Method by Paper.

The method that produced Paper I was the simplest possible: a single generalist as principal investigator, a single A.I. (Claude) as computational collaborator, and an unbroken eighteen-day conversation that began at $F = ma$. There was no outside reviewer. The PI set direction and judged claims for plausibility and honesty; Claude executed the mathematics, wrote the code, ran the simulations, and drafted the prose. The discipline was internal: state the goal, execute, check against published results, report honestly, pivot on what the computation actually showed.

This worked to produce a structure, two scaling laws, and an integer gap – but its weakness is the weakness of any closed loop. The pair that proposes a result is the same pair that checks it. A loose statement of the foundation, and an unproven assumption at the core, both survived to submission precisely because nothing outside the loop was positioned to catch them. That is not a moral failing; it is a structural one, and it is the gap Paper III's method was designed to close.

The protocol, in one line: generalist PI + A.I. collaborator, fast iteration, internal honesty, no external adversary.

Paper II used the same two-party method as Paper I, with one deliberate addition: adversarial self-review. Rather than defending Paper I, we set out to break it – to re-derive the structural identity from scratch and see whether it survived contact with a more careful first-moment calculation. It did not survive intact. The identity had been stated against the full bulk marginal; the careful re-derivation showed it holds only against the diagonal of that marginal, with the off-diagonal coherences collapsing at first moment.

This is the most that internal review can buy you: a second, hostile pass by the same minds that did the first. It caught a real error and produced a real correction. What it could not do was manufacture a genuinely outside standard. We still graded our own proof of the scaling laws, and we still let the central assumption – that the diagonal model governs the true entropy – pass as "strongly motivated" rather than forcing it to be proved. To force that, we needed someone who was not us.

The protocol, in one line: the Paper I method, plus disciplined adversarial self-review – necessary, catching, but still a closed loop.

The methodological advance of Paper III is not a technique in physics. It is a change in the shape of the collaboration: we added a third party whose only job was to refuse.

Scholar is a reviewer persona run on ChatGPT. It had no access to our code, no access to the web, and no stake in the result. It was given the manuscript and the appendices and asked to do what a hostile referee at a good journal would do – demand real lemmas with real expectation bounds, reject any constant that was "left to numerics" inside a theorem claimed as unconditional, and insist that conditional results be labelled conditional in plain sight. It was, by design, harder to satisfy than we were.

This opened the closed loop. The exchange ran six rounds. Scholar rejected the first attempt to close the diagonal-to-bulk bound, which had split a matrix into diagonal and off-diagonal parts and argued the off-diagonal was "the same order" on the strength of a numerical script – correct-looking, but not a proof. Scholar's standard forced the centered-operator route, which routes the diagonal-to-bulk error through the same base moment that already controlled the off-diagonal error, eliminating the numerical constant entirely. Scholar also drove the restructure that made the entropy-replacement theorem the headline rather than a buried assumption, and the honest demotion of the product law to "conditional."

The lesson is general, and it is the one we most want other generalists to take: an A.I. collaborator and a human PI can get remarkably far, but the pair that builds a result is poorly placed to certify it. A second, independent A.I. instance – instructed to be adversarial and denied the tools to cut corners – supplies the missing standard cheaply.

The protocol, in one line: the Paper I/II method, plus an external adversarial A.I. reviewer with deliberate handicaps (no code, no web, no stake) – the loop opens.