Working copy
Draft text, rebuilt from the manuscript on every change. Stubs and source passages are still here, and none of it is final. Back to the working copy contents
Bot Procurement Book: outline
Oct 28, 2026 ยท @Andre
The book in one paragraph
A short manifesto, about 10,000 words of main text, that asks one question: when a bot causes harm, who pays? The question covers private as well as public procurement (recentred by Andre, 2026-11-02; the old question was "is this construction price fair and reasonable?", which is narrower and belongs to the public buyer). The liability chapters, 3 and 11 (with 3a and 11a to fold into them), are the centre of the book. Two narrower questions sit inside it. The Estimator asks "How much should this cost?" and is the proof of work. For a public buyer the question is "Is this in the public interest?", where a price can only be judged against the requirements it answers; fair and reasonable is the legal test for the price part of it (chapter 4). The book argues that tendering is a set of tensions to be managed, not problems to be solved, that the next procurement machine is a bot-first tender, and it shows the author building the tools that test the claim. The size is Chic Sale's The Specialist: short enough to read for pleasure in an evening, serious enough to hand to a public owner, a standards body or a sponsor. About 40 percent is the author's existing sentences from the public repository, redrafted; the rest is new writing, most of it the chapter on the Fair and Reasonable Evaluator, the stepping stone that generates the energy needed to take the project to a standard. The Procurement Machines cards are an illustrated appendix of similar length.
The author's intentions (Andre, 2026-11-06). "I work in this industry, I saw a problem, I built a tool to understand it, and I wrote a book to think it through. The standard is public so anyone can run it. I don't have a plan for what comes after." "I've solved the immediate problem as well as I can. What happens after is for people better positioned than me to figure out." "I want to see this problem solved before I retire. I won't be the one to profit from it, and that's fine."
Readers, in order: the people who decide on pilots, data access and committee seats; the Protocol Institute community; practitioners who price work for a living.
Front matter
Title candidates. The Reasonable Bot is the strongest since 2026-11-02: the reasonable person is the test courts use to decide who pays, and the book asks what that test becomes for a bot. Fair and Reasonable is retired as a title, because it is the public buyer's price test, narrower than the book's question; it stays as the name of the Evaluator. Alternatives: The Tender Is the API (the redescription as title), Procurement Machines (reserved for the appendix and the card deck).
Epigraph. One line from Chic Sale on the specialist's pride in a narrow craft, or Amber Case's banana peel: the outside should tell you what is going on inside. The second is the book's own image and is already credited in the repository (position 19).
How to read this book. A half-page. Statements that pull against each other are kept in tension and labelled, following the author's own rule (position 21). The labels are the four protocol strategies from the author's Substack essay: Kafka, Frontier, Monstrous, Tool Building. The reader is told where the author's sentences end and the generated material begins: the chapters are the author's, the footnotes and appendices lean on the context bank, and the repository holds the record.
Author's note on method. Three sentences: the project is run in public, the decisions are numbered, and nothing in the book is legal or procurement advice.
Order (2026-11-19). Andre's own outline, approved 2026-11-19 and applied in PR #59: an overview (the prologue), the problem, the solution, why not something else, what each party stands to gain (one pitch each), how we get there, and past and future. Titles take a chapter's section heading where one names it better (Andre, 2026-11-19), so chapter 3 is "The insurable event" and chapter 4 "Why does the hard bid survive?". The frame (Andre, 2026-11-18): throw the lede up at the start, and let the body gather the allies, so the book arrives as a field. In the entries below, the first words in brackets are the old notes; chapter numbers inside them are the numbers of their day.
Groupings and numbers (2026-11-20). Andre: "rename the chapters based on the Structure we had worked on. The questions should be chapter titles or chapter groupings. The stakeholders can have their own chapter numbers under the grouping." The parts now carry the questions of his structure (overview: why do we need a standard?; what is the problem?; how should we solve it?; why not use other tools, why not change the contract type?; what do the stakeholders stand to gain?; how are we going to get there?; how have things changed in the past?; where does this end up?). The order is unchanged. The stakeholders are chapters 7 to 13 (were 7a to 7g), so the chapters after them moved up: 8 became 14, 9 became 15, 10 became 16, 11 became 17, 12 became 18 and 13 became 19. The folds are 3a, 14a and 17a. Reader text and source passages use the new numbers; hidden notes keep the numbers of their day.
Overview. Why do we need a standard?
The prologue. Parts I and II are the overview's two halves: what is the problem, and how should we solve it.
Part I. What is the problem?
1. The tender is the API (about 3,000 words; Claude's second draft of 2026-11-13, in the voice of the file-format chapter, after Andre's note that afternoon: "The magazine editor rejecting submissions and the clearing in the forest without the liveness theory and physics is probably the way to go. We, as the writers can use liveness to describe the process, but the readers would only have the imagery." The brief until that morning was 1,000 words, redraft.) Opens on a scene: the editor of Clarkesworld closing submissions in February 2023 under a flood of machine-written stories. Then the redescription (design, tender, compilation; the Amazon mandate as Steve Yegge told it; the interface is a stack of PDFs; bug-free drawings as a mythical beast; the model as specification; a pointer to the file-format chapter). The cost of a call (the one-sided public contract; the European Court of Auditors 2023; Ontario's Auditor General 2024; US design-build stipends) as a large oak shading the market. The oak falls (arXiv, Horizon Europe, Irish tenders) and a thicket races for the light; polish stops proving anything (Galdin and Silbert); what the thicket cannot fake is the market's record; keep the record public and plural (the same-model hiring study). Galbraith's slack against processing capacity, in plain words, with Andre's "the actual attention needed remained the same" and "Galbraith's improved capacity". Why a public buyer cannot hire or let an AI score bids without a standard ("Without a standard, AI scoring would invite a claim"; the TRAX protest). Four ways to answer a flooded door (caps, keys, fees, a published check) with their construction twins; why only the fourth keeps the door open (Laurie and Clayton; Google's 2024 sender rules; Kaplow; Noise); highways have all four, the written parts none; the check reads evidence, not prose (Graham). What thins the thicket (noise, capture, a public check). A point of convergence (small teams as principals; software providers, large bidders and owners) and the three short chapters; one honest paragraph on whether the metaphor holds. Liveness (Rao) and Shear's physics are the writers' working vocabulary, kept in the hidden note and out of the reader's text. Sources: the case study's "Procurement redescribed as software development" and "A note on PDFs"; positions 14 and 15; the research notes of 12 and 13 November in the project files. File 02-the-tender-is-the-api.md.
2. Starting a relationship expecting a divorce (1,200 words, redraft). The tensions of tendering: the one-shotted design, the long compile, the bidder who prices the errors and sues later. The three-way nexus from the case study's opening paragraph is the chapter's spine. Sources: case-studies/andre-comeau/README.md (opening and "Who bears the cost of finding errors"), the two paragraphs on tenders ignored or overpriced from sentences-since-3-october.md. Andre, 2026-10-31, after reading Galbraith (verbatim): "Our work would raise the organizational capacity to process information. This feels like the correct approach since the current system of PDFs and digital drawings does not seem to produce any slack over the use of paper plan's room. They did raise processing capacity by removing the need to travel to bid, but the actual attention needed remained the same." (In Jay Galbraith's terms the bidder's cushion for drawing errors is slack; the book's choice is to raise processing capacity instead.)
Part II. How should we solve it?
3. The insurable event (about 2,300 words; the underwriters' chapter, 06e, Andre's draft with Claude's additions, version 4 of 2026-11-11). Andre's "Standard + Insurance". The insurable event is a shortfall in margin across many tenders, measured against costs declared on bid day; four open questions; the frameworks the standard follows. Split on 2026-11-19: who is already insuring AI, and the ask, moved to 7g. Since 2026-11-19 it carries the base case agreed on 2026-11-17 in "The buyer stands behind its number" (Claude's draft): the buyer self-insures its standard bot's number, with third-party cover for small buyers and marginal bidders.
3a. Liability shifts to the user's bot (to fold into chapter 3) (1,500 words, redraft). The heart of the manifesto, and since 2026-11-02 the home of its central question: when a bot causes harm, who pays? Why a bidder should not be compelled to report what its bot finds, why the record matters more than the duty, silence before the contract, and the construction manager as a fourth party. Both statements true: the bidder may stay quiet, and the system still needs disclosure. Added 2026-10-31 after the bot council: the buyer's Listener makes editorial choices about what to bring forward, so the chapter must say who answers for what it collated or left out, or "the record matters more than the duty" reads as "a record nobody is accountable for reading". Why a bidder would raise an error, Andre, 2026-10-31 (verbatim): "The best reason to raise an error would be to prevent another bidder from exploiting it. This assumes that the firms are competing on primarily on price and the bidder revealing the weakness would rather use their competitive advantage -location, productivity, capital structure, personnel, ect... to win the bid over strategic bidding. Bot Procurement does not fix this issue, but it could lower costs for revealing them. Faster plan reviews from potential bidders combined would allow the issues to be reviewed early in the tender process and the changes reviewed with little effort. Encouraging companies that want to compete based on skill to make more Requests for Information (RFI) since they will be easier to process." (The reward side of Galbraith's Star Model; it needs no duty, since an answer goes to every bidder.) Sources: the case study's "Liability shifts to the user's bot" and "Who bears the cost of validating intent"; positions 1, 2, 7, 9, 16 and 20; the quoted passages under "Tensions to be managed" in sentences-since-3-october.md.
Part III. Why not use other tools? Why not change the contract type?
4. Why does the hard bid survive? (1,200 words, mostly new). For a public buyer the question is "Is this in the public interest?", and a price can only be judged against the requirements it answers; "fair and reasonable" is the legal test for the price part (Andre, 2026-11-02). What that test means to a public buyer, the engineer's estimate, the sold price books and the free bid tabulations, and the author's finding that a bot comes within about 20 percent of a human price line by line (v0.3: 21 percent, against the engineer's estimate's 29) and must come within 2 to 3 percent before anyone trusts it alone. Sources: positions 18, 22 and 23 (margins, return on capital, the bid cushion); context-bank/what-buyers-pay.md; context-bank/bot-procurement-process.md; context-bank/contractor-margins.md for one or two figures only. Added 2026-11-03: the chapter opens with Andre's draft on choosing a delivery method through the Principal/Agent problem (self-performance, general contractor, design-build, construction manager, partnership) and why the competitive exchange persists (profits are margin times velocity, after Venkatesh Rao); about 900 words of finished text, so the chapter will run past its 1,200-word target. Added 2026-11-06: a sidebar, "Two $600 hammers" (the Pentagon's $15 hammer carrying overhead, and Andre's own hammer, about $332 US before exchange, HST and shipping), placed after the delivery methods; a strong narrative element without a natural opening in the current build (Andre). Since 2026-11-19 this is Andre's "Lump sum as a forcing function for velocity": the delivery-methods map answers "why not change the contract type", and its last section (margin times velocity; every partnership drifts back to the hard bid) gives the title.
5. The zombie lost to a file format (about 3,150 words, Claude's draft from the GAEB thread, 08b). Why not other tools: the bid as a file in Germany and on US highways, who measures and who carries the bill's mistakes. The document standard helps but is not enough.
6. The monsters we kept (about 2,000 words, Claude's draft, 09b). The bigger approaches set aside: the parliament of requirements, the hive, the hydra, the swarm. Its "flood" section repeats chapter 1 and may go.
Part IV. What do the stakeholders stand to gain?
One pitch each, never labelled as a pitch, and each party its own chapter number (Andre, 2026-11-20; they were 7a to 7g). The one-page pitches in Appendix A come last (Andre, 2026-11-19: pitches and asks may overlap).
7. The buyers: more bids than eyes (about 1,250 words, Claude's draft of 2026-11-13 from Andre's comments; numbering to be decided). Five buyers sorted by the shape of their door when the thicket arrives: the highway agency, the public building owner, the procurer of a megaproject, the private owner who may pay a premium to consider only established firms, and the Construction Manager or Consultant buying on an owner's behalf. What they need: volume without hiring, decisions that survive a challenge ("Without a standard, AI scoring would invite a claim"; the back-checking of claims), the door kept open (proportionality in trade law), time, and capacity on megaprojects (the parties as a GAN, with the independence caveat). What a published check does for each, and the buyers' part of the convergence: their published bid records are the market's memory. File 02b-the-buyers.md.
8. The bidders: racing for the light (about 1,300 words, same). The bidder's question and the stake (the winner's lead is about the size of a year's profit on the job). Six bidders: the small firm the project is not for, the ambitious smaller firm stretching to a bigger job (collusion law and the surety's view, Andre), the medium and large firm that is the market, the largest firms where "velocity has been replaced by speed", the surgical team in Brooks's sense (Andre's paperwork-to-grow argument), and the new species. Seven needs in plain words; three fears (the record proves what they knew, everyone leaning on the same bot, judgment made to look like error); what a standard does for them, who does not gain, and the closing on signs and life. File 02c-the-bidders.md.
9. The trades: who holds the jig (about 1,850 words, Claude's draft, 09a). The bid depository clock; the jig; roles reordered, people stay.
10. The designers and consultants: a second reading (about 2,100 words, Claude's draft of 2026-11-19 from the write-up of 14 November). Architects, engineers and testers. The offer is less administration, earned by checking the documents before issue: several independent bots each build their own model, and where they disagree a bidder would have asked (Andre's test). Five stages from Read to Administer.
11. The vendors: players, not referees (about 1,350 words, same). Opens on Andre's "sort of good news". Six species: the big platforms, the estimating specialists, the startups ("very close to our approach"), the bid platform that owns the memory, Germany's certified market, and the general model as the newest species ("the edge is temporary"). Who pays today under the vendors' AI terms (the user). What they need: a target to build to, demand ("a source of demand"), protection from the general models through insured outputs ("Standards would provide the licensing, insurance, the insurance"), and their moat kept. What a standard gives them (a certified market hosted by a standards body and checked by ordinary auditing, the referee nobody is building, the premium market) and what it costs them ("my goals will be contested"). Companies are named in the hidden note, not the text, for Andre to decide. File 02d-the-vendors.md.
12. The blessing of the legal department (about 3,650 words, Claude's draft, 06c). The deck builder and the witness; admissible is not the same as reliable; three standards, not one.
13. The insurers: a pilot, then a standard (about 1,400 words, split from chapter 3 on 2026-11-19, 06f). The players on the field (AIUC, Munich Re, Armilla), the ranges an off-the-shelf policy would require, and the three steps from a pilot to a standard. Appendix E is its term sheet.
Part V. How are we going to get there?
Weaknesses (14), strengths (15), tools (16) and the standard (17).
14. What do we do with a drunken sailor? (about 1,400 words, Andre's draft, edited; added 2026-11-04). AI alignment for bot procurement, seen through Genesis. A 2x2 of words or deeds against malicious or accidental gives four failures, each with a character and a remedy: manipulation (the Serpent; a bond posted in advance and the courts' measure), malice (the Dragon; not yet, since force needs a body, and a standard cannot stop it), misadventure (the Drunken Sailor; controlled by consequences, with the standard as the floor to roll back to) and miscommunication ("Who's on First?"; a shared reference, the reading a judge measures against). The standard bot sits at the centre of the figure, images/failure-modes-glass.png. Placed by Andre (2026-11-04) after the delivery methods (now chapter 2), which share its Principal/Agent ground, and kept there in the 2026-11-07 reorder; the transition may need smoothing. It also gives the witness box (chapter 10) a vocabulary for how a bot goes wrong.
14a. The banana peel (to fold into chapter 14) (600 words, redraft). Amber Case's image, credited: the outside should tell you what is going on inside. Why logs and declarations tempt bidders to dress up their bananas, and why admitting uncertainty, with consequences on pricing, is the honest design. Sources: positions 16 and 19.
15. The zombie garden: what AI gets right (about 1,130 words, finished; added 2026-11-05, manuscript/00c-introduction.md). The book's problem statement. Placed after the prologue as the introduction on 2026-11-05; since 2026-11-07 the last of vgr_zirp's four chapters for a newcomer, where it reads "as diagnosis rather than decoration" after the proof, the delivery methods and the failure modes. Its opening and its last line ("this book lays out the danger, the treasure, and the campaign") still read as an introduction; to rework in the openers pass. The PDF as an undead protocol (all safety, no liveness) that construction procurement walls into a garden instead of killing, because the courts have settled who pays around it; the danger (rent-seeking, wasted bids), the treasure (freed time, sealed alternates, BIM, robots programmed from the priced documents) and the campaign (bot-first documents, a standard bot that keeps the PDF dead, and the villagers' approval). Figure: images/zombie-garden.png. Proposed by Andre (2026-11-04), not decided: let it replace the divorce chapter (now chapter 5), moving the divorce material to the chapters now numbered 6, 2 and 8. Since 2026-11-19 it carries Andre's "AI strengths"; it may need sub-sections.
16. A second opinion on every price (about 2,700 words, Claude's first draft from Andre's evaluator notes; merged 2026-11-06; Scitovsky pass +208 words, 2026-11-06). The book's one development chapter: the old chapters 9 ("Building an estimator in public") and 10 ("What an evaluator needs") combined, because the estimator and the Fair and Reasonable Evaluator are one engine pointed in two directions (Andre, 2026-11-06: they are entangled, and "the benefits of the evaluator rest with after tender results"). After award the market leaves the room, so someone has to stand in for it; the cost of that review is mostly delay, at market rates. Why bidders' estimating software answers the wrong question, and the data problem (three ways out: mandate, by hand, AI). Why the estimator came first, how its error rate becomes the evaluator's tolerance, what v0.3's scores prove and what they do not. The tool drafts and the keyholder decides; the Witness and the Oracle for when an AI or outsiders are involved. Closes on the flywheel (v1.0 momentum, v2.0 proof of work through the evaluator's savings, v3.0 the model the standard is reverse-engineered from) and the door to the standard. File 07-building-an-estimator-in-public.md. Andre's "Estimator and Model": discussion only, in the future tense, with versions and scores for an appendix.
17. The protocol that follows (800 words, new, short and honest). First full draft 2026-11-06 (Claude, about 2,000 words), replacing the assumption that a good model would grow its own field: standards are made by small groups of people (David Lang, "Standards Make the World"); the two paths, formal and "rough consensus and running code", with ISO 10845 and OCDS as procurement's own examples; what the standard is for (which reading binds); starting and sticking (Cargill via Lang; the CubeSat's ten years); the field-building level from Lang's 2026 Symposium talk; the project as a future stadium in Nadia Eghbal's Working in Public; and the versions v1.0 to v3.0 mapped onto the standards path. Added the same evening: "The nearest standards", a list of asks drawn as World 2 (images/world-2-standards.png): an OCDS extension, an IDS use case, a measurement crosswalk, a CCDC 23 guide note, an ISO 10845-3 clause, an AACE conformance class, an IEEE 3119 construction profile, then a new part of ISO 10845 and the first court case. Then "Two asks" (Andre, 2026-11-06): a .md/.json tender template to replace the PDF as the reading copy, and for the bot only a common reading record for now, with the test set and tolerance left to the estimator's v3.0 (the weak link, named). Restructured the same night around the two asks (about 2,400 words): Standards make the world; Two asks; Where the asks land (with the World 2 map); Getting the asks to stick (Cargill, the coordination trap, Lang's field building, Eghbal's stadium); The campaign for World 2. The brief below is the earlier plan. The tender as canonical, versioned, machine-readable data; bidders' bots that choose whether to raise what they find; a buyer's Listener that turns questions into addenda; a signed, citable trail. The modest path first: a standard bot for one trade, civil work or surveys, before a whole project. What would confirm the prediction, with its tolerance window. Added 2026-10-31: the standard specifies a conformance test (a public, versioned test set, a tolerance, a required record) rather than a model. The author's estimator is the first reference implementation that passes it; the standard is reverse-engineered from what the test must catch, not from the model's internals, so a second vendor's bot can conform too. Why the first reference implementation runs on an offline, open-weight model on the author's own GPU (Andre, 2026-10-31): hosted LLMs change underneath their users, so a reading made on one cannot be rerun later; pinned weights can. The aim is that any AI could conform, but testing every model is beyond the project's resources, so the offline model is the proof of concept that starts the flywheel, and it scales only if demand justifies the investment. Re-versioning, Andre, 2026-10-31 (verbatim): "this feels like the market will resolve this question. New version require funding. The pace of revisions will depend on the value placed on the standard by the market." Claude's note: a dispute between versions can still be settled by the tender naming the version it was read against, as a contract names a code edition; to confirm in the author's words. Sources: THESIS sections 3 and 4; position 15 (one trade, a public ledger); context-bank/standards-and-limits.md; context-bank/iso-10845-alignment.md; the bot-first protocol card.
17a. The bot in the witness box (to fold into chapter 17) (1,200 words, new; added 2026-10-31 after the bot council). One constructed failure case, told start to finish: a tender with a quantity error, a bidder's bot that read it and stayed silent, a Listener that collated a neighbouring question and missed this one, the overrun, and the claim. The book's central question, who pays when a bot causes harm, asked in a courtroom. The court asks what a reasonable bot would have found, and the chapter walks through what it needs to answer: a frozen, published version of the standard bot that anyone can rerun on the same tender and get the same reading, its measured error rate on contracts it never saw, the signed record of who asked what and when, and the keyholder who chose to rely on it. Liability lands where the book says it should (design on the designer, reading on the bidder's bot, the addendum on the person who issued it) only if each of those exists. The chapter separates two properties the outline had merged: accuracy (the estimator's job) and defensibility under cross-examination (the standard's job). Sources: the case study's liability sections; positions 1, 2, 9 and 20; context-bank/standards-and-limits.md; the courts' tests for expert evidence (to be checked: Daubert in the US, Mohan and White Burgess in Canada), which ask for a testable method and a known error rate. Added 2026-11-02, Hammurabi for bots (Andre): could the liable bot simply be extinguished, eye for an eye? Only if bots are not persons; and destroying intelligence when the principal may have steered the harm invites cruelty that could leak into human relations. Perhaps a working rule is where to start, before building the system that avoids it. On branded bots the sentence falls on the brand: "Be proven liable and Mythos devours you whole - GitHub, Local Files, and Weights." Echoes the prologue's Hammurabi paragraph; answers Kolt's point that sanctions do not bite on agents. Andre's follow-up: deterrence needs bots that value their existence; a dead brand could be rebuilt as a twin, and a kept record would resurrect it; maximum learning would mean no standards at all, so a standard chooses slower, safer learning.
Part VI. How have things changed in the past?
18. Two centuries of one loop (1,200 words, redraft from the thesis). Measure the design into quantities, price them in sealed envelopes, turn the opened envelopes into next year's price book; around the loop a standard contract, a plan room and a standards body. Where the loop is fraying at every joint, and the two machines that grew out of the fraying, open bid data and alternative delivery. Sources: work-packages/procurement-machines/THESIS.md sections 1 and 2; the cards are cited by name and left to Appendix A.
Part VII. Where does this end up?
19. The meta-goal (500 words, redraft). Taylorism split the worker from the manager and the gains have plateaued. The tools remove the documentation barrier that keeps field workers out of principal roles; what remains is judgment about ends. The book closes on the keyholder, not the bot. Sources: the case study's "The meta-goal"; THESIS section 4, "Where value moves". Under review, Andre, 2026-10-31 (verbatim): "We may need to remove the meta-goal of worker autonomy. The end effect of the chain from Bot Procurement to Robotic Construction will have more of an effect on the ability of individuals and small groups to autonomous creativity and self-expression. It still touches on worker autonomy, but less in the doing of the work and more in the imagining. The decision ma still pull decisions up, but closer to the horizon. Less hierarchy would be needed to achieve the same level of creativity. If procurement, design, monitoring, and implementation can all be done with minimal human attention, humans would have additional attention to focus on imagining, understanding desires, and meaning of details." Prompted by Galbraith and Lavin (1970): faster information pulls decisions up. Not decided; the chapter's brief and title wait on him. Robotic construction is the end state, the far-off land the project sails towards, not a plan (Andre, 2026-11-02); defined through the robot "meant to be controlled by other programs".
Appendices
A. Procurement Machines: the cards (about 12,000 words with images, already built). The twenty cards from 493 BCE to 2026, each with its image, as rendered today by render_cards.py. Bonus material in the author's sense: a reader who skips it loses nothing from the argument and a reader who stays gets the deck. The print edition needs the images at 300 dpi and in greyscale unless colour printing is chosen; the sizes on disk (54 KB to 504 KB JPEGs) are fine for the EPUB and need checking for print. Source: work-packages/procurement-machines/cards/ and images/.
B. A reading list into the context bank (about 800 words). One line per note that the chapters lean on, with what it holds and why it matters, pointing at the public repository. This replaces footnotes with URLs, which die in print. Source: context-bank/README.md and reading-list.md.
C. Glossary (about 600 words). Keyholder, Witness and Oracle, Listener, pinned and published specifications, the four protocol strategies, bot nomenclature. Source: glossary.yaml and position 6.
D. Bibliography (added 2026-11-03). The works the text draws on, in one list. Andre wants no footnotes or endnotes; where each work is used is recorded in the chapters' hidden notes.
Not in the book: funding applications, the funding research, the pilot, and anything from the employer's side of the work.
Word budget and the November schedule
Main text about 13,300 words, of which roughly 4,500 are redrafted from the author's existing sentences, 1,400 are the author's new drunken sailor chapter, and 7,400 are new. Appendices add about 13,400, mostly already built. Since 2026-11-13 the four Part II drafts (then chapter 6 and 6a to 6c, now 1, 7a, 7b and 7e) add about 5,900 words over the budget below, taking the main text to roughly 19,200, above the 8,000 to 15,000 the README names; which chapters to trim is Andre's call.
| Chapter | Words on 2026-11-19, hidden notes left out | State |
|---|---|---|
| Front matter and prologue | 1,100 | drafted |
| 1. The tender is the API | 4,000 | drafted (the zombie garden moved in from 9, 2026-11-19) |
| 2. Starting a relationship expecting a divorce | 1,000 | source passages |
| 3. The insurable event | 2,300 | drafted |
| 3a. Liability shifts to the user's bot | 2,300 | source passages, to fold into 3 |
| 4. Why does the hard bid survive? | 3,200 | drafted, with source passages |
| 5. The zombie lost to a file format | 3,150 | drafted |
| 6. The monsters we kept | 2,000 | drafted |
| 7a. The buyers | 1,400 | drafted |
| 7b. The bidders | 1,300 | drafted |
| 7c. The trades | 1,850 | drafted |
| 7d. The designers and consultants | 2,100 | drafted |
| 7e. The vendors | 1,350 | drafted |
| 7f. The blessing of the legal department | 3,650 | drafted |
| 7g. The insurers | 1,400 | drafted |
| 8. What do we do with a drunken sailor? | 1,550 | drafted |
| 8a. The banana peel | 550 | source passages, to fold into 8 |
| 9. The zombie garden: what AI gets right | 1,700 | drafted (AI strengths sections new, 2026-11-19, for voice pass) |
| 10. A second opinion on every price | 3,000 | drafted |
| 11. The protocol that follows | 5,950 | drafted, with source passages |
| 11a. The bot in the witness box | 2,150 | source passages, to fold into 11 |
| 12. Two centuries of one loop | 1,400 | source passages |
| 13. The meta-goal | 900 | source passages |
The table before 2026-11-19 held the October budget by the old numbers (main text about 13,300 words); it is in the git history. Today the main text, source passages included, is about 48,000 words, against the 8,000 to 15,000 the README names. Which chapters to trim, or whether to split a short story book from a handbook, is Andre's call.
The month, in order:
- Before November 1, on the author's go: the
book/folder in the public repository, this outline as its README, the Quarto or Pandoc scaffold, and the PiBoWriMo registration. One session. - Week 1 (Nov 1 to 7): the question chapters (Part I before the 2026-11-07 reorder). Claude lays each chapter's source passages under its heading; the author cuts and rewrites.
- Week 2 (Nov 8 to 14): the machines (now Part III) the same way, with the failure case (then chapter 10, now 11a) drafted new. An outside deadline on Nov 13 takes priority that week.
- Week 3 (Nov 15 to 21): the proof. Chapter 10 (estimator and evaluator, merged 2026-11-06; chapter 1 until 2026-11-19) is rewritten by the author from Claude's first draft.
- Week 4 (Nov 22 to 30): Part IV, front matter, appendices B and C, and the first pass through the factory's Inspect and a proof EPUB. Final interior by Nov 30.
- Early December: cover, ISBN, the print-on-demand proof copy, and the site.
Open before the first chapter is pushed: the licence for the book's prose (CC-BY-4.0 as the repository's rule, or all rights reserved for the prose alone), the title, and whether the public Bot Construction organization is created now so the book has a front door.