Skip to content
Contents
Decision AI

    Type to search. Planned chapters match by title only.

    Reference

    Changes

    Chapters opened, revised and corrected.

    This page lists what readers can see change: chapters opened, chapters revised, and corrections. Each entry gives the date and the chapter, and a correction is stated as a correction.

    Value against labels added, October 2026

    Chapter 1.3, From probabilities to actions and Chapter 2.1, Resolving and routing customer complaints now compare a recalibrated endpoint and a fitted classifier on the same labels. Jev was run on the training samples to make this possible.

    Revision pass, October 2026

    Opening summaries were shortened, and several claims were narrowed: calibration, the cost of labels, randomized experiments and which agent policies were run. The complaint chapter now follows the desk from routing to a limit on review time, with money back as a companion section. The agent chapter walks through one recorded run, and the last chapter asks the reader to choose before giving its reasoning.

    Site opened, October 2026

    The site opened with the plan of the booklet. Every chapter page shows the question the chapter will answer and the data it will use.

    Opening chapter drafted, October 2026

    The introduction and Chapter 1.1, Decision AI in applications and agents opened as first drafts. Both are conceptual and report no measured result; the case chapters still show only their questions.

    Review chapters drafted, October 2026

    Chapter 1.2, From evidence to typed questions and Chapter 1.3, From probabilities to actions opened as first drafts, with engine runs on sampled coffee reviews. Their labels are the reviewers' own star ratings; no human rated the reviews. Chapter 1.1, Decision AI in applications and agents gained two animated illustrations.

    Code-review chapter drafted, October 2026

    Chapter 2.2, Allocating review attention to code changes opened as a first draft, using patches written for this booklet for the budget lesson and real agent patches as a benchmark. Chapter 4.1, Comparing decision systems now includes this case.

    Synthesis chapter drafted, October 2026

    Chapter 4.1, Comparing decision systems opened as a short first draft that draws on the cases so far, with a decision guide. It adds no new runs.

    Agent chapter drafted, October 2026

    Chapter 3.1, Deciding when an agent is finished opened as a first draft on tasks written for this booklet, scored at checkpoints against grader rules.

    Opening summaries and comparison added, October 2026

    The three data chapters now open with a short summary and state their limits there. Chapter 1.2, From evidence to typed questions compares training a classifier with asking a decision endpoint. Charts now label the Claude Haiku adapter as a general language model. Wording that read a wide interval as a tie was corrected to say the comparison shows no clear advantage.

    Complaint chapter drafted, October 2026

    Chapter 2.1, Resolving and routing customer complaints opened as a first draft, redesigned without hand labels: routing is scored against the consumer's own product choice and the money-back question against the company's recorded response.

    Restructured and corrected after a metrics review, October 30, 2026

    An independent review of the metrics code changed results in every case chapter, and the book now has four parts and seven chapters; the planned routing, placement and optional cases are off the first edition.

    Corrections. In Chapter 2.1, Resolving and routing customer complaints, the classifier is now fit to the actual complaint mix, which removes Jev's earlier lead on agreement and reverses the raw cost comparison. In Chapter 1.3, From probabilities to actions, intervals on net value were too wide because the known star mix varied across bootstrap draws, and the earlier statement that all model policies earn about the same is withdrawn. In Chapter 1.2, From evidence to typed questions, the wording comparison now uses single-question controls, and the effect of removing the no-problem option is described as forced. In Chapter 2.2, Allocating review attention to code changes, runs that produced no change are excluded, which removes the models' clear lead over sorting by size on real patches. In Chapter 3.1, Deciding when an agent is finished, the grader was amended for false failures on date format and punctuation, and the simple checks no longer read a harness flag.

    Additions. Chapter 3.1, Deciding when an agent is finished now runs each stopping policy to the end of the task. Chapter 1.2, From evidence to typed questions and Chapter 2.1, Resolving and routing customer complaints show what labels buy a classifier, and the complaint chapter adds a review queue with fixed reviewer time. Chapter 4.1, Comparing decision systems was rewritten from these results.

    30 October 2026

    29 October 2026