diff --git a/README.md b/README.md
index b8d0c6b..f05f5cc 100644
--- a/README.md
+++ b/README.md
@@ -18,6 +18,10 @@ management and garbage collection experiments for Boa.
 The current API model investigation for Boa issue #2631 is documented in
 [`notes/gc_api_models.md`](./notes/gc_api_models.md).
 
+The current precise-tracing API redesign proposal (post `gc_allocator`) is
+documented in
+[`notes/precise_tracing_api_plan.md`](./notes/precise_tracing_api_plan.md).
+
 ## Project structure
 
 The current project structure is as follows.
diff --git a/notes/precise_tracing_api_plan.md b/notes/precise_tracing_api_plan.md
new file mode 100644
index 0000000..1b7e0fe
--- /dev/null
+++ b/notes/precise_tracing_api_plan.md
@@ -0,0 +1,229 @@
+# Precise-Tracing API Redesign Proposal (Post `gc_allocator`)
+
+Date: 2026-03-19
+
+## Context
+
+After removing the `Collector: Allocator` experiment (`#54`), the next step is
+to propose a concrete API redesign shape that can be discussed and tested.
+
+This note builds on:
+
+- `notes/gc_api_models.md` (model-family investigation for boa#2631)
+- `docs/boa_gc_api_surface.md` (current Boa-facing GC contract)
+- Tracker issues `#26`, `#27`, `#28`, `#30`
+
+The core target is a precise-tracing API that does not use root/reference
+count arithmetic as the liveness authority.
+
+## Problem statement
+
+Today, `Gc<T>` ergonomics and liveness accounting are tightly coupled to root
+counting. That gives simple usage but makes collector internals harder to
+evolve and reason about.
+
+For redesign work to be useful to Boa, we need an API that:
+
+1. uses tracing as the single source of liveness truth,
+2. keeps weak/ephemeron/finalizer semantics explicit,
+3. remains adoptable against the current Boa-facing API surface.
+
+## Proposed API (draft v0)
+
+This proposal uses an explicit root-table model with scope handles.
+
+### Core types
+
+```rust
+pub struct Gc<T: Trace + ?Sized> {
+    ptr: GcErasedPointer,
+    _marker: core::marker::PhantomData<T>,
+}
+
+pub struct Root<'scope, 'gc, T: Trace + ?Sized> {
+    ptr: Gc<T>,
+    slot: RootSlotId,
+    _scope: core::marker::PhantomData<&'scope Scope<'gc>>,
+}
+
+pub struct WeakGc<T: Trace + ?Sized> {
+    ptr: GcErasedPointer,
+    _marker: core::marker::PhantomData<T>,
+}
+
+pub struct GcContext {
+    /* collector state + root table + weak queues (interior mutability) */
+}
+
+pub struct Scope<'gc> {
+    cx: &'gc GcContext,
+}
+```
+
+### Allocation and rooting
+
+```rust
+impl GcContext {
+    pub fn scope<R>(&self, f: impl for<'gc> FnOnce(Scope<'gc>) -> R) -> R;
+    pub fn collect(&self);
+}
+
+impl<'gc> Scope<'gc> {
+    pub fn alloc<T: Trace + 'static>(&self, value: T) -> Gc<T>;
+    pub fn root<'scope, T: Trace + 'static>(
+        &'scope self,
+        value: &Gc<T>,
+    ) -> Root<'scope, 'gc, T>;
+    pub fn downgrade<T: Trace + 'static>(&self, value: &Gc<T>) -> WeakGc<T>;
+}
+
+impl<'scope, 'gc, T: Trace + ?Sized> Root<'scope, 'gc, T> {
+    pub fn gc(&self) -> &Gc<T>;
+}
+```
+
+Safety note: `Root<'scope, 'gc, T>` is lifetime-branded to the borrowed scope,
+so safe code cannot hold rooted references after the owning scope/context is
+dropped.
+
+Implementation note: `Scope::root(&self, ...)` assumes root-slot registration is
+handled via internal mutability in collector internals, so multiple roots can
+coexist without requiring an exclusive borrow of `Scope`.
+
+Collection note: `collect(&self)` here represents a high-level API shape for a
+single-threaded runtime. Implementations should guard collection phase entry so
+collection cannot race with active mutable interior borrows.
+
+### Pointer identity and casts (parity-preserving)
+
+```rust
+impl<T: Trace + ?Sized> Gc<T> {
+    pub fn ptr_eq<U: Trace + ?Sized>(a: &Gc<T>, b: &Gc<U>) -> bool;
+    pub fn into_raw(this: Gc<T>) -> GcRaw;
+    pub unsafe fn from_raw(raw: GcRaw) -> Gc<T>;
+
+    pub fn downcast<U: Trace + 'static>(this: Gc<T>) -> Option<Gc<U>>;
+    pub unsafe fn cast_unchecked<U: Trace + 'static>(this: Gc<T>) -> Gc<U>;
+    pub unsafe fn cast_ref_unchecked<U: Trace + 'static>(this: &Gc<T>) -> &Gc<U>;
+}
+```
+
+### Weak behavior
+
+```rust
+impl<T: Trace + ?Sized> WeakGc<T> {
+    pub fn new(value: &Gc<T>) -> WeakGc<T>;
+    pub fn upgrade(&self) -> Option<Gc<T>>;
+}
+```
+
+### Runtime helpers
+
+```rust
+impl GcContext {
+    pub fn finalizer_safe(&self) -> bool;
+}
+```
+
+`force_collect()` compatibility can be provided by wiring to
+`MarkSweepGarbageCollector::collect` in integration mode.
+
+## Semantics and invariants
+
+### I1. Tracing is authoritative
+
+Reachability is determined only by tracing from root slots and strong graph
+edges in the same cycle.
+
+### I2. Root slots replace root counts
+
+A value is rooted when at least one root slot references it. Slot lifetime is
+explicit (`Root<'scope, 'gc, T>` drop unregisters slot), and root handles are
+branded by the scope/context lifetime so they cannot outlive the owning GC
+context. No per-object root/refcount math is used for liveness.
+
+### I3. Weak upgrade semantics
+
+`WeakGc::upgrade` succeeds only if the referent is marked live in the current
+collector state.
+
+### I4. Ephemeron semantics
+
+Ephemeron values are traced only when keys are independently reachable via
+strong edges.
+
+### I5. Finalizer ordering
+
+Finalize-before-drop ordering is preserved, and collector teardown runs
+finalizers before destructors for tracked live values.
+
+### I6. Teardown safety
+
+Collector drop does not free values in an order that can cause UAF via
+finalizer-triggered graph activity.
+
+## Feasibility and adoption path
+
+This proposal is intentionally structured in two layers:
+
+1. Collector-native API in Oscars (`GcContext`, `Scope`, `Root<'scope, 'gc, T>`).
+2. Boa-compat layer that preserves current surface where required.
+
+### Boa compatibility mapping
+
+1. `Gc::new(value)`:
+   - compatibility shim calls
+     `with_gc_context(|cx| cx.scope(|s| s.alloc(value)))`.
+2. `WeakGc::new/upgrade`, raw-pointer helpers, and cast helpers:
+   - keep the same signatures.
+3. `force_collect()`:
+   - routed to collector `collect()`.
+4. `finalizer_safe()`:
+   - routed to collector phase state.
+
+This keeps migration incremental and avoids an all-at-once engine rewrite.
+
+## What this proposal deliberately does not include
+
+1. A new incremental/generational/concurrent algorithm.
+2. Full Boa integration in one milestone.
+3. Allocator-framework redesign in the same proposal.
+
+## Validation plan for this API proposal
+
+1. Contract tests:
+   - rooting slot lifetime and misuse resistance,
+   - weak upgrade behavior across collections,
+   - ephemeron key/value reachability behavior.
+2. Safety checks:
+   - Miri on root registration/unregistration paths and raw round-trips.
+3. Compatibility checks:
+   - parity checklist against `docs/boa_gc_api_surface.md`.
+4. Benchmarks:
+   - Boa workloads plus targeted GC stress cases.
+
+## Open review questions
+
+1. Should `Scope<'gc>` be mandatory for all allocations, or should we keep a
+   global-context fallback for Boa compatibility?
+2. Should `Root<'scope, 'gc, T>` be cloneable (multiple slots) or explicitly
+   unique?
+3. Should scope entry be closure-only (`GcContext::scope`) or should we also
+   support an explicit long-lived scope handle for ergonomics/integration?
+4. Is `cast_ref_unchecked` still desirable, or should compatibility rely on
+   value-consuming casts only?
+5. Which minimal Boa integration slice gives the best signal first:
+   pointer API parity, weak semantics parity, or runtime helpers?
+6. Should reading `T` from `Gc<T>` require an explicit scope/root token, so
+   stale handles cannot be dereferenced after collection in safe code?
+
+## Relationship to tracker work
+
+This proposal is intended to feed directly into:
+
+- `#26` Tracking issue for Boa integration
+- `#27` Coverage of `boa_gc` API surface area
+- `#28` Integration into Boa
+- `#30` Benchmark MarkSweepGarbageCollector in Boa with `arena3` allocator
+
+And it is aligned with `#54` (remove `gc_allocator` supertrait direction).