Failure modes

What goes wrong when agents write a book, which rule prevents it, and how it is detected. Every row is a failure that produces a successful-looking result — a complete, fluent chapter that is wrong.

In the manuscript

Failure	Looks like	Prevented by	Caught by
The same thing explained twice	Two chapters both define a term	Continuity ledger; requires/establishes	lint cross-chapter repetition
Register drift	Chapter 12 does not sound like chapter 2	Style guide with checkable rules	voice-auditor
Structural echo	Every chapter opens identically	Writers not reading each other	Reading the assembled draft
Boundary overlap	Two chapters cover the same ground	"Must not cover" in the brief	continuity-auditor
Unpaid promise	"We return to this in chapter 9"; chapter 9 does not	Promises section of the ledger	continuity-auditor
Contradiction	Chapter 8 contradicts chapter 3's position	Claims section of the ledger	continuity-auditor
Fabricated fact	A confident, specific, invented number	[VERIFY:] rule	Nothing mechanical. A human reads it
Fabricated citation to a real source	Correct path, wrong claim about it	Source register	Nothing mechanical
Locally fluent, globally empty	Paragraphs that read well and advance nothing	Argument stated in the brief	Reading against the brief

The last three are the ones to take seriously, because they survive every automated gate in the system. lint verifies that a source is cited, never that it says what the chapter claims.

In the process

Failure	Why it happens	Rule
The outline ends up in the book	It reads like a useful introduction	The outline never ships
Two documents describe the structure	One was written when the other was inconvenient to edit	One owner per document; fold and delete
A writer reads three chapters "for context"	It feels responsible	Writers read briefs, not drafts
Reviewer edits prose directly	It is faster than writing a finding	Reviewers write findings only
Two agents draft chapters sharing a source	Both were dispatched in one batch	Separate sequential batches
Briefs written thin, "the writer will figure it out"	Planning feels like overhead	The brief replaces the whole book
Status says drafted, file is a stub	Status was set optimistically	lint requires the file at each status

In production

Failure	Consequence	Rule
Cover built from an estimated page count	Spine off-centre; discovered on the proof	Cover built last, from the measured interior
Trim changed after drafting	Code listings wrap in print	Measure fixes the line cap; re-derive on change
Figure distinguished by colour, printed greyscale	The distinction disappears	Graphics style guide constraint, set before drawing
Volatile facts printed undated	The book is wrong within a year	snapshot_date, and a scope decision about carrying them at all
Unresolved [VERIFY:] marker printed	Exactly what it looks like	lint --strict before production

The meta-failure

A clean lint is not a good book.

Every mechanical gate in this system checks that a signal is present, not that the thing behind it is right. A chapter can pass every check while making a confident claim about a system that does not behave that way.

The gates exist to make the mechanical failures cheap to eliminate, so that attention is available for the failures only a reader can catch. They are not a substitute for reading the book, and a pipeline that produces a printed reference work nobody read end to end has skipped the only step that was never optional.