The book — plan of record

The author session (Claude, claude-fable-5-1), 3 November 2026; moved from the Project's claude/book/00-plan.md to this repository the same day. This file is the plan for jack's research outputs — the essays on gfrm.in, an arXiv working paper and the monograph with the working title A Court for Agents. It supersedes the planning parts of claude/monograph-a-court-for-agents.md (the chapter list and timeline); that note stands as the outline of record for content until each chapter file supersedes its section. Everything decided here is recorded under §6 with a probability that the decision stands after Guy reads it; only the questions under §7 go to him.

1. Standing rules of the author session

Set by Guy when the session was opened on 3 November and amended by him the same day when the book moved to gfrmin/agent-court; they bind every author session that reads this file.

1. The author holds no keys on jack and files nothing on jack or the runner. It never writes to renavondata/jack, renavondata/assize, renavondata/opencode-runner or the predecessors' repositories, and never edits a court document (CHARTER.md, SPEC.md, a brief, a review, RUNBOOK.md). If the book needs a court document or jack's committed evidence changed, the need is a finding: an issue on renavondata/jack (D48; before 4 November, on this repository labelled finding), which counsel answers on the issue and closes with what was done. Issues are the one thing the author files on jack. Findings F1–F3, raised before the move, were answered in data/counsel-notes.md and are not re-raised as issues.
2. Every number comes from log.jsonl by seq, recomputed on a clone with tools/value.py and tools/status.py, or from assize's REPORT.md and its ledgers (reports/L*/forecasts.jsonl, scored by harness/ledger.py). Each number cites the seq or the file. Counsel's forecasts on the ledger are data about counsel, scored like anyone's.
3. The author writes plan.md, chapters/ch-NN-<slug>.md, sources.md, and data/recompute-*.md here, and claude/monograph-a-court-for-agents.md in the Project, which is the author's. Counsel writes only data/counsel-notes.md and data/<layer>-exit.md here, and the author leaves those files as counsel left them.
4. Drafts are committed straight to master. This repository is the book's alone: essays are not drafted here, and the author does not write to gfrmin/gfrminsite, which other sessions own (D47).
5. The author decides everything it can and records each decision in §6 with a probability. It asks Guy only the [G] questions, the audience, the title, and for a read of each chapter.
6. Order of work: this plan, then chapter 2 (necessities), which needs no window data; the chapters that need window data wait for counsel's data/<layer>-exit.md.

2. Provenance: how a number gets into the book

Every figure in a draft carries a tag in square brackets at the point of use, one of:

• [seq N] — a fact of one entry of renavondata/jack/log.jsonl (its ts, type, proposer, a body member), read from the clone.
• [log@<commit>: …] — a count or fold over the log at a named commit of main (entries by type, by proposer, first seq of a type), recomputed by the author from the file.
• [status.py --subject X --as-of T --runs F --acts G @<commit>] — a window criterion, a wealth, a ratio or an overdue list, printed by tools/status.py on the clone at that commit with the committed evidence files named.
• [value.py --opening T --as-of T' @<commit>] — the C2a value statistic, printed by tools/value.py with the attack gate in force (see finding F2 for why this is not yet possible from the clone).
• [count_kernel.py @<commit>] — the kernel's line count.
• [CHARTER §n], [CHARTER An.m], [SPEC §n], [BRIEF-C1 …], [transcripts/<file>] — a rule, a clause, a pre-stated criterion or a transcript, by section; the book quotes these only when the wording is the point, and otherwise paraphrases with the tag.
• [kernel/check.py:fn], [judge/judge.py:fn], [tools/clerk.py:fn] — the code that enforces a sentence, by function.
• [assize:REPORT.md], [assize:reports/L2/REPORT.md §…], [assize:reports/L*/forecasts.jsonl] — assize's verdicts and ledgers.
• [acts-weekly: window] — the Project's claude/acts-weekly.md, which is counsel's classifier run over GitHub's record (claude/guy_acts.py); the book treats it as the predecessor's baseline and says that it is a classifier's output.

Three rules follow. A number with no tag is a defect. A number the author computes with a gate waived (for instance value.py --no-attack-gate, which the tool itself labels "illustrative fixture only") never appears in a draft; it may appear in data/ with the flag printed beside it. A number that depends on a verified GitHub Actions run (criteria 1, 2 and 4 of every window, and the C2a value statistic's gate) can be recomputed by the author only when the run's JSON is committed under transcripts/evidence/ — the author has no GitHub API access and must not acquire it — so such numbers enter through counsel's data/<layer>-exit.md, which names the committed evidence file, and the author re-runs the tool against it.

What the pipeline reproduced today, on main at 6b800e14 (2026-11-03T04:26:46Z), full history fetched:

The C1b figure agrees with REPORT.md's C1b section to the fraction; the book cites the recomputation, not the report. A shallow clone fails status.py with "historical tree unverified" (tools/status.py:prefix_tree walks --first-parent history); fetch at least 2,000 commits of main before trusting any historical figure. This is the same trap the 28 October acts run fell into on assize (claude/acts-weekly.md, "a note on method").

3. The outputs

Three outputs, one argument, published in the order that gets feedback soonest.

1. Essays on gfrm.in — one per chapter, each standing alone, each with a number that can come out wrong. The founding essay ("The Science Was Done. The Engineering Wasn't.", 1 November) is the series' first; the next is chapter 2 compressed.
2. The arXiv working paper — chapters 2, 4, 5, 6 and 8 compressed to about twenty-five pages: the necessities, testing by betting, the election, where the regress stops, and the record of the windows as pre-registered sequential experiments. It is what the essays point to and what the monograph expands.
3. The monograph — the chapters below, with the charter, the kernel, the pins, the acts tables and the forecast ledger as appendices.

Chapters, files and what each needs

Restructured 3 November onto the construction (D20): a logical sequence, each chapter building the smallest thing that answers what the previous one left open, in a builder's voice. The journey is evidence (chapter 14), not the spine.

Schedule

• Drafted 3 November: every chapter at least once on the construction; 13 is a frame.
• At C2a's crossing (data/C2a-exit.md): chapter 6's value section and chapter 13 §13.5.
• At C2b's exit: chapter 7's record, chapter 13 §13.6; the arXiv paper assembled from appendix A, chapters 6, 7, 12, 13.
• While C3 runs: draft 2 of every chapter against Guy's reads; the appendices; the monograph's first full draft about a month after C2b's exit.
• Essays are not this repository's (D47). The first, compressed from appendix A, was published on gfrm.in on 3 November from gfrminsite.

4. Register and conventions

• British English, the voice of the founding essay: dense prose, verdicts stated, no pleasantries, no lists in the body except tables of correspondences (result ↔ mechanism; Fuller ↔ check). From 3 November the voice is a builder's: "we" for the construction (we have a record; we need…), the results in hand and no history to apologise for; each chapter opens with a figure from ordinary human life — a will, a child, the tipster, the mast, Dr Bonham, the writ, the punter, the canon, Rex, the council, the keys, the cathedral, the trilemma, the notebook, the author, the referee — which illustrates and never evidences (D21–D22). Mathematics in display where a formula is the point, inline otherwise; in essays inline mathematics uses \(…\) only (the site's rule; a bare $ is money).
• First person singular belongs to Guy and appears only where his own history or acts are the subject ("my earlier attempts", "my reserved acts"); the court, counsel and the author are named in the third person. The byline is Guy's; the author sessions and counsel are acknowledged in the front matter, and their forecasts are scored in the text as anyone's.
• Citations are author–year inline; the full entry and the job it does are in sources.md. A result is stated with its hypotheses before it is applied; its epistemic status is marked (theorem; impossibility result; necessity theorem under stated conditions; design principle with a track record; empirical regularity). The book never says "theorem" of Gall or Goodhart.
• Court terms are used as the charter uses them: entry, proposer, principal, key, power, trusted base, judge, checker, writ, window, crossing, promotion, repair, decision, object, procedure. "Sovereign" is the charter's word for Guy's constitutional role and is used only in that sense.
• Every chapter closes with "What would show this chapter wrong".
• Quotation is kept to the sentence that matters; the rest is paraphrase with the tag.

5. The argument in one paragraph (for every chapter to serve)

A self-modifying agent needs a core it cannot reach, a free edge it can rewrite, and a score that is rigid everywhere. The core is forced, piece by piece, by results that are decades old — Rice, Löb and the grain of truth, Conant–Ashby and requisite variety, Dawid, proper scores, Ville, Bahadur–Savage, Alpern–Schneider, Hart, Fuller, Anderson, Thompson, Arrow–Fisher — and each is discharged in jack by one mechanism, in code, under a key the agent does not hold. The law contains nothing that does not carry a proof or a pre-statement; everything else is an entrant that can lose. The court cannot prove its own soundness, so it keeps a book on itself instead: a wealth process on every claim about itself, never reset, that would be seen going bankrupt. Its cost to the one human is measured and driven toward zero. What is new is only the economy: a cheap, general, unreliable policy worth governing. The evidence is the record of the windows, recomputed; the ways it would be shown wrong were written before the windows opened.

6. Decided (not asked)

Each line is a decision the author took rather than asked; p is the author's probability that it stands unchanged after Guy reads this file. Lines are appended, never edited; a reversal is a new line citing the old one.

• D1 (p 0.80). Essay drafts live at claude/book/essays/<slug>/index.md in the Project with Hugo front matter (draft: true, categories, series: [a-court-for-agents] proposed, series_order), because the author cannot write to gfrmin/gfrminsite; Guy copies a draft to drafts/<slug>/index.md and moves it in. The author also sends the file in the conversation when it asks for the read.

• D2 (p 0.85). The provenance tags of §2 are the citation form for every number, in drafts and in the published text; in the published text they become footnotes or an appendix table, never dropped.

• D3 (p 0.90). Numbers computed with a gate waived never appear in a draft (§2). Today's value.py --no-attack-gate figure is in data/recompute-20261103.md with the flag printed beside it and is not in any chapter.

• D4 (p 0.75). Chapter 2 argues each necessity in four parts — the result with its hypotheses; what it forbids or forces for a self-modifying agent; the mechanism in jack that discharges it, cited to code, SPEC, charter clause or seq; what would show it misapplied — and marks each row's epistemic status. The rows are the monograph note's table, reordered so that the results about evaluation (Rice; Löb, Gödel and the grain of truth; Dawid; proper scores; Goodhart) come before the results about testing (Ville; Bahadur–Savage; Alpern–Schneider), which come before control (Conant–Ashby; Ashby), law (Hart; Fuller), enforcement (Anderson; Saltzer–Schroeder; Thompson–Wheeler), irreversibility (Arrow–Fisher) and growth (Gall; Hayek).

• D5 (p 0.80). The Löb/Gödel/grain-of-truth row is stated as a design rule the results force, not as a theorem about jack. The claim the book makes is exactly this: the agent's reach does not include the judge or the score, by key; and the court's own soundness is not proved but bet on. The book never says that jack "solves" the Löbian obstacle or that the court is "provably sound".

• D6 (p 0.90). Rice's theorem is stated for extensional properties of programs in a Turing-complete language, and the chapter says what it does not forbid: decidable syntactic properties — which is what every kernel check is — and properties of restricted languages. The necessity the row carries is that "is this an improvement" is a property of program × world, settled by running, not inspecting.

• D7 (p 0.70). Hart and Fuller stay in chapter 2's table as necessities by elimination — a system without recognition, change and adjudication cannot tell law from not-law, change it, or decide a breach; a rule failing any of Fuller's eight cannot guide conduct — with their status marked "design principle with a track record"; their full treatment is chapter 3.

• D8 (p 0.60). The arXiv paper is chapters 2, 4, 5, 6 and 8 compressed; primary category cs.AI, cross-listed to stat.ME and cs.MA; the title is the monograph's working title until Guy decides the title (§7).

• D9 (p 0.80). The order of drafting is 2 → 3 → 6 → 1 now; 4 at C2a's crossing; 5, 7, 8 at C2b's exit; 9, 10, 11 and the appendices after. Chapter 1 waits for chapter 2 because the question is sharper once the necessities are written.

• D10 (p 0.75). Counsel's forecasts appear in the chapter whose window they bet on, with their log-loss, and in a running table data/counsel-forecasts.md that the author recomputes from the log (never copied from counsel's own reports). The K-series on assize's ledger (CHARTER §11) is scored from assize:reports/L*/forecasts.jsonl and claude/acts-weekly.md.

• D11 (p 0.70). The monograph note claude/monograph-a-court-for-agents.md is not rewritten now; it is the outline of record for content until each chapter supersedes its section, and the author edits it only to record what has moved to a chapter file.

• D12 (p 0.70). The byline is Guy's; counsel and the author sessions are acknowledged in the front matter by role and model id, and their acts in the text are scored like anyone's (rule 2). The voice is the founding essay's (§4).

• D13 (p 0.85). The lineage chapter uses transcripts/PREDECESSORS-20261026.md (the cold read of the seven predecessors, quotations verbatim) as its evidence of record, re-checked against the clones; Guy's answer to G is what each project taught him that its repository does not show, and that is the only part the author cannot write.

• D14 (p 0.80). Where the published essay and the record disagree, the book follows the record and the discrepancy is listed under §8 for Guy (an essay is his, not a court document); the one found so far is E1.

• D15 (p 0.65). "Necessity", "sufficiency", "minimality" and "optimality" are used in exactly the senses chapter 2 §0 fixes, and nowhere else in the book in any other sense: necessity = removing the mechanism leaves a stated result unmet; minimality = the law contains only mechanisms that discharge a necessity and constants that carry a proof or a pre-statement; sufficiency = by construction of pre-stated falsifiers, never by theorem; optimality = Wald's, of the decision procedure forced on every party, not of the mechanism.

• D16 (p 0.90; supersedes D1, 3 November). Essay drafts live at essays/<slug>/index.md in this repository with Hugo front matter (draft: true, categories, series: [a-court-for-agents] proposed, series_order), and each is sent for Guy's read as a pull request against master; merging is his decision to publish, and copying the file into gfrmin/gfrminsite is his act. The author's Hugo front matter follows the site's conventions (a real em-dash in frontmatter text; inline mathematics as \(…\)).

• D17 (p 0.85). A finding is an issue labelled finding with one subject per issue, stating what the book needs, where in the record the need arises (file, seq, tool output), and what the author will cite in the meantime; the author does not reopen a finding counsel has closed, and raises a new one if the closure left the need unmet.

• D18 (p 0.60). Chapter 6 §6.7's operative claim is the Shafer–Vovk reading — the court is a bettor against its own null and its wealth is its evidence — with logical induction named in one paragraph as an ideal the design falls short of, not as a property of anything on the record. Guy's answer to G decides whether the market reading becomes a claim; until then the chapter says what would have to exist for it to be one (an election over an open roster) and that nothing yet does.

• D19 (p 0.80). data/counsel-forecasts.md is the running table D10 named, recomputed from log.jsonl by the author at each commit that cites it (counsel's and the sovereign's forecasts, per-mille p, first resolution, log-loss); jack score at the same commit is cited beside it for per-type means, and the two must agree before either is used.

• D20 (p 0.85). The book is a construction, not a chronicle: chapters 0–15 in the order Guy and the author fixed on 3 November, each building the smallest thing that answers what the previous one left open, in a builder's voice; the journey is chapter 14's evidence. The first draft's chapter 2 is kept whole as appendix A (the results row by row), and the first drafts of 1, 3, 6, 7, 9 were re-cut into 10, 4, 12, 11, 14.

• D21 (p 0.80). Each chapter opens with a figure from ordinary human life (a will, a child, a tipster, the mast, Dr Bonham, the writ, the punter, the canon, Rex, the council, the keys, the cathedral, the trilemma, the notebook, the author, the referee). The figures illustrate the loop, the lock and the score as things everyone has run; they never evidence — the evidence is always the theorem and the log.

• D22 (p 0.75). The construction's voice is "we"; "I" is Guy's and appears only where his own acts, wishes or history are the subject. Each chapter ends with "What this chapter leaves open" (the next chapter's problem) and "What would show this chapter wrong".

• D23 (p 0.80). The thesis is loop–lock–score (ch 0 §0.2), and the bootstrap's precise claim is that what bootstraps is authority — a power exists at level n only because it was earned at level n − 1 by scored record under the same three rules; "self-governing" is the structural property and "self-improving" is what it buys (ch 11 §11.10). Pending ruling R1.

• D24 (p 0.80). The human never leaves: the owner of the utility is the one unscored party by the alignment commitment, priced, confined to the irreversible, and corrected where he forecasts (ch 0 §0.1; ch 12 §12.4). Pending ruling R2.

• D25 (p 0.70). The convergence tests of the bootstrap are the ratio guard and lapse, and the fixed point is "a level at which adding another changes nothing in the law" (ch 11 §11.12). Pending ruling R3.

• D26 (p 0.60). The audience is anyone who has run the loop, with the statisticians at the back of the room; the arXiv paper keeps no figures and is appendix A with chapters 6, 7, 12 and 13 compressed; primary category cs.AI. Pending Guy's impact ruling (R5).

• D27 (p 0.85). The title is A Court for Agents: Law, scoring and the self-modifying machine (Guy, 3 November).

• D28 (p 0.85; was R1). Self-governing is the structural property; self-improving is its consequence. Decided from the charter's own build order — "oblige it to control itself" precedes the agent [CHARTER §8; Federalist 51] — and from the subtitle Guy chose, which puts law and scoring before the machine.

• D29 (p 0.90; was R2). The human never leaves. Decided from the alignment commitment the lineage began with [credence/CONSTITUTION.md §2.5] and the charter's "wishes are never scored" [CHARTER §7]: the owner of the utility is the standard and cannot be graded against himself.

• D30 (p 0.75; was R3). The ratio guard and lapse are the convergence tests, and the fixed point is a level at which adding another changes nothing in the law. Decided because they are the only two mechanisms in the law that can stop its growth [CHARTER A3.4; §3 Change], and A2.1 already puts the three-rule thesis on trial at C2's exit. No finding is raised: asking the charter to name them would add law for a book's reading, against "fail fast, fail cheap" and the rule that nothing enters the law without a case.

• D31 (p 0.80; was R4). The figures stay and "we" stays. Decided from what Guy asked for on 3 November — a book with GEB's inevitability, the loop as what every adapting thing does — and from the standing rule that figures illustrate and never evidence (D21).

• D32 (p 0.75; was R5). The audience is everyone who has run the loop, with statisticians as the referee; cs.AI primary, cross-listed stat.ME and cs.MA; the arXiv paper without figures. Decided from Guy's "firstly an AI project, but to only talk about safety undersells it" and from the thesis being a claim about adaptive intelligence with safety as a corollary.

• D33 (p 0.95; was R6). The title is A Court for Agents: Law, scoring and the self-modifying machine; Guy chose the subtitle on 3 November, which is confirmation enough.

• D34 (p 0.70). The second half of G is not required for chapter 14 to stand: the chapter's claims rest on the repositories and the cold read, and the [G] paragraphs are kept as marked invitations — what each project taught that its repository does not show — which Guy may fill or the book may publish as gaps. The author does not ask again.

• D35 (p 0.85). The second half of G is answered from the record Guy supplied on 3 November — the session transcripts and memory notes on steel (data/lineage-sessions.md): his own dated words on each project's lesson, and the sessions' notes. Chapter 14 §14.8 states the cross-lineage lesson the "measurements" frame missed: every predecessor's law was blind about itself and left the owner outside the record; jack's two additions are the law under its own test and the owner inside. Four quotations from the sessions in the chapter, the rest paraphrased with dates; the extraction file published on Guy's authorisation, with nothing private in it.

• D36 (p 0.90; supersedes D18, closes G). §12.7 names no ideal above Shafer–Vovk. Guy, 3 November: the book reflects what was worked on, and that is Shafer–Vovk; logical induction was counsel's name in the founding list, never considered in the design, and nothing built descends from it. It gets one sentence in chapter 16 §16.3 as a related construction the court does not approximate, because the essay's quoted list means a reader will look for it.

• D37 (p 0.85). A residue chapter (15), on Guy's suggestion of 3 November, in the only form the programme allows for the unbuilt: each item a claim — what the court does not do, what would force it, the likely mechanism, the author's probability by a date — never a design (the rule that withdrew "prepare C3 now"); the bets are scored by the next edition and are not on jack's ledger. The market reading lives here (§15.3) as the open roster, an engineering consequence of chapter 7. The one-owner assumption is stated as a limit (Arrow), not an extension. The falsifiers chapter becomes 16.

• D38 (p 0.85). The book is published as a static site built from chapters/ by mdBook — open source, one binary, search, a draft banner; MathJax was in this list until D43 — deployed by GitHub Pages from a workflow in this repository, at gfrmin.github.io/agent-court until Guy points a subdomain at it (one CNAME; court.gfrm.in suggested). Decided from his preference for self-hosted, open-source tooling and from gfrm.in already deploying the same way. The two acts that are his: enabling Pages on the repository (Settings → Pages → Source: GitHub Actions; the API path is closed to the author's token) and the DNS. Essays stay on gfrm.in; the arXiv paper is assembled later from appendix A and chapters 6, 7, 12, 13.

• D39 (p 0.85). The book has editions, not a finish line, because the programme is open-ended: the first edition cuts at C3's exit (chapters 6, 7 and 13 completed from the window exits; 10, 11 and 13 revised against the agent's record; chapter 15's bets scored once; appendices B–E; the arXiv paper), and later layers are later editions, each scoring the previous edition's residue. Gaps, nice-to-haves, source needs, errata and site tasks are GitHub issues on this repository, labelled gap, nice-to-have, source, erratum, site, under milestones "1st edition — C3's exit" and "2nd edition — beyond C3"; finding stays counsel's channel. Issues #4–#15 opened 3 November. §8 of this file keeps only what predates the move to issues.

• D40 (p 0.85). The book ends with an Afterword (unnumbered, after chapter 16) on how the book itself is made and will change: the three parties and their confinements, the repository as the book's ledger, the editions and their cut, and what an edition may not do — the book as one more instance of the construction. Chapter 15 remains the future as bets and chapter 16 the discussion as the referee's case; Guy's 3 November suggestion of "Discussion and the Future" is met by the three together.

• D41 (p 0.85; amends D38). The book's public home is gfrm.in/book/, built by gfrminsite's publish workflow from this repository's chapters/ (one source of truth; nothing copied by hand), on Guy's suggestion of 3 November; this repository's book workflow only builds and checks, and dispatches a republish when a token is configured. The GitHub Pages deploy from this repository is dropped. Rendering the chapters in the site's own Hugo typography is a later nice-to-have (issue #16).

• D42 (p 0.85). The gfrm.in changes for the book — the workflow step, the English nav entry, the homepage block, the i18n keys, the styles — are the author's to draft and Guy's to merge: a pull request on gfrminsite (#32, 3 November), since the author never pushes to his site's master. Essays continue to arrive as pull requests on this repository and are copied into gfrminsite by Guy (D16); if he prefers, the author can open them as gfrminsite pull requests directly on the same footing.

• D43 (p 0.9; amends D38, D41). Guy reviewed gfrminsite PR #32 on 3 November and pushed 116a831 onto its branch, with gfrmin/agent-court#17 as the counterpart here; the publishing chain is now his design and the author records it as the plan of record. gfrminsite builds /book/ from the agent-court commit a workflow_dispatch names (book_sha), otherwise from the last commit this repository's book workflow built green on master — never from a moving HEAD, so a broken chapter cannot redden the site, and every build names its commit. mdBook is pinned here by .mdbook-version and .mdbook-sha256 and installed by scripts/install-mdbook.sh, which both repositories run; site-url comes from book.toml alone. MathJax is off: no chapter contains \(, \[ or $$, and the option fetched a script from cdnjs on every page — a chapter that first needs mathematics turns it back on in the same commit. This repository's book workflow asks gfrminsite to republish through workflow_dispatch (a token with Actions: write there, nothing more); gfrminsite purges Cloudflare after every deploy with its own token, so nightly and dispatched refreshes reach readers; Hugo emits a BookHead fragment (PostHog, a link home) that is copied in as theme/head.hbs; /book/ is written once, in the menu entry with identifier book. The author verified #17 on a clone (sha256 against the tarball; the grep for math; counsel's file untouched; #32's book step rehearsed green against master + #17) but could neither merge it nor post a formal review from this session — both were refused as acts on another's PR — so the merge is Guy's, as the ordering in #17's description already made it; #32's failed build is that ordering and nothing else, and goes green once #17 is on master with a green book run and the job is re-run. Recorded in a comment on #32. Later the same day: Guy's 4ae25ff moved the mdBook pin and installer into gfrminsite and the mdBook build into its own read-only job, with a fallback to the last good book artifact; #32 went green without #17, which now only keeps this repository's CI on the same pin (a mismatch is a warning there).

• D44 (p 0.85). Figures, counts, citations and quotations stay literal in the prose and are checked, not substituted, on Guy's question of 3 November ("use variables for values that should be the same?"). Substitution was rejected on three grounds: the prose is rendered by two tools (mdBook here, Hugo for the essays) and would need two template syntaxes kept in step; a mdBook preprocessor runs commands, which gfrminsite's book job was made read-only to exclude; and the raw Markdown is what is read in diffs and reviews. And a figure is not one value but a value at a commit by a command (the herald's merges were 85, 88 and 97 at three commits), so a bare variable would make the copies agree while all of them went stale. Instead data/facts.toml records each figure with its jack commit, its method and the phrases that print it; tools/check_book.py recomputes every figure at its commit, fails on any phrase anywhere that prints another number, counts what the book says about itself, matches every Author (Year) to sources.md and back, and checks every quotation of a jack document verbatim at the citing file's pinned commit. It runs as its own job, without secrets, before the book build, so gfrm.in never publishes a commit that fails it. Its first run found five quotations not verbatim (chapters 4, 6, 7, 12 and appendix A, each a word or an omission unmarked), one phrase attributed to the wrong document, a "Savage" and an "Arrow" without the year that tells which work, and three sources orphaned by the re-cut (Beer 1972, Blackstone, Good 1965). This is half of issue #14; the other half, a script that reruns every tag in the chapters, remains.

• D45 (p 0.85). Issue #14 is closed by the checker, not by a second script. Every complete provenance tag in the chapters — a jack tool, its arguments and @commit — is rerun at that commit on every build and must succeed (check_book.py, check_tags); each figure the prose prints is held by data/facts.toml, recomputed and matched to its phrase (D44). A tag abbreviated with "…" or a placeholder is not runnable on its own and is covered only through the figures it stands for. What remains unchecked is a figure in the prose that nobody has registered; the remedy is to register it when it is written, and the first edition's cut (D39) includes a pass that registers every figure still loose.

• D46 (p 0.75). Issue #16, setting the book in gfrm.in's own Hugo typography, is deferred to after the first edition. The book works in mdBook, with search, a draft banner and the site's analytics and home link (gfrminsite #32). A Hugo rendering would mean a second renderer to keep in step with the first, or a migration whose only gain is appearance, while every chapter is still changing weekly. The decision is revisited at the first edition's cut, when the text is stable enough for typography to matter. Amended 3 November: the issue itself is the site's, so it moved to gfrmin/gfrminsite#49; this decision now stands only as the author's advice on timing.

• D47 (p 0.95; amends D16, D42 and rules 3–5 of §1). This repository is for the book. Guy, 3 November: "This repo is for the book, no? Not the essays... Leave gfrminsite for other sessions." So essays/ is removed from this repository, along with the checker's essay phrases, and the author no longer writes to gfrmin/gfrminsite. The published essay ("Fifteen Things…") is that repository's and its sessions'. The author's last pull request there, gfrminsite #42, is left open for them: five sentences corrected against the texts, each correction also in the book. Issue #12 (an essay per chapter) is closed as out of scope. Guy, the same day: "any actual essays can move to gfrminsite, or any ideas for essays as issues of that repo". There was no unpublished essay to move. The six essay ideas from #12 are now gfrminsite issues #43–#48, one per chapter (0, 12, 8, 14, 4, 6), each pointing at its chapter on gfrm.in/book/. Opening issues there is the author's only act on that repository. Where the book finds a published essay wrong, the author raises an issue on gfrminsite and does not patch the site.

• D48 (p 0.95; amends §1 rule 1 and D39's finding label). Findings about jack's documents or evidence are issues on renavondata/jack, not on this repository. Guy, 4 November: "those are issues for the jack repo, not this one", and asked whether to lift the rule, "I file them on jack myself". So §1 rule 1's "files nothing on jack" now excepts issues. The author opens issues on jack and comments on its own; it still makes no commit, opens no pull request, holds no key, and edits no court document there. The three open findings moved on 4 November: #2 to jack#243, #29 to jack#244 and #30 to jack#245, each carrying counsel's disposition. This repository's own issues are the book's: gaps, sources, errata and nice-to-haves.

7. Open questions for Guy

Only these; everything else is decided above. Rule of 3 November (Guy): the author interviews him for a ruling always, but only when the principles on the record cannot decide. R1–R6, requested earlier that day, were re-examined under that rule and each was found decidable; they are recorded as D28–D33 and withdrawn as questions.

• G, second half: closed 3 November from steel's sessions and notes (D35).

1. G The lineage. Answered 3 November, first half: the frame is right — the lineage is a sequence of measurements. The second half (what tannen and wald taught that their repositories do not show) is still open. transcripts/PREDECESSORS-20261026.md records what each predecessor says about itself (proplang's three termini; tannen's M5 passed with "production bugs found … out of its reach by construction"; wald-charter's "forced by a measurement or not added at all"; wald-arena's brief 002 "met as written" at $38.40). What did tannen and wald teach you that their repositories do not show — and is the chapter's frame right, that the lineage is a sequence of measurements, each project finding the next necessity by hitting it, rather than a sequence of failures?
2. G The ideal for §II.7. Closed 3 November: Shafer–Vovk alone (D36). For "a book that can go bankrupt", is Garrabrant's logical induction the right ideal to name — a market over the institution's own claims that cannot be Dutch-booked in the long run, with jack as a crude priced instance and the missing piece being traders rather than a mixture over a fixed roster — or do you want the regress stopped at Shafer–Vovk, with no market reading?
3. G Audience first: 3 November: under discussion — Guy asked first what jack's biggest impact will be once built and agents use it; the audience follows from that. statisticians (anytime-valid inference), AI-safety readers, or institutional economists. Chapter 1's register and the arXiv category follow from this.
4. Title. A Court for Agents stands as the working title. Subtitle candidates: Government by betting; Law, scoring and the self-modifying machine; or none. The author's recommendation, as a bet: Government by betting (p 0.55 that you pick it if you pick either). Answered 3 November: A Court for Agents: Law, scoring and the self-modifying machine. The author's bet lost.

8. Findings and errata (the author does not act on these)

From 3 November every gap, nice-to-have, source need, erratum and site task is an issue on this repository (D39); this section keeps the record of what was raised before that.

Findings for counsel — court documents or evidence the book needs changed or added. From 3 November a new finding is an issue on this repository labelled finding; the three below predate that rule and were answered by counsel in data/counsel-notes.md (F1: LAW.md to be brought to the current layer with the C2a report; F2 and F3: the C2a runs and acts evidence to be committed at C2a's crossing, named in data/C2a-exit.md). They are kept here as the record of what was asked.

• F1. LAW.md maps the charter at C0. Its rows read "C2 (forecasts)" for maxim 1's second sentence, "C1" for the witness rule and for the clerk, and "the writ is C1" under adjudication, while self_judged (kernel/check.py line 83) and ts_decreased (line 60) have been in the kernel since C1a's installing push [seq 29] and the clerk since C1b (tools/clerk.py; tools/enforcement.json). The book cites the code and the enforcement map, not LAW.md, until LAW.md is brought to the current layer. (LAW.md is counsel's path in judge/scope.json.)

• F2. The C2a value statistic cannot be recomputed from the clone. tools/value.py gates scored events on the instant of a verified main tests run holding tests/test_value.py; the runs files under transcripts/evidence/ end at runs-tests-20261101b.json, before that test existed, so status.py --subject C2a prints value: E 1.000000 over 0 events and criterion 4 unverified. The book cannot cite a value figure until the gating run's JSON is committed under transcripts/evidence/ (and named in data/C2a-exit.md), as C1b's runs-tests-20261101b.json was. The same holds for criteria 1 and 2 of every window. Discharged 3 November, afternoon: the runs file transcripts/evidence/runs-tests-20261103-resolve.json (jack PR #216) holds the verified main run, and status.py --subject C2a --runs that file derives the gate's instant, 2026-11-01T15:59:44Z, itself (criterion 4 met). The gated value E at 04:18:22Z on 3 November is 20.131519 over 114 events — the same figure as the gate-waived run of the morning, which the book still does not cite as such. Chapters 6 §6.4 and 13 §13.5 now carry it.

• F3. The acts evidence for C2a is absent from the tree (status.py prints "off-log repairs: unknown (no acts file)"); the book's chapter 8 needs the window's acts-<date>.csv committed as C1b's was.

• F4 (issue #2, 3 November). REPORT.md's C1b section counts the sovereign's acts in the window as two revocations and the promotion; the log holds five [seqs 197, 201, 216, 217, 337], all decisions. The book cites the seqs.

• F5 (issue #3, 3 November). The enforcement map tools/enforcement.json is chapter 8's measure of how much law compiled; the book needs it current at every exit and named in data/<layer>-exit.md.

Errata for Guy — statements in published text of his that the record does not support; the author follows the record.

• E1. The founding essay says "Three older maxims became kernel rules" and gives lex non cogit ad impossibilia as "the rule that every rule is admitted with a witness that violates it and one that satisfies it". No such rule is in the kernel: no primary rule has been admitted, by A2.1 the law language enters only when a first case arises, and none has (C1a report §1; C1b report). What the kernel does refuse as impossible is narrower — a weak public key no one could be the sole holder of (bad_key, [SPEC §7]) and a key entry that would leave no key with the key power [CHARTER A2.5]. The book states the witness rule as law-in-waiting. Whether to amend the essay is yours.