diff --git a/compiler/forget/src/HIR/InferReactiveScopes.ts b/compiler/forget/src/HIR/InferReactiveScopes.ts
new file mode 100644
index 0000000000..27af18aff5
--- /dev/null
+++ b/compiler/forget/src/HIR/InferReactiveScopes.ts
@@ -0,0 +1,99 @@
+/**
+ * Copyright (c) Facebook, Inc. and its affiliates.
+ *
+ * This source code is licensed under the MIT license found in the
+ * LICENSE file in the root directory of this source tree.
+ */
+
+import { HIRFunction } from "./HIR";
+
+/**
+ * This is a second (final) stage of constructing reactive scopes. Prior to this pass,
+ * InferReactiveScopeVariables infers the sets of identifiers that "construct together",
+ * assigning each identifier in each scope the same ScopeId and same MutableRange which
+ * describes that span.
+ *
+ * Note that at this point reactive scopes describe ranges based on specific instructions
+ * at arbitrary points in the control flow graph. However, reactive scopes must align
+ * with control-flow boundaries — we can't memoize half of a loop!
+ *
+ * This pass refines the reactive scopes as follows:
+ *
+ * ## Expanding each reactive scope to align with control-flow boundaries
+ *
+ * This corresponds with the shape of the AST: a scope that extends into an if consequent
+ * would expand across the alternate branch, A scope that extends partway into an if
+ * would expand to cover the full loop body, etc.
+ *
+ * ```javascript
+ * function foo(cond, a) {
+ *                    ⌵ original scope
+ *                         ⌵ expanded scope
+ *   const x = [];    ⌝    ⌝
+ *   if (cond) {      ⎮    ⎮
+ *     ...            ⎮    ⎮
+ *     x.push(a);     ⌟    ⎮
+ *     ...                 ⎮
+ *   }                     ⌟
+ * }
+ * ```
+ *
+ * ## Merging (some) overlapping reactive scopes
+ *
+ * Two scopes overlap if there is one or more instruction that is inside the range
+ * of both scopes. In general, overlapping scopes are merged togther. The only
+ * exception to this is when one scope *shadows* another scope. For example:
+ *
+ * ```javascript
+ * function foo(cond, a) {
+ *                                 ⌵ scope for x
+ *   let x = [];                   ⌝
+ *   if (cond) {                   ⎮
+ *                   ⌵ scope for y ⎮
+ *     let y = [];   ⌝             ⎮
+ *     if (b) {      ⎮             ⎮
+ *       y.push(b);  ⌟             ⎮
+ *     }                           ⎮
+ *     x.push(<div>{y}</div>);     ⎮
+ *   }                             ⌟
+ * }
+ * ```
+ *
+ * In this example the two scopes overlap, but mutation of the two scopes is not
+ * interleaved. Specifically within the y scope there are no instructions that
+ * modify any other scope: the inner scope "shadows" the outer one. This category
+ * of overlap does *NOT* merge the scopes together.
+ *
+ * The implementation is inspired by the Rust notion of "stacked borrows". We traverse
+ * the control-flow graph in tree form, at each point keeping track of which scopes are
+ * active. So initially we see
+ *
+ * `let x = []`
+ * active scopes: [x]
+ *
+ * and mark the x scope as active.
+ *
+ * Then we later encounter
+ *
+ * `let y = [];`
+ * active scopes: [x, y]
+ *
+ * Here we first check to see if 'y' is already in the list of active scopes. It isn't,
+ * so we push it to the stop of the stack.
+ *
+ * Then
+ *
+ * `y.push(b)`
+ * active scopes: [x, y]
+ *
+ * Mutates y, so we check if y is the top of the stack. It is, so no merging must occur.
+ *
+ * If instead we saw eg
+ *
+ * `x.push(b)`
+ * active scopes: [x, y]
+ *
+ * Then we would see that 'x' is active, but that it is shadowed. The two scopes would have
+ * to be merged.
+ */
+export function inferReactiveScopes(fn: HIRFunction) {}