diff --git a/AGENTS.md b/AGENTS.md
index 7638d00e..f985d31c 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -118,6 +118,11 @@ Additive-only within `/api/v1`:
 - New features expose capability endpoints for feature detection rather than relying on version
   sniffing. Contract strategy and tooling: issue #135.
 
+Treat this as binding. The one exception: `/api/v1` is not locked yet, so a removal taken before
+lock is in scope — but only when it is recorded in the pre-lock removals table in
+[docs/architecture/v1-scope.md](docs/architecture/v1-scope.md) and ships before the lock. Assume
+any removal not listed there is a mistake.
+
 ## Pull requests
 
 Conventional Commit subjects (`feat(playback): add realtime session hub`). One concern per PR.
diff --git a/docs/architecture/v1-scope.md b/docs/architecture/v1-scope.md
index 46e3d762..04a2b2a3 100644
--- a/docs/architecture/v1-scope.md
+++ b/docs/architecture/v1-scope.md
@@ -18,6 +18,21 @@ When the scope locks, this file becomes the source of truth and will contain:
 Until lock: treat any capability not tracked as `Proposed`/`Locked` on the project as out of scope
 for feature PRs (see the scope gate in `CLAUDE.md`).
 
+## Breaking removals taken before lock
+
+The additive-only rule in item 2 binds at lock. Before then a removal is in scope, and there is no
+amendment to write because the amendment process in item 3 does not exist yet. `CLAUDE.md` states
+the rule without that qualifier, which reads as a contradiction — it is not, but a removal taken
+now has to be recorded here so a reader after lock can tell a deliberate decision from a violation.
+
+Each entry names what goes, why waiting is worse, and the design that decided it. **Every removal
+listed here must have shipped before the scope locks.** One still outstanding at lock loses its
+justification and falls back to the Deprecation/Sunset flow like anything else.
+
+| Removed | Release | Rationale |
+|---|---|---|
+| String `GET`/`PUT`/`DELETE /api/v1/settings…`, the unknown-key extension bag, preference fields on profile/library/series DTOs | Cross-platform settings contract, [design](../superpowers/specs/2026-07-10-cross-platform-user-settings-contract-design.md) | Replaced wholesale by the typed settings contract. Deferring past lock would mean carrying the Deprecation/Sunset surface *and* the untyped key bag — which lets any client invent a production setting the server stores unvalidated — through the deprecation window, which is the exact surface the contract exists to close. |
+
 Feature-detection precedent: clients discover which metadata providers (including the
 built-in NFO provider, #216) apply to a library type via
 `GET /api/v1/libraries/provider-defaults` rather than version sniffing. New capabilities
diff --git a/docs/superpowers/specs/2026-07-10-cross-platform-user-settings-contract-design.md b/docs/superpowers/specs/2026-07-10-cross-platform-user-settings-contract-design.md
index 4aa90b0b..0ec73d25 100644
--- a/docs/superpowers/specs/2026-07-10-cross-platform-user-settings-contract-design.md
+++ b/docs/superpowers/specs/2026-07-10-cross-platform-user-settings-contract-design.md
@@ -521,6 +521,12 @@ the amendment process it describes only exists *after* lock. There is therefore
 write and no exception to request: before lock, removing the legacy settings surface is simply in
 scope.
 
+That argument does not live only here. Reasoning kept in a design doc is invisible to whoever reads
+the policy later and sees a removal that appears to break it, so the removal is recorded in the
+**pre-lock removals** table in `docs/architecture/v1-scope.md`, which is the file that governs it.
+The table also carries the deadline: **this work must ship before the scope locks.** If it has not,
+the justification lapses and the removal goes through Deprecation/Sunset like anything else.
+
 **This is an argument for doing the work now rather than after lock.** After lock, the same removal
 would need the Deprecation/Sunset flow the v1 policy mandates and the codebase already implements
 (`internal/api/handlers/legacy_read_routes.go`), which reintroduces exactly the transitional