chapter eleven

11 Patterns in the wild: Three end-to-end case studies

 

This chapter covers

  • Reading architecture cases without turning them into recipes
  • Binding each claim to its evidence class, workload, and operating point
  • Following Task, State, Authority, and Evidence when a design changes
  • Returning evidence to ASSESS, ROUTE, or SELECT
  • Diagnosing failure at the contract, pattern, seam, parameter, and workload layers
  • Keeping, repairing, recomposing, simplifying, or retiring an architecture decision
  • Operating architecture decisions as a versioned, reopenable ledger

A prototype agent built engineering models from a user's description, creating elements and the connections between them. The developer had added a review loop: generate a change, ask the model to critique it, then revise. But a connection to a nonexistent element needs no such debate; code can check it directly. The developer found that repeated reviews added delay and variability. The self-review loop was removed, with structural checks and at most one repair attempt placed before the write. Names and modeling detail still needed judgment; whether a connection had a valid endpoint did not.

Chapter 10 showed how to choose and test an initial composition. The three cases here follow what happens as a design is used: when results expose a weakness, what should you keep, repair, recompose, simplify, or retire? You will trace each choice from an observed problem through a design change to the evidence needed to judge it.

11.1 How to read an architecture case

11.1.1 Five evidence classes

11.1.2 Five architecture actions

11.1.3 A zero-pattern control: When evidence says stop

11.2 Case 1: From tool calls to a state-preserving execution chain

11.2.1 Bound the completed task

11.2.2 The smallest plausible baseline

11.2.3 Failure one: Fixed steps drift

11.2.4 Failure two: A valid argument loses its origin

11.2.5 The trace localizes the handoff

11.2.6 Three runtime structures for three kinds of truth

11.2.7 Replay the task with the new ownership

11.2.8 Evidence, decision, and reopen conditions

11.3 Case 2: Deciding what a modeling agent should judge

11.3.1 Start with the model behind the picture

11.3.2 Separate the work at its dependencies

11.3.3 Remove a review loop when the check can be computed

11.3.4 Make the write and its result observable

11.3.5 Reuse the division of responsibility

11.4 Case 3: Operating evidence reopens the architecture

11.4.1 One question crosses several boundaries

11.4.2 Two operating mismatches changed the design

11.4.3 Evals make drift visible

11.4.4 Decision and reopen condition

11.5 Cross-case analysis

11.5.1 What the cases teach us