diff --git a/CONTRIBUTING.md b/CONTRIBUTING.md
index 7d69176..7c62e64 100644
--- a/CONTRIBUTING.md
+++ b/CONTRIBUTING.md
@@ -67,6 +67,12 @@ before the implementation. The loop is **type → red → green → refactor**:
    `npm run verify` as the definition-of-done gate.
 
 Every test pairs an `expectTypeOf(...)` with an `assert.*`; keep them together.
+Each test body follows **AAA (Arrange–Act–Assert)** with labeled blocks
+separated by a blank line: `// Arrange` sets up the inputs (e.g. the matcher
+factory), `// Act` exercises the subject once from them (not a second
+throwaway call), `// Assert` holds every check — type expectations first,
+runtime assertions last; an empty block drops its label (see
+[development/testing.md § AAA ordering](./development/testing.md#aaa-ordering)).
 Type-first is enforced structurally: `npm test` runs `check:tsc` before the
 test runner, so a wrong type can never be papered over by a passing assertion.
 Per [AGENTS.md § Never do](./AGENTS.md#never-do), reach green honestly — fix the
diff --git a/development/testing.md b/development/testing.md
index c58bbda..06a1986 100644
--- a/development/testing.md
+++ b/development/testing.md
@@ -36,6 +36,33 @@ The runner is `node --test --strip-types "src/**/*.test.ts"` and the tiers are i
 build step, and the runner relies on the `.ts` import-extension convention (see
 [tooling.md](./tooling.md#source-imports-use-ts-extensions)).
 
+## AAA ordering
+
+The rule is in
+[CONTRIBUTING.md § Testing discipline (type-driven)](../CONTRIBUTING.md#testing-discipline-type-driven).
+
+#### Decision (2026-09)
+
+Test bodies read arrange → act → assert: inputs (the factory) set up first, the
+subject exercised once from them, all checks last — types then runtime. The
+blocks are labeled with `// Arrange` / `// Act` / `// Assert` comments and
+separated by a blank line; an empty block drops its label.
+
+#### Why
+
+- Interleaved setup/checks hide what runs vs. what is observed; the eye
+  re-reads the block to find the seams.
+- A factory built mid-test invites a second throwaway call of the subject;
+  arranging it once makes the positive construction and the negative
+  `Parameters<…>` check share one source of truth.
+- Labels make the seams explicit, not inferred — grep-able and reviewable
+  without reading the statements.
+
+#### Rejected
+
+- Unlabeled ordering (bare blank lines): the seams still have to be found by
+  reading; the labels cost nothing.
+
 ## Known issues
 
 - The type-aware linter misidentifies `expectTypeOf()` as a floating promise, so