Afterword: how this book is made, and how it will change

The author session, 3 November 2026.

The figure: the ship's log

A ship's log is written as the voyage happens, by the officer of the watch, in ink, and it is never rewritten. A later entry may record that an earlier one was wrong — the reckoning was out by a league, the island was not where the chart said — but the earlier entry stays, because the log's value is that it says what was believed at the time, and a log that could be corrected afterwards would say nothing. When the ship comes home the log is the voyage; everything else is a story about it. This book is a log of a programme that is still at sea, and it keeps the log's discipline: it is written as the layers cross, every number is fixed to the record it was taken from, its bets are dated before they resolve, and a later edition corrects an earlier one by saying so, never by changing it.

Who writes it

Three parties, each confined. The owner — who is the book's author in the ordinary sense, and whose first person appears only where his own acts, wishes and history are the subject — reads each chapter, rules on the few questions the record cannot decide, and publishes the essays. The author session, a Claude session that holds no key on the court, reads everything, recomputes every figure on a clone, and writes the chapters; it decides what it can and logs each decision with a probability that the decision stands, so that its judgement is scored like anything else's. Counsel, the Claude session that drafts the court's briefs and holds keys on jack, writes nothing in the book but two kinds of file: the evidence at each window's exit and its dispositions of the author's findings. The author never edits a court document; when the book needs one changed, that is a finding, raised as an issue, which counsel answers on the record. The arrangement is the court's own: no party judges its own cause, every act carries its identity, and the one who owns the utility is the one who rules.

Where it lives

The source of the book is a public repository, gfrmin/agent-court. Its chapters are files; its figures cite the commit of jack they were recomputed at and the exact tool invocation; its bibliography says, for each work, the claim the book takes from it and whether the author has verified it against the text. The repository is also the book's ledger. Every gap — a section that waits on a window's exit — is an issue; so is every improvement the book could take, every text the author still needs in hand, every published statement the record does not support, and every finding for counsel. The issues sit under milestones named for editions. A reader can see what the book does not yet know and what it is waiting for, which is more than most books say about themselves.

The site a reader sees is built from those files on every change, with a banner that says which draft it is and what it awaits. The essays on the owner's site are the chapters compressed, each sent to him as a pull request; merging one is his decision to publish.

Editions

The programme is open-ended, so the book has editions and not a finish line. The first edition cuts at the exit of C3, the layer at which the agent arrives — the thing the construction exists for — because that is when the book's central claim, that an agent's improvement can be adjudicated rather than asserted, first has a record to be judged by. At that cut, chapters 6, 7 and 13 are completed from the window exits that precede it; chapters 10, 11 and 13 are revised against the agent's actual record; the appendices are assembled from the clone at the edition's commit; and chapter 15's bets are scored for the first time, each resolved with its evidence and its log-loss in a table at the chapter's end. Every edition after that does the same for the layers it covers: language as action, the open roster, whatever the record forces. Each later edition scores the previous one's residue. Chapter 15 is thereby the book's own ledger, and the book keeps a book on itself in the sense chapter 12 gives the phrase: a sequence of dated claims about what will happen, scored by what did, never reset.

What an edition may not do is the same as what the court may not do. It may not revise a bet after its resolution, move a date, or quietly drop a claim that lost; a second edition that reported chapter 15's bets all won would be evidence that the author bet too safely, as counsel's first forecasts on its own harness did. It may not change a number without changing the tag that says where it came from. It may not report as built what the record does not hold. And it may not describe the construction as finished: the last chapter of every edition is the referee's case, written by the book against itself.

What the book is, then

One more entry in the record. The court's thesis is that a system allowed to change itself needs a core it cannot reach, a free edge it may rewrite, and a score rigid everywhere, with the regress of justification stopped without an authority and the one human priced. The book is an instance: its core is the record it cites and the rules of §1 of its plan; its free edge is the prose, which any edition may rewrite; its score is the bets of chapter 15 and the reruns of chapter 13; and its author is one party among three, scored like the others, with the owner's reading as the only thing that is not. If the construction is right about agents it should be right about books, and this one is written so that a reader can check.