How this was made
How do you write a book like this — seventy-six thousand words, every claim referenced, its figures built from real datasets — and know it is accurate? Not from a single prompt. That is far too much to get right in one go, and far too easy for a mistake to slip through. I built it instead the way you build anything this complex, with a lot of help from AI: step by step, as a system, with a check at every stage. This page is how that system keeps every fact traceable to a real source.
How it was made, step by step
Each step was small enough to get right on its own, and checked before the next began.
- Plan — decided what the book would argue and cover: the thesis, the shape, the edges.
- Research — AI agents gathered the science, and every load-bearing figure went into one master list, sitting next to the source it came from.
- Write — drafted the book chapter by chapter, in one consistent voice.
- Expand — added the sidebars, the epigraphs and the extra detail that give each chapter its texture.
- Draw — hand-built every figure and map in a single style.
- Design — set the colours and the type.
- Build — turned the finished book into this site.
- Publish — checked it all one last time, then shipped it as an ebook, and here.
One rule ran through all eight: if a fact couldn’t be traced to a real source, it was cut, not guessed. Automated checks flag anything broken: a citation that leads nowhere, a cross-reference that points at nothing. Then a second round of AI “re-readers” goes back over the text to catch mistakes that slipped through.
The AI once invented a reference — Walton et al., 2017, a paper that was never written. The checks caught it, and out it came.
The before and after of one real sentence
The AI is a fine first-drafter, and a confident writer of sentences that are subtly wrong, or merely robotic. One of the passes hunts the robotic ones. Here is an AI draft, and the human version that replaced it. Tap between them:
"Not because the tool is reliable — it isn't — but because the process around it is."
"The tool never became trustworthy. The process did."
Same meaning, every fact intact. Pared back to something a person would say.
What the checks can’t catch
This method has a clear limit. At build time the automated checks verify that every citation resolves to a real, findable source. What they cannot check is whether that source says what the sentence claims. That last stretch is human reading, and it is never quite finished. Where the evidence was thin, the book hedges on purpose: a range where the data won’t carry a single number, a claim softened rather than overstated. A fact that could not be traced at all was left out.
You can’t make any one AI output trustworthy. You can make the whole system trustworthy.
The engineering
The method in full, for the technically curious
Everything above, with the machinery shown: version control, the agent swarms, and the automated build that refuses to ship a broken citation. Skip it, or dig in.
The pipeline
The work ran as one long pipeline, from a blank repository to this site. Each stage fed the next, and a single rigour gate ran the length of it: the same discipline applied to a chapter, a figure, and a page of this companion.
The rigour framework
The gate is a set of concrete rules and tools, written down in the repository and reused book to book.
- Version control (git). The manuscript, the research notes, the code that draws the figures: the whole project lives in one repository under git, the version-control system software teams use to record every change as a small, dated, reversible step. Nothing joins the book until I have read exactly what changed, line by line, the way a colleague signs off on code before it merges. Because each step is small and dated, a mistake can be traced to the change that caused it and undone on its own. The history is the audit trail.
- The facts register. A single file,
research/facts.md, holds every load-bearing figure in the book: the age of the volcanoes, the koala's 11.4% of wildlife-hospital admissions, the depth of the giant podzols, each with its source. When prose and register disagree, the register wins. Figures go in here first, then into the writing. - Cite-or-cut. A load-bearing claim is either sourced or gone. There is no third option, and no expert sign-off step: traceable referencing is the whole warrant.
- Four editorial passes. Packaged to run on any document: two lift the prose into the house voice (Lift, Tighten), two make sure it is true (Cite-check, Invention-check). The last is a “Draft No. 4” tightening, after John McPhee.
- The audit-swarm. Every few chapters, a swarm re-reads the prose against the register. Cheap agents flag disagreements, frontier agents adjudicate the flags, and a synthesizer returns a ranked report, catching drift before it compounds.
The four editorial passes
Each of the four passes, shown on a real example from this book: the draft that came in, and what replaced it.
Lift make flat writing sound like the book
Takes a stiff, textbook sentence and warms it into the book's voice without changing a single fact. Rhythm, plain words, a little life.
An estuary is highly productive owing to the accumulation of nutrients and sediment delivered by tidal and riverine processes.
An estuary works because the tide and the river between them deliver a constant subsidy of nutrients and fine sediment.
Tighten cut the machine's tells, find the exact word
Hunts the giveaways an AI leaves: the seesaw phrases, the rule-of-three padding, the same word twice in three lines. Then tightens to the plainest true version.
That granularity isn't fussiness; it's the safety net.
That fine grain is the safety net.
Cite-check every claim has to trace to a real source
Reads the finished text against its own list of sources and flags anything that doesn't resolve: a number with nothing behind it, or a citation to a paper that was never written.
… the Cooloola dune ages, confirmed by optical dating (Walton et al., 2017).
There is no such paper. The AI produced a real-looking citation for dating work other authors did do. Chase it and you hit a dead end. Removed.
Invention-check flag made-up detail about real people
The most dangerous thing an AI does is invent something confident and specific that reads as completely real: a quote, an anecdote, a figure. This pass hunts them down.
As one early naturalist wrote, "we stood and watched a whole world quietly slipping away."
A quotation put in a real person's mouth, with no source anywhere. The AI writes quotes that sound perfect. Unless each one traces to a record, it comes out.
How the swarm works
The writing and the site were built by swarms of AI agents, directed by me and orchestrated as deterministic workflows. Many narrow agents, each with one job, run in a fixed shape. The pattern repeats:
- Fan-out. A phase spawns many agents at once, typically one agent per file, so no two write the same file and there are no conflicts.
- Barrier. The phase waits for all agents before the next begins, so later work (a pathway) can safely reference earlier work (the guides it strings together).
- The gate. A final agent runs the build validator,
validate-content.mjs, which fails if any citation does not resolve to the bibliography, or any cross-reference points at nothing. It is the compiler a book never had.
Models are chosen per job: a voice-forward model for prose that has to sing, a precise model for structured, citation-heavy work. A book chapter, the 26 concepts, the 72 guides, the 103 questions: each was authored as one of these fan-outs, then gated.
The companion, as data
Under the surface, this site is a small, cross-linked knowledge graph, and the gate holds it together. Guides teach concepts; pathways string guides into field walks; questions check them. Every factual edge traces to a citation that must resolve, or the build stops.