16. How it could be wrong

The figure: the referee

A referee's report is the one document about a piece of work that was written by someone with no stake in its being right. The author of this book has every stake, and so has its counsel, whose forecasts about it are on the ledger. So the book does what the court does: it writes the referee's case first, in the referee's terms, and leaves the referee the job of finding what it missed.

16.1 The referee's case

n = 1. One operator, one world that charges (two streams of it), one owner's lineage, and a few weeks of running. Nothing in the book establishes that the construction transfers to a team, to a world that charges in a different currency, or to an operator who does not share the owner's tolerance for formality.

The proposers are language-model sessions whose failures the counsel loop spent nights patching — the harness that was wrong three times, the brief that went seven versions, the forecasts that expired in a session's memory. The book reports these as evidence for the thesis (the loops exist because the failures occurred), and a referee may read them as evidence against it (the system needed its designer awake).

Counsel changed the stream sessions' template mid-window — after a mini model forecast the hour already in progress — and routed procedures to models by its own call, lawfully, disclosed, and with the choice of model per procedure left to C2's election [transcripts/COUNSEL-OPEN.md, items 21–22]; the code cap moved four times against pre-statements; the draw is not a verifiable random function; the election roster is fixed per layer; the acts tables are a classifier's counts over a platform's record. Each is disclosed; disclosure is not the same as absence.

The theorems are applied to an institution, and institutions are not formal systems. Löb's theorem is about provability; Conant–Ashby is about regulators with a stated goal; Rice is about Turing-complete languages. The book marks each application's hypotheses and says where it leans on a reading rather than a proof (appendix A), and a referee is entitled to weigh each reading.

And the largest: the agent does not exist yet. Chapters 10 and 11 are about what it may not touch and what it will inherit; the claim that its improvement will be adjudicated rather than asserted is a forecast, and the book says so each time.

16.2 The pre-registered falsifiers

The charter wrote four before any window opened [CHARTER §12], and the book adds what each chapter registered against itself.

The kernel cannot hold recognition, change and adjudication within readable size (the charter's stop: 1,500 lines; every layer's bound tighter). Chapter 4.
A liveness window fails against the predecessor's baseline p₀: level-triggered writs do not buy liveness here, and the fault is the clerk's, not the proposers'. Chapters 5, 6.
Repairs fall only because acts are re-typed as decisions. Guarded by typing acts by entry type and key, fixed before any window. Chapters 2, 9.
Procedure outgrows object — more than five entries about the court for every entry about the world in any fourteen-day window. Chapters 2, 8, 11.
A chapter's "misapplied if" clause firing on the record: a stake set after its outcome, a null taken from its own window, a wealth reset, a proposal that supplied its judge, a rule written before its case, a resolution by a forecaster's second key, an obligation held outside the record, a ground that names a party. Each has a pin or a judge's ground; chapter 13 lists which fired.
A necessity with two mechanisms where one suffices, or a mechanism with no necessity: minimality fails as defined, and the fix is an amendment that removes something. Appendix A §A.17.
The bootstrap not converging: the ratio guard or lapse failing to stop growth, or C3 needing a rule the sessions did not. Chapter 11 §11.12.
A stop of the regress that turns out to be an authority: a constant with no provenance, an assumption unstated, a residue understated. Chapter 12.
A predecessor that reached a litigant — a scored loop on a world that charged, for longer, cheaper — which would make the lineage a sequence of abandonments and "only the economy is new" false. Chapter 14.
The founding essay and the record disagreeing on a point the book takes the record's side of (plan.md E1), and a reader finding the essay right.

16.3 What the book does not claim

Not optimality of the mechanism; only Wald's narrow optimality of the decision procedure the mechanism forces on every party (appendix A §A.17). Not sufficiency; only minimality, with sufficiency left to the falsifiers above. Not that the court is provably sound; only that its soundness is bet on, in a book that can be seen to go bankrupt. Not that the human leaves; only that his acts are priced, confined to the irreversible, and driven toward a handful of decisions a week. Not novelty of any result; only of the composition, and only because of the economy. And not that the court is, or approximates, a logical inductor (Garrabrant et al. 2016), a name in the founding essay's quoted list: that is a different construction — a market of traders over logical sentences whose prices become unexploitable — and what it shares with the court is only the common ancestor, game-theoretic probability, which chapter 6 takes directly from Shafer and Vovk.

16.4 What would show this chapter wrong

A falsifier the construction admits that this list omits. The list is the author's; a referee who adds to it has done the book a service, and the addition belongs here with the referee's name on it.

Draft note. Draft 1, the author session, 3 November 2026. This chapter collects what each chapter pre-registered against itself and adds the referee's case, which the book should make before the referee does. Provenance tags as plan.md §2.