diff --git a/docs/decisions/17670-bidi-is-an-implementation-mechanism.md b/docs/decisions/17670-bidi-is-an-implementation-mechanism.md
new file mode 100644
index 0000000000000..637b5b6d2187f
--- /dev/null
+++ b/docs/decisions/17670-bidi-is-an-implementation-mechanism.md
@@ -0,0 +1,89 @@
+# 17670. BiDi is an implementation mechanism, not a public API
+
+- Status: Proposed
+- Discussion: https://github.com/SeleniumHQ/selenium/pull/17670
+
+## Context
+
+Bindings use WebDriver BiDi to implement and extend the Selenium API. This record
+settles BiDi's relationship to the public API: what users program against, and where
+protocol code lives. CDP was managed differently across the bindings, but all of them
+resulted in users needing to understand CDP implementation details to work with it
+to one extent or another. This is happening in BiDi the same way.
+
+No binding conforms yet: each exposes BiDi implementation details through the driver,
+and none treats its BiDi namespace as internal.
+
+| Binding    | Current behavior |
+|------------|------------------|
+| Java       | Driver leaks the raw BiDi connection (`HasBiDi.getBiDi()`) — internal module plumbing surfaced on the driver; protocol types are public |
+| Python     | `driver.network` / `driver.script` are the right high-level names, but return the low-level BiDi modules (`bidi`-namespaced) instead of neutral wrappers, exposing the low-level API |
+| Ruby       | Driver exposes a BiDi accessor (`driver.bidi`) |
+| .NET       | Driver extension returns a BiDi type (`AsBiDiAsync()` → `IBiDi`) |
+| JavaScript | Driver exposes a BiDi accessor (`driver.getBidi()`) |
+
+## Decision
+
+BiDi is an implementation mechanism, not a public API. New capabilities are exposed as
+high-level, protocol-neutral APIs, idiomatic per language, subject to our standard
+deprecation policy. Every binding must:
+
+1. never reference BiDi in its supported API — no BiDi types in signatures, and no driver method or property that exposes BiDi or hands back a BiDi object; and
+2. keep BiDi implementation code (including everything generated from CDDL) visibly internal — marked per language convention as not intended for public consumption.
+
+This relocates low-level access, it does not remove it: the BiDi namespace stays reachable
+by naming it directly (`BiDi::Network.new(driver)`) — technically reachable, but unsupported,
+undocumented, and subject to change without warning — and what is removed is the driver as a
+path into BiDi. Each binding in the table above shows that one fault — the driver offering a
+way in. `HasBiDi.getBiDi()` is the starkest: not access anyone uses, just internal plumbing
+that surfaced on the driver.
+
+Marking a surface *Beta* (or any "stabilizing toward public" status) does not satisfy
+requirement 2: Beta signals an API on its way to becoming supported, which is the opposite
+of internal.
+
+Illustratively (Ruby):
+
+Supported, high-level — `driver.network` is the protocol-neutral API; it returns a neutral
+wrapper, not a BiDi type, and offers no path down to the protocol:
+
+```ruby
+driver.network.add_request_handler(...)
+```
+
+Internal, low-level — reached only by naming the BiDi namespace explicitly; accessible but internal/unsupported:
+
+```ruby
+BiDi::Network.new(driver).add_intercept(phases: [:before_request_sent], url_patterns: [...])
+```
+
+Not allowed — the supported API providing a path to the low-level protocol, whether through a
+dedicated accessor or through the high-level object itself:
+
+```ruby
+driver.bidi.network.add_intercept(...)   # a BiDi accessor on the driver
+driver.network.add_intercept(...)        # driver.network leaking the low-level call
+```
+
+## Considered options
+
+1. Expose the whole protocol as a public API (Rejected)
+   * the protocol is generated from a living CDDL spec and regenerated as the spec changes, so we cannot manage deprecations on it — it can't be committed to as supported, public API
+   * exposing the raw protocol is not user-friendly syntax anyway
+2. A supported, separate public API for conveniently working with the low-level BiDi implementation — a "mid-level" tier beside the high-level API (Rejected)
+   * protocol coupling — users build on BiDi-shaped concepts, the same trap CDP created
+   * mixing a protocol-agnostic high-level API with a BiDi-specific mid-level one leaves users unsure which to reach for
+   * it raises a large set of unanswered questions about which level owns which orchestration functionality
+3. Allow methods and classes to reference BiDi (Rejected)
+   * users shouldn't need to know the underlying implementation mechanism; things should just work without additional ceremony
+4. Internal implementation mechanism only (Accepted)
+
+## Consequences
+
+- Users never need to know which protocol services a command.
+- Each new capability requires API design work across bindings before it ships.
+- Existing public surfaces that fail the conformance tests are brought in line per the
+  deprecation policy, tracked in the ADR implementation-tracking issue linked from this
+  record's PR.
+- How the internal BiDi layer is produced and structured — generated definitions, their
+  separation from orchestration — is decided separately.