How This Textbook Is Built and Maintained

A book with a continuing editorial record

This textbook brings literature discovery and editorial revision into the same workflow. A new paper can be connected to the exact passage it might change, investigated alongside the surrounding chapter, and presented to an editor with a proposed revision and the arguments for or against it. The published book changes only after human approval and a separate publication step.

The 72 chapters are stored as structured sections with identifiable references. This lets the editorial tools retrieve a passage, follow its citations, and read related chapters before proposing a change. Each investigation starts from a fixed version of the book. Proposed wording, source material, model assessments, and editorial decisions remain linked to that starting point, so an editor can see what was examined and how the proposal developed.

Finding evidence is the start of an investigation

Search agents use literature tools to find potentially relevant studies, guidelines, corrections, and retractions. In the October 2026 book-wide run, every chapter received both a search for recently indexed literature and a historical search without a date restriction. Older evidence can still expose an omission or an inaccurate reference.

An agent is an AI model with a bounded task and tools for carrying it out. Here, those tools let it search for sources, retrieve article text, read chapter sections, and save a structured assessment. Search results must be checked for source identity and relevance to a particular passage. A promising abstract is enough to justify further investigation; it does not establish that the chapter should change.

Article investigations require the source text specified by the case. The system records which text was returned to each agent and checks that supporting quotations occur in that text. Loading a document without reading its contents does not satisfy this check. Required figures or supplements that remain unread must be identified as an evidence gap. These records establish access and traceability; the interpretation still needs review.

Independent judgement before the draft

The multi-agent review separates the writing task from criticism of the proposed change. The most recent book-wide exercise used the following sequence for each selected case:

1. A critic examined the sources and book context and saved an initial assessment without seeing the author's draft.
2. A separate author agent investigated the case and produced a proposed revision, a reason to leave the text unchanged, or a request for more evidence.
3. The critic then read the author's result, returned to the sources, and compared it with its own saved assessment. It had to explain any change in judgement.

Saving the first assessment before exposing the draft is intended to reduce the tendency to accept a plausible answer simply because it is already on the page. The editor can inspect all three assessments, earlier drafts, and the exact proposed before-and-after text. Disagreement and decisions to hold a case are preserved.

These are AI assessments. Separate sessions and different models do not make them independent clinical experts, and agreement between them is not evidence that a recommendation is correct. Human editors remain responsible for judging the evidence and deciding whether the wording belongs in the book.

Pi and the durable agent harness

Pi is a software toolkit for building AI agents, developed by Earendil and open-source contributors. It connects an AI model to the tools it can use and keeps a record of the conversation. In this textbook, those tools let agents search the literature, read sources, and prepare editorial assessments. The project's durable harness surrounds these sessions with saved inputs, tool results, progress, and limits on model calls and spending. “Durable” means that this work is stored beyond a single running process. An interruption can be inspected and, where recovery is supported, continued from recorded state.

The hosted service uses Pi Durable, the framework for running and saving these agent sessions, with Cloudflare Durable Objects, a cloud service that combines application processing with persistent storage. Together they retain each investigation's state. The local operator workflow also saves Pi conversations and checkpoints; it was used for the October book-wide run. Both feed the editorial review workspace, while retaining the provenance of the original investigation.

Recovery needs care. An interrupted request may have an uncertain outcome, and repeating it can consume resources again. The operator workflow records unresolved requests and carries their reserved budget into a continuation. In the completed run, paused work was resumed without changing its earlier records; two later interruptions required inspected continuations. Persistence made that recovery auditable, but did not eliminate the failures.

An editor controls what reaches the reader

The review workspace gathers the source support, reading gaps, model arguments, and proposed wording in one case. Editors can keep a case on hold, reject it, or request reconsideration with a reason. A model's recommendation does not approve the text. References and a publication version must also be prepared before an approved change can be released.

Version checks protect against applying a proposal to a chapter that has changed since the investigation began. Published editions retain a record of their changes. Audio is tied to the chapter version it describes, allowing the reader to avoid being offered an older recording as though it covered newly published text.

Readers contribute through chapter feedback. A report of an unclear passage, a citation problem, or newer evidence can enter editorial review, where it needs the same source checks as other proposed changes.

What has been demonstrated so far

As of 5 October 2026, a bounded literature-search round has completed across all 72 chapters. Eighteen selected cases then received three saved model assessments each. The final assessments recommended 11 revisions, held six cases for further evidence, and recommended no change in one. All 18 cases were delivered to the human review workspace; none was clinically approved or published as part of that run.

The search was not exhaustive. Complete section text was returned for 399 of 432 sections, and 55 chapters had at least one failed or partially failed source query despite completing both required search types. The 18 detailed cases were selected for concrete passage problems and available source text. Their outcomes do not establish a clinical error rate or the quality of all 72 chapters.

The monthly Cloudflare research service was enabled on 5 October 2026. A new research round is scheduled for the first of each month at 06:00 UTC, and the cloud service continues saved work between checks. Incomplete searches and technical failures remain visible in the operational record. The service emails the editor when a completed round has actionable proposals. Scheduling does not approve the first round or any later clinical change: human review and publication remain separate decisions.