diff --git a/IDEAS.md b/IDEAS.md
new file mode 100644
index 0000000..6c0ef4d
--- /dev/null
+++ b/IDEAS.md
@@ -0,0 +1,45 @@
+# Ideas
+
+## Declared lazy-boundary enforcement
+
+`import-conflicts` finds modules reached through both runtime-static and dynamic
+paths from the same entrypoint. It cannot identify every module that should be
+lazy: whether a modal, panel, or specialized renderer should load on activation
+depends on product behavior and callsite UX.
+
+Add an opt-in contract for boundaries whose lazy intent is already known:
+
+```ts
+// codescythe-require-lazy -- expensive structured-answer dispatcher
+const StructuredAnswer = Dynamic(async () => import("./StructuredAnswer"));
+```
+
+For each declaration, Codescythe should:
+
+- resolve the dynamic target;
+- find every configured entrypoint that can reach the dynamic importer;
+- fail when any of those entrypoints can also reach the target through a
+  runtime-static path;
+- print every violating entrypoint, static importer, and shortest proof path;
+- ignore type-only imports because they do not affect bundling;
+- fail clearly when the annotated statement no longer contains a supported
+  dynamic import.
+
+This separates discovery from enforcement. Bundle analysis and human judgment
+find worthwhile boundaries. Codescythe keeps those boundaries detached after
+they are declared.
+
+### Non-goals
+
+- Inferring that a UI surface is rarely opened.
+- Deciding whether loading latency is acceptable.
+- Estimating emitted or transferred bundle bytes.
+- Treating every static import of a dynamically imported module as invalid
+  across unrelated entrypoint graphs.
+
+### Open questions
+
+- Should configuration support the same contract for generated callsites?
+- Should one annotation cover every dynamic import in a statement?
+- Should violations support reasoned exceptions, or require narrowing the
+  declared entrypoint scope?