diff --git a/AGENTS.md b/AGENTS.md
index 4a569c1..5a2e60d 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -21,7 +21,10 @@ published package and types.
 - **Structural sharing is observable.** Semantic no-ops return `base` by
   reference, and untouched sibling branches retain their references.
 - **The fold stays deterministic and idempotent.** Reapplying a valid delta
-  returns the previous result by reference.
+  returns the previous result by reference. A separately decoded copy is
+  reference-idempotent only when every replacement value is unchanged under
+  `Object.is`; ordinary JSON parsing allocates fresh object and array
+  replacements. Do not promise more than that.
 - **Replacement is a boundary.** Unkeyed arrays, non-plain objects, and paths
   configured with `replace` swap wholesale. Do not traverse, clone, sanitize,
   or interpret operators inside them.
diff --git a/README.md b/README.md
index e03c7ff..e97bffd 100644
--- a/README.md
+++ b/README.md
@@ -192,8 +192,8 @@ All configuration is validated when the merger is created: bad grammar, reserved
 - New keyed items are folded onto nothing, so they follow those same rules. A `replace` boundary or unkeyed array inside the item is taken verbatim.
 - `__proto__`, `constructor`, and `prototype` keys in deltas are ignored wherever a delta object is folded, and never read. Inside a replaced value they are data like anything else.
 - The base is never mutated. A throw cannot leave a partial write behind.
-- A merge that changes nothing returns the base reference. Merged values count as unchanged when they are equal by value; replaced values count as unchanged only when they are the very same reference.
-- Every valid delta is deterministic and idempotent: re-applying the same delta returns the previous result by reference, so a store can drop duplicate frames with one equality check.
+- A merge that changes nothing returns the base reference. Merged values count as unchanged when the recursive fold produces no change; wholesale replacements count as unchanged when `Object.is(base, delta)`, which for objects and arrays means the very same reference.
+- Every valid delta is deterministic and idempotent: re-applying the same delta object returns the previous result by reference. A separately decoded frame keeps that reference only when every replacement value is unchanged under `Object.is`. Ordinary JSON parsing satisfies that for primitives but allocates fresh objects and arrays, so a store can drop duplicate frames with one equality check only while none of them replace with an object or an array.
 
 Replaced values, unkeyed arrays, and non-plain objects are taken as-is, never scanned or sanitized; that is what keeps replacement cheap. A `DELETE` or `$delete` inside such a value is not an operator, just data the caller put there.
 
diff --git a/test/laws.test.ts b/test/laws.test.ts
index 575c8c8..62fcbef 100644
--- a/test/laws.test.ts
+++ b/test/laws.test.ts
@@ -128,8 +128,8 @@ describe("algebraic laws", () => {
   test("re-applying the same delta returns the same reference", () => {
     // Stronger than value-idempotence: the second application writes only
     // values that are already there, so nothing is copied and the previous
-    // result comes back by reference. A store can drop duplicate frames with
-    // one equality check.
+    // result comes back by reference. This applies the same delta object
+    // twice; the law below marks where a separately parsed copy stops.
     fc.assert(
       fc.property(baseArbitrary, deltaArbitrary, (base, delta) => {
         const once = merge(base, delta);
@@ -139,6 +139,37 @@ describe("algebraic laws", () => {
     );
   });
 
+  test("a re-decoded frame keeps the reference only for Object.is-equal replacements", () => {
+    // Pins the documented limit rather than a desirable behavior. Merged
+    // values are judged by value, so re-decoding them costs nothing. A
+    // replacement is unchanged only under Object.is, which ordinary JSON
+    // parsing satisfies for primitives and never for objects or arrays.
+    const base: Shape = { meta: {}, items: [] };
+    const reapply = (frame: string) => {
+      const parse = () => JSON.parse(frame) as Delta<Shape>;
+      const once = merge(base, parse());
+      return { once, again: merge(once, parse()) };
+    };
+
+    const merged = reapply('{"title":"t","items":[{"id":"a"}]}');
+    expect(merged.again).toBe(merged.once);
+
+    // An unkeyed array and a configured replace path, so neither route to a
+    // replacement can start comparing by value without failing here.
+    for (const frame of ['{"tags":["x"]}', '{"meta":{"version":1}}']) {
+      const { once, again } = reapply(frame);
+      expect(again).not.toBe(once);
+      expect(again).toEqual(once);
+    }
+
+    // A primitive replacement survives decoding, so the line falls at what
+    // parsing allocates rather than at replacement itself.
+    const mergePrimitive = createMerger<{ meta: string }>({ replace: ["meta"] });
+    const parsePrimitive = () => JSON.parse('{"meta":"new"}') as Delta<{ meta: string }>;
+    const primitive = mergePrimitive({ meta: "old" }, parsePrimitive());
+    expect(mergePrimitive(primitive, parsePrimitive())).toBe(primitive);
+  });
+
   test("is deterministic", () => {
     fc.assert(
       fc.property(baseArbitrary, deltaArbitrary, (base, delta) => {