refactor(book): frame the three release pipelines as one discovery loop

[bojieli]

Implements item 5 of the structural review — but with the framing corrected.

Stacked on #861 (which stacks on #860). Retargets to main as those merge.

The correction

My original reading was: chapters 3, 7 and 9 each build the same release pipeline with three different vocabularies, so converge the vocabulary. That reading was too coarse, and the correction came from the author.

The three do not use the same mechanism at all:

Chapter	Job	Machinery
7	locate the first error in a failed trajectory, classify it	rubrics + first-error localization
3	turn one piece of evidence into a minimal grounded change	Proposer-Reviewer
9	decide whether a change actually improved the system	boundary + retention sets, canary, rollback

Forcing their vocabularies together would hide exactly the differences that matter.

What they actually share

A position. Each occupies one segment of a single loop:

Evidence (Ch 7): failed trajectory → first error + error class
  → Proposal (Ch 3): evidence → one minimal, reviewable, reversible change
  → Experiment (Ch 9): measure on boundary and retention sets, ship to a canary
  → Feedback: results and new failed trajectories return to the evidence segment

A new Chapter 1 section states this and names the loop after Discovery Loop — the company Jeff Dean, Sanjay Ghemawat, Quoc Le and Oriol Vinyals announced on 5 August 2026.

On the citation. The company has published its mission and no reproducible technical results. The section says so explicitly and cites it for the naming only, not as evidence — which is the same distinction chapter 7 spends a whole chapter making. Citing a ten-day-old announcement as if it were a result would have contradicted the book's own methodology.

Where it raises a level

Putting an Agent system's loop next to a pure research loop yields two constraints, and they are what much of the rest of the book is about:

• The experiment must be grounded in real observation. A research experiment can be a training run; an Agent experiment modifies a system that is currently serving users, so the verdict has to come from the environment's actual state — do the tests pass, the final database state, what the tool returned — not from the model's account of its own behaviour.
• Every experiment must answer both "what did this fix" and "what did this break". A research loop wants the metric to go up; an Agent system must also show that behaviour which was already correct still is. That is why the boundary set and retention set from refactor(book): name the five cross-cutting design patterns once in Chapter 1 #861 exist.

Chapters 3, 7 and 9 each gain one sentence in their chapter summary naming which segment they build, so no chapter restates the loop. (The summary was chosen over an inline anchor because it is the one location that is structurally locatable in all 13 editions.)

Scope

Applied to the Chinese edition and all 12 translations.

Verified: experiment, figure and thought-question inventories still match the Chinese across all 13 editions; chapter 1 has six sections everywhere; no dangling footnotes, missing images, or unbalanced code fences.

Sources for the citation: TechCrunch, 5 Aug 2026 · Jeff Dean's announcement

🤖 Generated with Claude Code