diff --git a/docs/typed-contract-audit.md b/docs/typed-contract-audit.md
new file mode 100644
index 00000000..3a1454b3
--- /dev/null
+++ b/docs/typed-contract-audit.md
@@ -0,0 +1,89 @@
+# Typed Contract Audit
+
+Date: 2026-03-31
+
+## Goal
+
+Audit the current `Typed*` / `typed_*` surface and evaluate whether we can move to one unified model without prefix/suffix naming and wrapper-style APIs.
+
+## What is actually public today
+
+The *public* `Typed*` API names are concentrated in two areas:
+
+1. `solverforge-core`
+   - `TypedEntityExtractor<S, E>` in `domain/entity_ref.rs`.
+2. `solverforge-solver`
+   - `TypedValueSelector<S, V>` trait.
+   - `StaticTypedValueSelector<S, V>`.
+   - `FromSolutionTypedValueSelector<S, V>`.
+
+Most other `typed` occurrences are comments, test names, filenames, or descriptive prose (e.g. “typed getter”, “fully typed”), not additional public type wrappers.
+
+## Inventory (high-signal)
+
+### Core descriptor boundary
+
+- `EntityExtractor` is the trait-object boundary used by descriptors.
+- `TypedEntityExtractor<S, E>` is a concrete generic implementation used to adapt typed closures into that boundary.
+- Wireframe explicitly documents this as an intentional erasure boundary.
+
+### Solver selector layer
+
+- Module filenames currently encode `typed_`:
+  - `heuristic/selector/typed_value.rs`
+  - `heuristic/selector/typed_move_selector.rs`
+- Public API currently carries “Typed” names for value selectors, while move selector trait is already unified as `MoveSelector`.
+
+### Scoring layer
+
+- Scoring terminology still frequently says “typed” in docs/comments, but the public primary name is already `ScoreDirector` (without a `Typed` prefix).
+
+## Assessment of your claim
+
+Your claim is directionally correct for naming:
+
+- The prefix is mostly historical signaling (“this is the zero-erasure path”) and no longer adds much disambiguation where this has already become the default architecture.
+- The solver already demonstrates the desired end-state in places (`MoveSelector`, `ScoreDirector`) where generic APIs exist without `Typed` prefixes.
+
+However, there is one structural nuance:
+
+- `TypedEntityExtractor` is not cargo-cult inheritance; it is an adapter from strongly typed field accessors to the intentional runtime-erased descriptor boundary (`Box<dyn EntityExtractor>`).
+- That adapter can be renamed to a neutral name, but removing the adapter concept entirely would require changing descriptor storage and runtime polymorphism design.
+
+## Recommended unified naming target
+
+If we pursue cleanup, the target model can be:
+
+- `TypedEntityExtractor` -> `EntityCollectionExtractor` (or `EntityFieldExtractor`).
+- `TypedValueSelector` -> `ValueSelector`.
+- `StaticTypedValueSelector` -> `StaticValueSelector`.
+- `FromSolutionTypedValueSelector` -> `FromSolutionValueSelector`.
+- Module file rename:
+  - `typed_value.rs` -> `value_selector.rs`
+  - `typed_move_selector.rs` -> `move_selector.rs`
+- Method rename:
+  - `iter_typed(...)` -> `iter_values(...)` (or simply `iter(...)` if unambiguous).
+
+This keeps zero-erasure guarantees but removes terminology noise.
+
+## Migration strategy (no compatibility shims)
+
+Because repository policy forbids transitional shims, migration should be done as a single breaking sweep:
+
+1. Rename types/modules/methods in `solverforge-core` and `solverforge-solver`.
+2. Update all internal call sites and tests in the same commit.
+3. Update crate wireframes in the same commit (required by AGENTS policy).
+4. Update facade re-exports and macro-generated paths.
+5. Update README/examples/changelog references.
+
+## Risk summary
+
+- **Low runtime risk**: mostly symbol/file renaming and call-site updates.
+- **Medium integration risk**: downstream users will need to migrate imports/method calls in one release.
+- **High mechanical scope**: many files reference the names directly, especially tests and wireframes.
+
+## Conclusion
+
+- Yes: we can and likely should converge on prefix-free API naming.
+- No: we should not remove the extractor adapter *concept* unless we also redesign descriptor polymorphism.
+- Best next step is an atomic, no-shim rename sweep across core + solver + macros + wireframes.