How this book was made

This book was researched, drafted and assembled with AI systems over two days in October 2026. A reader is entitled to know how, because the method shapes what kinds of error the text may contain and how far its numbers can be trusted. This appendix describes the pipeline, gives its counts, and states its limits.

The corpus

The work began the night before, on 15 October 2026, with a survey of the literature on detecting faint moving objects, made for a separate research question: whether an order-of-magnitude faster blind search on one modern GPU was possible at equal sensitivity. That survey screened 13,493 arXiv papers for relevance and downloaded the full LaTeX source of 11,782 of them, so that later readers could search and quote the papers themselves rather than their abstracts. From these, 1,083 papers were read closely and summarised as structured reading cards, each recording the paper’s method, its key numbers and where in the paper they appear. The cards were then combined into 12 theme syntheses, on topics such as stacking, linking, template banks and GPU kernels, and finally into a frontier report on where a large speed-up might be found. That report, its evidence-checked digest and the design studies behind it are the “project” this book refers to, mainly in Chapter 38, Chapter 40 and Chapter 42.

The corpus was strongest on the Solar System and on the analogue fields (pulsars, gravitational waves, radar, space debris) and weakest on exoplanets, because it had been built for the moving-object question.

The research dossiers

On 16 October 2026 the book’s plan was divided into 76 research topics, from “Detection theory for planets” and “Discovery history to 1900” to “Optimal image coaddition”, “The compute of discovery” and “Open problems across the field”. Each topic went to one research worker: an instance of the GPT 6.1 Sol model running in the Codex command-line agent, with access to the local corpus, the arXiv API, NASA ADS, Semantic Scholar, software repositories and the open web. Thirty workers ran at a time on a local subscription pool and the remaining 46 on a relay key. Each took about 13 minutes on average.

Every worker received the same instructions, followed by its own topic and scope. The instructions asked for a dossier in a fixed order: overview, history and people, physics and mathematics, methods with pseudocode and complexity, landmark results, implementations, limitations, a section on where the algorithmic room lay and which limits were physical, connections to other topics, figure ideas with sourced data, and numbered references. They asked for every number to carry its source, down to the section or table where possible. And they set three rules: quote numbers only from sources actually opened; mark the worker’s own calculations “(estimate)”; and never invent a citation, saying so when an arXiv identifier or DOI could not be verified. Each worker also wrote a machine-readable file of the references it used and, optionally, the data series for its figures.

The chapters

The chapters were written by 20 writers, each an instance of Claude Opus 5.5 running headless in Claude Code. Each writer was assigned one or more chapters and the dossiers behind them (the table of contents maps every chapter to its dossiers), and was required to read every assigned dossier in full before writing, and to open the primary sources where a dossier was thin or a claim needed checking. Local full texts made that practical: a writer could check a number against the paper’s LaTeX in seconds.

All writers worked from one style guide. It fixes the reader (a physicist, ML researcher or GPU engineer who is clever but not an expert in the subfield), the voice (connected prose, plain and exact English, no hype), the shape of a chapter (history, physics, methods, landmark results, limitations, and a closing “Where the room is”), the notation of Chapter 1, and the rules on facts: use only numbers that are in the dossiers or verified in a primary source, keep every “(estimate)” qualification, distinguish measured results from forecasts and models, and cite every number that matters. Charts had to be drawn from real, sourced data, written as small data files and rendered by a script; schematic diagrams were written by hand as SVG.

Each chapter passed a linter before it was accepted. The linter renders every equation with KaTeX and reports errors, checks that every figure the text calls for exists and is valid, catches blank lines that would break an inline diagram, rejects links to chapters that do not exist, and flags stray dollar signs. It checks form, not truth.

Claude (Opus 5.5) planned the book, built this site and edited the result. The chapter writers ran in parallel, and the synthesis chapter, this appendix and the preface were written last, after the writer had read the closing section of every other chapter.

What can go wrong

The process has known weaknesses, and the book should be read with them in mind.

The dossiers can contain errors. A research worker can misread a table, attach a number to the wrong configuration, or summarise a paper’s caveat away. Writers checked many claims against the primary sources, but not every one. Where dossiers disagreed and a writer noticed, the text says so; where both were wrong in the same way, it may not.

Numbers are as good as the sources the workers opened. The rule was to quote only what was read. That rule excludes invented numbers, but not numbers that were correctly copied from a source that was itself mistaken, out of date, or a preprint later revised. Several of the most important results discussed here are 2025 and 2026 preprints, and some software documentation changed while the book was being written.

Citations should be checked before reuse. References were to come only from sources a worker or writer had opened, and writers verified arXiv identifiers where they could. Bibliographic details such as volume and page numbers are less reliable than the identifiers, and an occasional reference may point to the right paper with an imperfect title or author list. Anyone citing a result from this book should cite and check the original.

Syntheses can echo one another. The theme syntheses and some dossiers drew on the same earlier reports, so agreement between them is not always independent confirmation. The project’s own evidence digest found exactly this in its design studies, and Chapter 42 discounts them accordingly.

The project’s GPU numbers are first measurements. The tensor-core benchmark in Chapter 40 was run on one A100 at NERSC on 16 October 2026, on synthetic data, with timings that mixed two memory variants of the card and a comparison kernel that was not optimised. Its sensitivity retention on the GPU has not been tested, and nobody else has reproduced it. Every H100 figure in the book is a model, and the design studies behind the frontier chapters are unpublished.

AI writers share blind spots. Models trained on similar text may agree on a mistaken commonplace. The defence the book relies on is the primary source: the link next to each number. Where it matters, follow it.

None of this makes the book unusual among technical surveys, which are also assembled from sources their authors could not all re-derive. What is different is the speed and the scale, and the fact that the method is stated. The book describes the field as of 16 October 2026. Corrections are welcome.