diff --git a/docs/adr/0020-browser-sdk-testing-strategy.md b/docs/adr/0020-browser-sdk-testing-strategy.md
new file mode 100644
index 0000000..2c5ddc7
--- /dev/null
+++ b/docs/adr/0020-browser-sdk-testing-strategy.md
@@ -0,0 +1,95 @@
+# 20. Browser SDK testing strategy
+
+Date: 2026-06-18
+
+## Status
+
+Accepted.
+
+## Context
+
+The TypeScript browser SDK (`clients/passkeys-browser/`) exercises ceremony flows that span several
+cooperating pieces: the WebAuthn browser API, the fetch layer, CBOR/base64url encoding, and the
+pk-auth server's registration and authentication endpoints. There are three distinct ways to test
+this code:
+
+1. **Pure unit tests with mocks** — each component is tested in isolation; collaborators are
+   replaced with mock objects that assert call sites.
+2. **Simulated-environment unit tests** — a lightweight in-memory HTTP server (no real Java process)
+   handles `/auth/**` requests while a Node.js runtime stands in for the browser. Tests drive the
+   full client stack against this fake server, with per-test control over server behavior.
+3. **Integration / end-to-end tests** — the real Java server runs (via Docker or a local process)
+   and a headless Chromium browser executes the WebAuthn flow over the network.
+
+The developer-experience constraint that anchors the comparison is: **a developer must be able to
+open an IDE and run unit tests without unreasonable setup**. "Reasonable" means `npm install`.
+"Unreasonable" means Docker, a running JVM, or a Chrome binary — those are prerequisites for
+integration work, not for exploring a specific client code path.
+
+Pure mocks (option 1) let each component be exercised in total isolation and are excellent for
+corner cases that are hard to trigger otherwise, but they bind tests to implementation details
+(which function was called, in what order) and leave the seams between components untested.
+
+A simulated environment (option 2) replaces only the I/O boundary — the HTTP server and the
+browser WebAuthn API — while running the full SDK stack in Node.js. It tests the integration
+points between components without Docker or a real browser. It is easy to step-through-debug in
+an IDE and straightforward to trigger failure paths (server timeouts, unexpected status codes,
+authenticator errors) that would be cumbersome to produce against a real server.
+
+End-to-end tests (option 3) give the strongest production signal — real crypto, real browser
+security boundaries, real server persistence — but are the most expensive to set up, the slowest
+to run, and the hardest to steer toward a specific code path or failure mode. Debugging requires
+coordinating client and server concurrently.
+
+The project already runs Playwright end-to-end suites under `examples/` (opt-in via
+`PK_RUN_E2E=1`). The question this ADR resolves is what the *primary* unit testing strategy for
+the SDK itself should be and how the three approaches relate to one another.
+
+## Decision
+
+**Prefer the simulated-environment approach (option 2) as the primary unit-testing strategy for
+the browser SDK.** Option 1 (mock-based) is a valid complement for edge cases. Option 3
+(end-to-end) is the existing Playwright suite under `examples/`.
+
+Concretely:
+
+- **Simulated environment** — a lightweight Node.js HTTP server stubs the pk-auth `/auth/**`
+  endpoints in memory. Per-test fixtures can replace any server response (status code, body,
+  timing) to exercise specific client behavior without a running JVM. The WebAuthn browser API
+  (`navigator.credentials.create` / `.get`) is replaced with a fake authenticator that produces
+  structurally valid CBOR/COSE artifacts. No Docker. No Chrome. Tests run with `npm test` from
+  any IDE that can drive `vitest`.
+- **Mock-based unit tests** remain appropriate when a failure mode is difficult to reproduce in
+  the simulated environment (e.g. a network socket close mid-response, a CBOR encoding edge case)
+  or when a very specific internal behavior needs pinning. They are not the default.
+- **End-to-end tests** (`PK_RUN_E2E=1`) provide the integration gate against the real Java server
+  and a real browser. They gate CI but are not expected to cover the full error-path matrix — that
+  is the simulated environment's job.
+
+The simulated environment should remain flexible (tests can inject arbitrary server responses) but
+is intentionally *not* production-hardened (no persistence, no rate limiting, no challenge
+store). Its only contract is behavioral fidelity to the real server's happy path and
+well-defined error responses.
+
+## Consequences
+
+- **Positive — seam coverage without infra cost.** The integration points between the fetch layer,
+  encoding, and ceremony orchestration are exercised in every unit test run; regressions at those
+  seams surface immediately without Docker or a browser.
+- **Positive — controllable failure paths.** Server timeouts, 4xx/5xx responses, and
+  authenticator failures can be injected deterministically, producing test coverage that is
+  impractical to achieve against a real server.
+- **Positive — IDE-friendly.** `npm install` is the only prerequisite; tests run and debug in
+  any IDE with a Node.js adapter.
+- **Positive — loose coupling.** Tests assert observable outcomes (resolved/rejected promises,
+  returned credential objects) rather than internal call sequences, so internal refactors do not
+  require test rewrites.
+- **Negative — the fake server and fake authenticator must stay in sync with the real
+  implementations.** If the Java server changes a response shape, the simulated server must be
+  updated or tests will silently diverge from production. The end-to-end suite is the backstop.
+- **Negative — the Node.js runtime is not a browser.** Subtle browser-specific behaviors (CSP,
+  secure-context enforcement, platform authenticator UX) are not exercised by the simulated
+  environment; only the end-to-end suite can catch those.
+- **Constraint — end-to-end tests remain opt-in** (`PK_RUN_E2E=1`) and are not expected to cover
+  the full failure matrix. They validate the happy path and critical flows against real infra;
+  the unit suite owns error-path breadth.
diff --git a/docs/adr/README.md b/docs/adr/README.md
index a84e29b..1227e29 100644
--- a/docs/adr/README.md
+++ b/docs/adr/README.md
@@ -27,3 +27,4 @@ ADR is added and the prior one is marked `Superseded by NNNN`. See
 | [0017](./0017-sonarqube-cloud-static-analysis.md) | Superseded by [0018](./0018-remove-sonarqube-cloud.md) | 2026-06-11 | SonarQube Cloud for static analysis and coverage tracking |
 | [0018](./0018-remove-sonarqube-cloud.md) | Accepted | 2026-06-11 | Remove SonarQube Cloud; enforce coverage with native JaCoCo line + branch gates |
 | [0019](./0019-crypto-agility-post-quantum-readiness.md) | Accepted | 2026-06-17 | Crypto-agility and post-quantum readiness for passkey algorithms |
+| [0020](./0020-browser-sdk-testing-strategy.md) | Accepted | 2026-06-18 | Browser SDK testing strategy |