impossible to blur.
Read the record for nothing. Learn the model at one desk in Joinville. Or have it applied to a corpus of your own. Every way in arrives at one inbox.
The open package — every un-gated letter, the data, the workshop, this method — costs nothing, and always will. A free subscription only keeps your place across the volumes.
The centaur model, taught at one desk: the student navigates the manifold while the soul document is held outside, and nothing is cited that can be derived. Built from Joinville materials, about Joinville children.
The discipline applied to a corpus or a school of your own — both-ways verification, provenance floors, framing that resists flattery. Scoped per problem, never sold by the tier.
No prices are dressed up. The reading is free and stays free; the course and the consultancy are scoped by conversation, not by tier card.
There is no intake portal, no ticketing queue, no auto-reply wearing a person’s name. What you send is routed, as written, to a single dashboard a single person reads.
Expect an overflow, not an immediate return. Pedro is a datum here, not a company — the reply comes when there is one to give.
The method earns the figures the right to be believed — and then opens the same door it walked through: read it, learn it, or commission it.