About › The making

Colophon · method

How this was made

A book has no compiler. Prose cannot be run, so nothing automatically catches the wrong date, the drifted figure, the citation that points nowhere. This project set out to fix that — to treat a natural-history book like a codebase: a single source of truth for every fact, automated audits that fail loudly, and swarms of AI agents held, at every step, to a rule that a claim you cannot trace is a claim you cannot keep.

1 · The pipeline

The work ran as one long pipeline, from a blank repository to this app. Each stage fed the next, and a single rigor gate ran the length of it — the same discipline applied to a chapter, a figure, and a database of wildflowers.

2 · The rigor framework

The gate is not a metaphor. It is a set of concrete rules and tools, adapted from the author's earlier book and sharpened here.

• The facts register. A single file, research/facts.md, holds every load-bearing figure in the book — the salinity of the gulf, the clearance percentages, the fault-slip rates — 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 flagged. There is no third option and no expert sign-off step — traceable referencing is the whole warrant.
• Editorial passes. Three repeatable line-edits: a fact-check pass (find every unsourced claim and broken citation), a voice pass (lift flat prose into the house voice), and a “Draft No. 4” pass (strip AI writing tics and tighten, 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.

3 · How the swarm works

The writing and the app were built by swarms of AI agents, directed by the author and orchestrated as deterministic workflows — not a single model improvising, but many narrow agents run in a controlled shape. The pattern repeats:

• Fan-out. A phase spawns many agents at once — typically one agent per file, so no two ever 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, any deep passage lacks its grounding, 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 61 guides, the 63 questions — each was authored as one of these fan-outs, then gated.

4 · The companion, as data

This app is not a website of pages; it 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 — and every factual edge traces to a citation that must resolve, or the build stops.

Some data was hard-won. The wildflower finder began as a 41-page government nursery catalogue in which the flower colour and flowering season were drawn as tiny icons, not text. An agent reconstructed all 866 species from the PDF’s vector layer — decoding each colour from its exact fill and each season from its silhouette, self-calibrating against the legend and checking its work against rendered pages — then that was filtered to the Adelaide botanical region and cross-checked, park by park, against wildflower records to the 74 common species you can actually go and find.

5 · What the gates do and don’t catch

Honesty about the method matters as much as the method. The build gate guarantees that every citation resolves to a real source — it cannot guarantee that the source says what the sentence claims. That last mile is the audit-swarm and human review, and it is never finished. The book carries deliberately hedged figures where the evidence is thin; the flowering times here are labelled indicative, not verified wild phenology; the terrain map still wants its elevation base. Where a fact could not be traced, the rule held: it was left out rather than invented.

Built by the author directing AI agents (Claude), under the rigor framework above. The full method is documented in the book’s repository — FRAMEWORK.md, the facts register, and the workflow scripts behind every swarm.