1. What is being written, and against what 1
2. Why a process at all 2
3. Statements first: the spine 2
4. Style rules (binding, partly enforced by a program) 3
5. Who does what 3
6. The cycle for each section 4
7. Automatic checks 5
8. Records 5
9. What this has caught so far 6
10. What the process does and does not establish 6
Changes to this process 7
This page describes, in detail, how the manuscript is being written: who does what, what is checked, by whom, against what, and what the checks do and do not establish. The process is still in use and is revised when it fails; the changes are recorded at the end.
The manuscript gives a complete natural-language proof of the result formalized in OpenAI's Lean proof of the construction of four-dimensional quantum Yang–Mills theory, at the fixed commit 3456e5e. It has two parts:
Part I (about 30 pages): the setting, the definitions, the main theorems, the principal statements on which the proof rests, each with the idea of its proof, and the complete deduction of the main theorems from them.
Part II (about 120 pages): the proofs, in ten sections, numbered 10–19 after the nine sections of Part I.
The authority is the Lean. The manuscript claims nothing the Lean does not prove. A statement may be weaker than its Lean counterpart, or stated in a more natural form, but never stronger. "The Lean" means the declarations that the main theorem actually depends on, as computed from the compiled files (about 74,000 declarations; see the method). The source also contains alternative routes that the main theorem does not use; these are not evidence for anything in the manuscript.
The rest of this repository (the summary, the unit outlines, the dependency graph) serves as a map for finding things in the Lean. It is never cited as evidence.
The obvious approach is to ask a capable model to write the paper from the Lean. We tried that, and it failed in ways that are easy to miss:
Statements take the shape of the formal code. The first drafts were faithful but unreadable. Propositions ran to nearly a page because they copied the Lean's long conjunctions and packaging structures. One section was 92% definitions, one of them 776 words long. Those drafts were rejected.
Length grows when nobody measures it. The unit outlines in this repository grew to 4.4 times their intended length over several rounds of review, because each round added detail and none measured the total.
Prose sounds true without being true. An explanatory sentence ("this follows from the bound of Lemma X") can be plausible, fluent and wrong.
Conditions disappear quietly. When a statement is shortened, a hypothesis or a binding of variables can be lost, and nothing breaks: the text still compiles and still reads well.
A model that wrote a proof is a poor judge of it. It tends to confirm its own reading.
The process below is built against these failure modes. Much of it comes from an earlier project of the same kind, a paper written from a completed Lean proof, where the same failures occurred.
Before any prose is written, every statement of Part I is fixed in a single document, the spine:
the definitions Part I needs;
the main theorems: Theorem A, the existence statement exactly as formalized, and Theorem B, the detailed construction as far as the Lean proves it;
fifteen principal statements, from the one-step renormalization bound (RG) to the contact extensions of the operator product expansion (CT);
the deduction of Theorems A and B from them;
the list of Part II lemmas that prove them.
Each statement is written in natural form: one result per statement, quantifiers first, conclusion displayed. The objects it uses are set up in the text before it, and side conditions of the formal version go to Part II lemmas. A clause map then checks that nothing was lost: every clause of the Lean's main statements is assigned to a statement of the spine, and the union of the natural statements must equal what the Lean proves.
Freezing. The project lead approves the spine, after which the principal statements and the main theorems are frozen. A frozen statement can change only through the drift protocol: the old text, the new text and the reason go to the project lead, and nothing changes without approval. Part II lemmas are not frozen. They can be split, added or reworded as the proofs are written, and every such change is recorded in a drift log.
| rule | limit |
| statement (theorem, proposition, lemma) | target 100 words, hard cap 150 |
| definition | target 120 words, hard cap 200 |
| ideas per statement | one |
| "there exists a family such that the following assertions hold" | never: the conditions go in the setup |
| definitions | concrete, by formula, with one sentence of purpose; never a translation of a Lean structure |
| Part II | about one numbered lemma per page; a proof longer than three pages is split at a genuine intermediate statement |
| standard results | cited exactly: name, hypotheses, numbering, checked against the source |
| Lean names, file paths, process vocabulary | never in the text |
| references between numbered items | always by cross-reference, never by a typed number |
The exposition follows the presentation of the renormalization group in the work of Bauerschmidt, Brydges and Slade, in this order: the objects, the inductive step, the multiscale theorem, the continuum results, the deduction of the main theorem, the constructions.
| role | who | does |
| project lead | a person | sets the goals, approves the spine and every change to a frozen statement, reads each finished part and sets the level of detail and length |
| editor | an AI model (Claude) | writes the tasks, integrates proposals into the manuscript (the only one who edits it), runs the checks, keeps the records, decides on findings |
| writers | AI models (GPT-6.1 Sol, high or extra-high reasoning effort) | write proposals for one section at a time, in a separate directory; never edit the manuscript |
| reviewers | AI models (Claude Opus), each in a fresh context | read a frozen snapshot and report; never edit anything |
The separation is deliberate. Writers propose and the editor integrates, so that one reader sees the whole manuscript and its notation stays consistent. Reviewers have no access to the writer's reasoning, only to the text, the spine and the Lean, so they cannot inherit its mistakes.
Task. The editor writes the task: what the section must contain (its statements from the spine), its sources, its page allowance and the rules above. Over the allowance means not done.
Proposal. A writer drafts the section in its own directory. Alongside the text it delivers a table pairing each statement with the Lean declarations it rests on, notes recording what it could not do and its least certain sentence, and the results of its own checks.
Integration. The editor reads the proposal in context, integrates it into the manuscript, compiles, and runs the automatic checks.
Frozen snapshot. The section, everything before it, the spine and the style rules are copied into a snapshot with a manifest of SHA-256 hashes. A review verdict applies to that snapshot only.
Independent reviews, each by a separate reviewer:
Owner of each principal statement. One reviewer per statement. It reads the manuscript in order up to the statement before opening any source, then maps each clause to the spine and the Lean and tries to break it. It judges three criteria: (i) the statement is correct against the spine and the Lean; (ii) it can be understood from the text before it; (iii) it serves the statements that use it.
Mathematical review (Part II): every proof step is checked against the Lean. Is it correct, complete at the stated level of detail, and no stronger than what is proved? Gaps written around are the main target.
Fresh reader: an expert reader with no prior context, and no access to the Lean, reads the text linearly. It reports every place where it had to guess an object, a normalization, a quantifier or the next step, judges the length, and ends with a two-sentence summary of what the section proves. The summary is compared with what the section is meant to prove.
Prose truth: every explanatory sentence outside the statements is marked true, false or unsupported, with evidence. Every condition removed in a rewrite is traced to the places that used it.
Response table. Every finding gets a row: the finding, the decision (applied, or declined with a mathematical reason), the exact edit, and how it was checked. Findings are never deleted or rewritten. The fixes are made by a writer as a new proposal and integrated by the editor.
Re-check. The same reviewers check the new version against their own earlier findings and mark each one resolved or not resolved. They also check the reasons given for declining findings, and report anything the fixes broke.
Delivery review. A last reviewer checks:
that the compiled PDF matches the reviewed sources;
that every finding is closed or deferred to a named owner;
that every principal statement passes all three criteria;
that every proof address exists;
that all references resolve.
It looks at every rendered page and states what its review does not establish.
The project lead reads it. Comments on level and length become binding style rules for the sections not yet written.
Part II starts with two pilot sections, one on the ultraviolet side (§12, path actions and the coupling reset) and one on the infrared side (§15, infinite volume). They go through the whole cycle and are read by the project lead before the other eight sections are written. Each later section then follows the same cycle.
Run after every integration:
compilation: no errors, no undefined or duplicate references, no overfull lines;
shape: the word count of every statement and definition against the limits of §4;
vocabulary: no Lean names, no file paths, no process words, no words on a list of habitual machine phrasing; the check also covers TeX comments;
Lean links: every Lean name in the correspondence table must be among the declarations the main theorem depends on. For Part I: 657 rows, 807 names, all of them dependencies of the main theorem.
Passing these checks means none of these mechanical faults is present. It says nothing about whether the mathematics is right.
The statement register: one row per numbered statement, giving its owner, mathematical status, review status, rendering status, where in Part II it is proved, and its open findings.
The correspondence table: each numbered statement and definition, with the Lean declarations it corresponds to and the relation between them (direct, weaker, composition, cited, none). It will be published with the manuscript.
The response tables of every review round, and the drift log of every change to the spine.
The working records (tasks, review reports, response tables, the statement register) are kept in a separate working repository.
Examples from the first rounds. None of them would have been caught by compiling the text:
Theorem B was weaker than the Lean. As first written, it said that hierarchy data exist, not that they are the objects constructed. Its owner rejected it on criterion (i). The repair changes the wording of a frozen statement, so it goes to the project lead under the drift protocol.
A pilot section assumed what it had to prove. The first draft of §15 used two inputs, the correction estimate and a rank comparison, as hypotheses with no statement to cite. The mathematical review found the Lean declarations that prove them. The revision states them as labelled lemmas matching the Lean.
A hypothesis was missing. The first draft of §12 omitted a bound on a kinetic constant that the Lean carries explicitly. Without it, one of the section's identities cannot hold. The bound now appears in the definition, in Part I and in the table of constants.
The editor's own edits introduced errors, and were caught in turn:
axiom labels stripped from definition titles during integration (Theorem A refers to them);
a sentence left broken by a deletion;
Part II numbers that collided with addresses used in Part I.
The third led to the rule that numbered items are referred to only by cross-reference.
True-sounding prose that was false. For example: a sentence attributing a smooth extension to the wrong estimate; a unit stated so that, read naturally, the axiom would be false for fields of positive degree.
It does establish that every numbered statement has been compared, clause by clause, with the Lean declarations it rests on, by a reviewer who did not write it. It also establishes that every explanatory sentence has been checked for truth, and that every finding has a recorded decision.
It does not establish correctness in the sense of a human referee's report. The reviewers are AI models; they are independent of the writers but not of each other's training, and they can share blind spots. The correctness of the theorem rests on the Lean, which is machine-checked. The faithfulness of the manuscript to the Lean is checked by the reviews above, not by a machine.
It is not finished. Until a section is marked checked in the status table, it is a draft. Corrections from readers are welcome as issues.
11 November 2026. Statements first, in natural form, with the length limits of §4. This was adopted after the first drafts were rejected for their shape.
11 November 2026. Numbered items in Part II are referred to only by cross-reference. Part II sections are numbered 10–19, continuing the numbering of Part I.
11 November 2026. After the first draft: (1) a cross-section check compares every place where one section cites another's result with what that result states, and mismatches are repaired on both sides (three rounds, the last confirming all items resolved); (2) an independent referee review of the whole manuscript (three referees, on Part I, Sections 10–14 and Sections 15–19); every suggestion is adjudicated against the text and the Lean (accept, accept modified, reject, or left to the project lead) before writers apply it, and a confirmation review checks the applied changes; (3) main theorems are measured too, and a statement may cite at most four other results.
12 November 2026. Second draft: an editorial pass merging two expert human reviews, six reviewer reports (four Opus at maximum effort, two GPT-6 Astra) and a comparison with a collaborator's draft into one plan, with page budgets per section (ceiling 265 pages); confirmation reviews of every changed statement against the Lean and a fresh reader asked to explain five mechanisms of the proof; targeted re-checks before publication. The statement-to-declaration correspondence is published as correspondence.tsv.
13 November 2026. Third draft: an exposition pass on the second draft from a handoff of reading experiments: six writers, one per group of sections, each within a page allowance (ceiling 286 pages); every new example, table and computation then checked by confirmation reviewers of a different model family against the text and the Lean, with a targeted re-check of the one item they required.