diff --git a/README.md b/README.md
index 1fbea7c..1bebee9 100644
--- a/README.md
+++ b/README.md
@@ -13,10 +13,54 @@ Planned: clang-format, code coverage policies.
 
 ## Sanitizers
 
-ASan, UBSan, LSan, and TSan as independently configurable Bazel `cc_feature`s, with
-ready-to-use `--config=` aliases (`asan`, `ubsan`, `lsan`, `tsan`, `asan_ubsan_lsan`,
-`tsan_ubsan`), suppression files for known false positives, and `target_compatible_with`
-constraints for skipping or restricting tests under specific sanitizers.
+| Config | Sanitizers | Notes |
+|--------|-----------|-------|
+| `--config=asan` | AddressSanitizer | Memory errors, buffer overflows |
+| `--config=ubsan` | UndefinedBehaviorSanitizer | Integer overflow, null deref |
+| `--config=lsan` | LeakSanitizer | Memory leaks |
+| `--config=tsan` | ThreadSanitizer | Data races, deadlocks — cannot combine with ASan/LSan |
+| `--config=asan_ubsan_lsan` | ASan + UBSan + LSan | **Recommended default for CI** |
+| `--config=tsan_ubsan` | TSan + UBSan | Threading + undefined behavior |
+
+## Sanitizer Combination Compatibility
+
+| Combination | Valid? | Notes |
+|---|---|---|
+| ASan + UBSan | ✅ Yes | Standard — included in `--config=asan_ubsan_lsan` |
+| ASan + LSan | ✅ Yes | Included in `--config=asan_ubsan_lsan` |
+| TSan + UBSan | ✅ Yes | Use `--config=tsan_ubsan` |
+| ASan + TSan | ❌ No | Incompatible runtime libraries (`libasan` vs `libtsan`) |
+| LSan + TSan | ❌ No | TSan has built-in leak detection; enabling both causes runtime conflicts |
+
+Invalid combinations are enforced at three layers, strongest first:
+
+1. **Feature level (primary)** — the sanitizer `cc_feature`s declare
+   `mutually_exclusive` categories (`asan_tsan`, `lsan_tsan`), so
+   enabling `score_asan`+`score_tsan` or `score_lsan`+`score_tsan` through the
+   toolchain fails at **analysis time** with an explicit error
+   (`Symbol ...:asan_tsan is provided by all of the following features: score_asan score_tsan`).
+   This protection is intrinsic to feature resolution and applies to every
+   consumer automatically — no extra build-graph dependency required.
+2. **Bazel build target (secondary)** — the
+   `//sanitizers/flags:sanitizer_combination_check` genrule catches the
+   flag-driven path and prints an actionable message. The CI test suite depends
+   on it automatically.
+3. **Compiler level (backstop)** — Clang emits an explicit error (e.g.
+   `error: invalid argument '-fsanitize=address' combined with '-fsanitize=thread'`).
+
+## Usage
+
+### Add Dependency
+
+```python
+bazel_dep(name = "score_cpp_policies")
+```
+
+### Configure Sanitizers
+
+Copy [`sanitizers/sanitizers.bazelrc`](sanitizers/sanitizers.bazelrc) into your repository's `.bazelrc`.
+
+### Run Tests
 
 ```bash
 bazel test --config=asan_ubsan_lsan //...   # recommended default for CI
diff --git a/sanitizers/README.md b/sanitizers/README.md
index 4811611..2280536 100644
--- a/sanitizers/README.md
+++ b/sanitizers/README.md
@@ -77,6 +77,44 @@ library (`libclang_rt.ubsan_cxx`) is linked, which is required for C++ ABI
 error handlers. GCC does not support this flag but also does not need it (GCC
 links UBSan runtime automatically).
 
+#### Mutually exclusive runtimes (primary enforcement)
+
+TSan uses a different runtime library than ASan/LSan, so those pairs cannot be
+enabled together. This incompatibility is modeled directly in the feature layer
+via `cc_mutually_exclusive_category` targets:
+
+| Category | Members |
+|---|---|
+| `asan_tsan` | `score_asan`, `score_tsan` |
+| `lsan_tsan` | `score_lsan`, `score_tsan` |
+
+When a toolchain enables both features in a mutually exclusive category, Bazel
+fails at **analysis time** with an explicit error, e.g.:
+
+```
+Error in configure_features: Symbol
+@score_cpp_policies//sanitizers/features:asan_tsan is provided by all of
+the following features: score_asan score_tsan
+```
+
+Because this lives in the feature definitions themselves, it protects every
+consumer automatically — no dependency on a separate check target is required.
+
+> **Note — two activation surfaces:** Sanitizers are driven through two
+> independent surfaces that this repo keeps in sync via the `--config=` aliases
+> in `sanitizers.bazelrc`:
+>
+> - **Toolchain features** (`score_*`) — carry the actual compiler/linker flags.
+>   The `mutually_exclusive` categories guard this surface: enabling an invalid
+>   pair fails at analysis time with the error shown above.
+> - **`//sanitizers/flags:*` bool_flags** — drive `config_setting`,
+>   `config_setting_group`, and `target_compatible_with` wiring. The
+>   `//sanitizers/flags:sanitizer_combination_check` genrule (see below) guards
+>   this surface as a **secondary check**.
+>
+> Enable sanitizers through the `--config=` aliases, which set both surfaces
+> together, rather than toggling `--features=score_*` or the bool_flags directly.
+
 ### `constraints/` — `target_compatible_with` aliases
 
 Use these in `BUILD` files to skip tests that are incompatible with a sanitizer:
@@ -145,9 +183,14 @@ The following table shows which flag combinations are **supported**:
 | ✓ | | | ✓ | ❌ **Invalid** | ASan+TSan: incompatible runtime libraries |
 | | | ✓ | ✓ | ❌ **Invalid** | LSan+TSan: TSan has built-in leak detection |
 
-Invalid combinations are caught at **build time** by
-`@score_cpp_policies//sanitizers/flags:sanitizer_combination_check` (the CI
-test suite depends on this target automatically via `tests/BUILD.bazel`).
+Invalid combinations are enforced primarily at the **feature level**: the
+`score_asan`/`score_lsan` and `score_tsan` features declare mutually exclusive
+runtime categories, so enabling an invalid pair fails at Bazel **analysis time**
+(see [`features/`](#features--compilerlinker-feature-definitions) above). As a
+secondary guard for the flag-driven path,
+`@score_cpp_policies//sanitizers/flags:sanitizer_combination_check` also catches
+these at build time (the CI test suite depends on this target automatically via
+`tests/BUILD.bazel`).
 
 ### `wrapper.sh` — Runtime environment loader