diff --git a/changelog.d/maintenance/quality-policies-ws5.md b/changelog.d/maintenance/quality-policies-ws5.md
new file mode 100644
index 0000000000..990c7e7c20
--- /dev/null
+++ b/changelog.d/maintenance/quality-policies-ws5.md
@@ -0,0 +1 @@
+- **Docs**: `QUALITY_GATES.md` now codifies the per-runner test retry policy (Playwright 1 CI retry with trace; Vitest per-test explicit quarantine only; node:test never) with target flake SLOs, and the release-level ratchet-drift rule (combination drift on the pure tip is the release captain's to fix once on the branch — never pushed onto contributor PRs, never rebaselined per-PR)
diff --git a/docs/architecture/QUALITY_GATES.md b/docs/architecture/QUALITY_GATES.md
index 3686b9eef4..c2970af8f9 100644
--- a/docs/architecture/QUALITY_GATES.md
+++ b/docs/architecture/QUALITY_GATES.md
@@ -173,6 +173,31 @@ pending implementation).
 
 ---
 
+## Test Retry Policy (WS5.4, v3.8.49)
+
+Retry is per-runner, never a global blanket — a blanket retry converts real regressions
+into invisible flakes:
+
+| Runner | Policy | Why |
+| --- | --- | --- |
+| Playwright (e2e) | `retries: 1` in CI only, with `trace: on-first-retry` | Browser/network timing is genuinely nondeterministic; one retry with a trace turns a flake into a diagnosable artifact |
+| Vitest | NO global retry. A proven-flaky test gets an explicit per-test retry (visible in the diff, reviewed in PR) | Keeps the quarantine list in the repo, never opaque |
+| node:test (unit) | NO retry, ever | A flaky unit test is a bug in the test — fix it, don't re-roll it |
+
+Target SLOs once flake telemetry lands (WS5.2/5.3): <1% flake rate per test
+("fix now" threshold), ≥95% pass rate per pipeline. Industry reference values —
+recalibrate against our own measurements.
+
+## Release-Level Ratchet Drift (WS5.5, v3.8.49)
+
+When a ratchet (file-size, complexity, eslint warnings) regresses on the PURE release
+tip — i.e. the COMBINATION of merges regressed it, and no single PR reproduces the
+regression on its own branch — the fix belongs to the **release captain, once, on the
+release branch**: prefer extraction/refactor; rebaseline only with the documented
+justification entry. Never push combination drift onto a contributor PR, and never
+rebaseline per-PR (that hides real regressions). Discriminate first: reproduce the
+red against the pure tip in a probe worktree before assuming your PR caused it.
+
 ## Allowlist Policy
 
 Every gate that cannot fail on pre-existing violations uses a frozen allowlist