diff --git a/.github/workflows/bench.yml b/.github/workflows/bench.yml
index 5aa9a53c8..4dd9c7b12 100644
--- a/.github/workflows/bench.yml
+++ b/.github/workflows/bench.yml
@@ -43,7 +43,7 @@ jobs:
           name: Arm Cortex-A55 (Snapdragon 888) benchmarks
           bench_pmu: PERF
           archflags: "-mcpu=cortex-a55 -march=armv8.2-a"
-          cflags: "-flto -static -DFORCE_AARCH64 -DFIPS202_NATIVE_PROFILE=\"aarch64/profiles/cortex_a55.h\""
+          cflags: "-flto -static -DFORCE_AARCH64 -DMLKEM_NATIVE_FIPS202_BACKEND=\\\\\\\"aarch64/cortex_a55.h\\\\\\\""
           bench_extra_args: -w exec-on-a55
         - system: bpi
           name: Bananapi bpi-f3 benchmarks
diff --git a/examples/bring_your_own_fips202/Makefile b/examples/bring_your_own_fips202/Makefile
index 1a80951e9..e4152b7b0 100644
--- a/examples/bring_your_own_fips202/Makefile
+++ b/examples/bring_your_own_fips202/Makefile
@@ -19,7 +19,6 @@ INC=
 INC+=-Imlkem_native/mlkem
 INC+=-Imlkem_native/mlkem
 INC+=-Imlkem_native/mlkem/native
-INC+=-Imlkem_native/mlkem/sys
 
 # Part B:
 #
diff --git a/examples/bring_your_own_fips202/custom_fips202/fips202.h b/examples/bring_your_own_fips202/custom_fips202/fips202.h
index 5ffe60fae..4bebebeb0 100644
--- a/examples/bring_your_own_fips202/custom_fips202/fips202.h
+++ b/examples/bring_your_own_fips202/custom_fips202/fips202.h
@@ -11,7 +11,7 @@
 #ifndef FIPS202_H
 #define FIPS202_H
 
-#include "namespace.h"
+#include "common.h"
 #include "tiny_sha3/sha3.h"
 
 #define SHAKE128_RATE 168
diff --git a/examples/bring_your_own_fips202/custom_fips202/namespace.h b/examples/bring_your_own_fips202/custom_fips202/namespace.h
deleted file mode 100644
index fab249afa..000000000
--- a/examples/bring_your_own_fips202/custom_fips202/namespace.h
+++ /dev/null
@@ -1,30 +0,0 @@
-/*
- * Copyright (c) 2024 The mlkem-native project authors
- * SPDX-License-Identifier: Apache-2.0
- */
-#ifndef NAMESPACE_H
-#define NAMESPACE_H
-
-#include "config.h"
-
-#if defined(MLKEM_USE_NATIVE_X86_64)
-#define MLKEM_NATIVE_BACKEND X86_64
-#elif defined(MLKEM_USE_NATIVE_AARCH64)
-#define MLKEM_NATIVE_BACKEND AARCH64
-#else
-#define MLKEM_NATIVE_BACKEND C
-#endif
-
-#define ___FIPS202_NAMESPACE(x1, x2, x3) x1##_##x2##_##x3
-#define __FIPS202_NAMESPACE(x1, x2, x3) ___FIPS202_NAMESPACE(x1, x2, x3)
-
-/*
- * NAMESPACE is PQCP_MLKEM_NATIVE_FIPS202_<BACKEND>_
- * e.g., PQCP_MLKEM_NATIVE_FIPS202_AARCH64_
- */
-#define FIPS202_NAMESPACE(s) \
-  __FIPS202_NAMESPACE(PQCP_MLKEM_NATIVE_FIPS202, MLKEM_NATIVE_BACKEND, s)
-#define _FIPS202_NAMESPACE(s) \
-  __FIPS202_NAMESPACE(_PQCP_MLKEM_NATIVE_FIPS202, MLKEM_NATIVE_BACKEND, s)
-
-#endif
diff --git a/examples/mlkem_native_as_code_package/Makefile b/examples/mlkem_native_as_code_package/Makefile
index 1a6aed46a..127afd7d2 100644
--- a/examples/mlkem_native_as_code_package/Makefile
+++ b/examples/mlkem_native_as_code_package/Makefile
@@ -21,7 +21,6 @@ INC+=-Imlkem_native/mlkem
 INC+=-Imlkem_native/mlkem/native
 INC+=-Imlkem_native/fips202
 INC+=-Imlkem_native/fips202/native
-INC+=-Imlkem_native/mlkem/sys
 
 # Part B:
 #
diff --git a/fips202/fips202.h b/fips202/fips202.h
index 48429c271..333a959d7 100644
--- a/fips202/fips202.h
+++ b/fips202/fips202.h
@@ -7,9 +7,8 @@
 
 #include <stddef.h>
 #include <stdint.h>
-#include "namespace.h"
-
 #include "cbmc.h"
+#include "common.h"
 
 #define SHAKE128_RATE 168
 #define SHAKE256_RATE 136
diff --git a/fips202/fips202_backend.h b/fips202/fips202_backend.h
new file mode 100644
index 000000000..24e6593b2
--- /dev/null
+++ b/fips202/fips202_backend.h
@@ -0,0 +1,18 @@
+/*
+ * Copyright (c) 2024 The mlkem-native project authors
+ * SPDX-License-Identifier: Apache-2.0
+ */
+
+#ifdef MLKEM_NATIVE_FIPS202_IMPL_H
+#error Only one FIPS202 assembly profile can be defined -- did you include multiple profiles?
+#else
+#define MLKEM_NATIVE_FIPS202_IMPL_H
+
+/* Include to enforce consistency of API and implementation */
+#include "native/api.h"
+
+#if defined(MLKEM_NATIVE_FIPS202_BACKEND_IMPL)
+#include MLKEM_NATIVE_FIPS202_BACKEND_IMPL
+#endif
+
+#endif /* MLKEM_NATIVE_FIPS202_IMPL_H */
diff --git a/fips202/fips202x4.h b/fips202/fips202x4.h
index 33a732a88..1bb0d08c7 100644
--- a/fips202/fips202x4.h
+++ b/fips202/fips202x4.h
@@ -7,9 +7,9 @@
 
 #include <stddef.h>
 #include <stdint.h>
+#include "common.h"
 #include "fips202.h"
 #include "keccakf1600.h"
-#include "namespace.h"
 
 #include "cbmc.h"
 
diff --git a/fips202/keccakf1600.c b/fips202/keccakf1600.c
index 7b653fca9..5049999dd 100644
--- a/fips202/keccakf1600.c
+++ b/fips202/keccakf1600.c
@@ -15,7 +15,7 @@
 #include <stdint.h>
 
 #include "config.h"
-#include "fips202_native.h"
+#include "fips202_backend.h"
 
 #include "cbmc.h"
 
diff --git a/fips202/keccakf1600.h b/fips202/keccakf1600.h
index 7341df0bf..3d59e0b19 100644
--- a/fips202/keccakf1600.h
+++ b/fips202/keccakf1600.h
@@ -6,10 +6,8 @@
 #define KECCAKF1600_H
 
 #include <stdint.h>
-#include "fips202_native.h"
-#include "namespace.h"
-
 #include "cbmc.h"
+#include "common.h"
 #define KECCAK_LANES 25
 
 /*
diff --git a/fips202/namespace.h b/fips202/namespace.h
deleted file mode 100644
index fab249afa..000000000
--- a/fips202/namespace.h
+++ /dev/null
@@ -1,30 +0,0 @@
-/*
- * Copyright (c) 2024 The mlkem-native project authors
- * SPDX-License-Identifier: Apache-2.0
- */
-#ifndef NAMESPACE_H
-#define NAMESPACE_H
-
-#include "config.h"
-
-#if defined(MLKEM_USE_NATIVE_X86_64)
-#define MLKEM_NATIVE_BACKEND X86_64
-#elif defined(MLKEM_USE_NATIVE_AARCH64)
-#define MLKEM_NATIVE_BACKEND AARCH64
-#else
-#define MLKEM_NATIVE_BACKEND C
-#endif
-
-#define ___FIPS202_NAMESPACE(x1, x2, x3) x1##_##x2##_##x3
-#define __FIPS202_NAMESPACE(x1, x2, x3) ___FIPS202_NAMESPACE(x1, x2, x3)
-
-/*
- * NAMESPACE is PQCP_MLKEM_NATIVE_FIPS202_<BACKEND>_
- * e.g., PQCP_MLKEM_NATIVE_FIPS202_AARCH64_
- */
-#define FIPS202_NAMESPACE(s) \
-  __FIPS202_NAMESPACE(PQCP_MLKEM_NATIVE_FIPS202, MLKEM_NATIVE_BACKEND, s)
-#define _FIPS202_NAMESPACE(s) \
-  __FIPS202_NAMESPACE(_PQCP_MLKEM_NATIVE_FIPS202, MLKEM_NATIVE_BACKEND, s)
-
-#endif
diff --git a/fips202/native/aarch64/cortex_a55.h b/fips202/native/aarch64/cortex_a55.h
new file mode 100644
index 000000000..3d9d92879
--- /dev/null
+++ b/fips202/native/aarch64/cortex_a55.h
@@ -0,0 +1,24 @@
+/*
+ * Copyright (c) 2024 The mlkem-native project authors
+ * SPDX-License-Identifier: Apache-2.0
+ */
+
+/* FIPS202 assembly profile targeting Cortex-A55 */
+
+#ifdef FIPS202_NATIVE_PROFILE_H
+#error Only one FIPS202 assembly profile can be defined -- did you include multiple profiles?
+#else
+#define FIPS202_NATIVE_PROFILE_H
+
+/* Identifier for this backend so that source and assembly files
+ * in the build can be appropriately guarded. */
+#define MLKEM_NATIVE_FIPS202_BACKEND_AARCH64_A55
+
+#define MLKEM_NATIVE_FIPS202_BACKEND_NAME AARCH64_A55
+
+/* Filename of the C backend implementation.
+ * This is not inlined here because this header is included in assembly
+ * files as well. */
+#define MLKEM_NATIVE_FIPS202_BACKEND_IMPL "aarch64/src/cortex_a55_impl.h"
+
+#endif /* FIPS202_NATIVE_PROFILE_H */
diff --git a/fips202/native/aarch64/default.h b/fips202/native/aarch64/default.h
new file mode 100644
index 000000000..1386c6440
--- /dev/null
+++ b/fips202/native/aarch64/default.h
@@ -0,0 +1,24 @@
+/*
+ * Copyright (c) 2024 The mlkem-native project authors
+ * SPDX-License-Identifier: Apache-2.0
+ */
+
+/* Default FIPS202 assembly profile for AArch64 systems */
+
+#ifdef FIPS202_NATIVE_PROFILE_H
+#error Only one FIPS202 assembly profile can be defined -- did you include multiple profiles?
+#else
+#define FIPS202_NATIVE_PROFILE_H
+
+/* Identifier for this backend so that source and assembly files
+ * in the build can be appropriately guarded. */
+#define MLKEM_NATIVE_FIPS202_BACKEND_AARCH64_DEFAULT
+
+#define MLKEM_NATIVE_FIPS202_BACKEND_NAME AARCH64_DEFAULT
+
+/* Filename of the C backend implementation.
+ * This is not inlined here because this header is included in assembly
+ * files as well. */
+#define MLKEM_NATIVE_FIPS202_BACKEND_IMPL "aarch64/src/default_impl.h"
+
+#endif /* FIPS202_NATIVE_PROFILE_H */
diff --git a/fips202/native/aarch64/common.i b/fips202/native/aarch64/src/common.i
similarity index 100%
rename from fips202/native/aarch64/common.i
rename to fips202/native/aarch64/src/common.i
diff --git a/fips202/native/aarch64/profiles/cortex_a55.h b/fips202/native/aarch64/src/cortex_a55_impl.h
similarity index 76%
rename from fips202/native/aarch64/profiles/cortex_a55.h
rename to fips202/native/aarch64/src/cortex_a55_impl.h
index 3ca536bf4..90b857f66 100644
--- a/fips202/native/aarch64/profiles/cortex_a55.h
+++ b/fips202/native/aarch64/src/cortex_a55_impl.h
@@ -5,12 +5,12 @@
 
 /* FIPS202 assembly profile targeting Cortex-A55 */
 
-#ifdef FIPS202_NATIVE_PROFILE_H
+#ifdef FIPS202_NATIVE_PROFILE_IMPL_H
 #error Only one FIPS202 assembly profile can be defined -- did you include multiple profiles?
 #else
-#define FIPS202_NATIVE_PROFILE_H
+#define FIPS202_NATIVE_PROFILE_IMPL_H
 
-#include "../fips202_native_aarch64.h"
+#include "fips202_native_aarch64.h"
 
 /*
  * On Cortex-A55, we use lazy rotation assembly for Keccak-x1,
@@ -22,4 +22,4 @@ static INLINE void keccak_f1600_x1_native(uint64_t *state)
   keccak_f1600_x1_scalar_asm_opt(state);
 }
 
-#endif /* FIPS202_NATIVE_PROFILE_H */
+#endif /* FIPS202_NATIVE_PROFILE_IMPL_H */
diff --git a/fips202/native/aarch64/profiles/default.h b/fips202/native/aarch64/src/default_impl.h
similarity index 95%
rename from fips202/native/aarch64/profiles/default.h
rename to fips202/native/aarch64/src/default_impl.h
index 3a9da28b7..4d28e35bf 100644
--- a/fips202/native/aarch64/profiles/default.h
+++ b/fips202/native/aarch64/src/default_impl.h
@@ -5,12 +5,12 @@
 
 /* Default FIPS202 assembly profile for AArch64 systems */
 
-#ifdef FIPS202_NATIVE_PROFILE_H
+#ifdef FIPS202_NATIVE_PROFILE_IMPL_H
 #error Only one FIPS202 assembly profile can be defined -- did you include multiple profiles?
 #else
-#define FIPS202_NATIVE_PROFILE_H
+#define FIPS202_NATIVE_PROFILE_IMPL_H
 
-#include "../fips202_native_aarch64.h"
+#include "fips202_native_aarch64.h"
 
 /*
  * Default logic to decide which implementation to use.
diff --git a/fips202/native/aarch64/fips202_native_aarch64.h b/fips202/native/aarch64/src/fips202_native_aarch64.h
similarity index 91%
rename from fips202/native/aarch64/fips202_native_aarch64.h
rename to fips202/native/aarch64/src/fips202_native_aarch64.h
index 72e7c9710..d62b2c958 100644
--- a/fips202/native/aarch64/fips202_native_aarch64.h
+++ b/fips202/native/aarch64/src/fips202_native_aarch64.h
@@ -6,11 +6,8 @@
 #define FIPS202_AARCH64_NATIVE_H
 
 #include <stdint.h>
-#include "config.h"
-#include "namespace.h"
-#include "params.h"
+#include "common.h"
 
-#ifdef MLKEM_USE_NATIVE_AARCH64
 #define keccak_f1600_x1_scalar_asm_opt \
   FIPS202_NAMESPACE(keccak_f1600_x1_scalar_asm_opt)
 void keccak_f1600_x1_scalar_asm_opt(uint64_t *state);
@@ -38,6 +35,5 @@ void keccak_f1600_x4_scalar_v84a_asm_hybrid_opt(uint64_t *state);
 #define keccak_f1600_x4_scalar_v8a_v84a_hybrid_asm_opt \
   FIPS202_NAMESPACE(keccak_f1600_x4_scalar_v8a_v84a_hybrid_asm_opt)
 void keccak_f1600_x4_scalar_v8a_v84a_hybrid_asm_opt(uint64_t *state);
-#endif /* MLKEM_USE_NATIVE_AARCH64 */
 
 #endif /* FIPS202_AARCH64_NATIVE_H */
diff --git a/fips202/native/aarch64/keccak_f1600_x1_scalar_asm_opt.S b/fips202/native/aarch64/src/keccak_f1600_x1_scalar_asm_opt.S
similarity index 99%
rename from fips202/native/aarch64/keccak_f1600_x1_scalar_asm_opt.S
rename to fips202/native/aarch64/src/keccak_f1600_x1_scalar_asm_opt.S
index 3db9e2046..3655a219d 100644
--- a/fips202/native/aarch64/keccak_f1600_x1_scalar_asm_opt.S
+++ b/fips202/native/aarch64/src/keccak_f1600_x1_scalar_asm_opt.S
@@ -28,12 +28,12 @@
 // Author: Hanno Becker <hanno.becker@arm.com>
 // Author: Matthias Kannwischer <matthias@kannwischer.eu>
 
-#include "config.h"
-#if defined(MLKEM_USE_NATIVE_AARCH64)
+#include "common.h"
+#if defined(MLKEM_NATIVE_FIPS202_BACKEND_AARCH64_DEFAULT) || \
+    defined(MLKEM_NATIVE_FIPS202_BACKEND_AARCH64_A55)
 
 // Needed to provide ASM_LOAD directive
 #include "common.i"
-#include "namespace.h"
 
 /********************** CONSTANTS *************************/
     .data
@@ -457,4 +457,5 @@ initial:
     free_stack
     ret
 
-#endif /* MLKEM_USE_NATIVE_AARCH64 */
+#endif /* defined(MLKEM_NATIVE_FIPS202_BACKEND_AARCH64_DEFAULT) ||
+          defined(MLKEM_NATIVE_FIPS202_BACKEND_AARCH64_A55) */
diff --git a/fips202/native/aarch64/keccak_f1600_x1_v84a_asm_clean.S b/fips202/native/aarch64/src/keccak_f1600_x1_v84a_asm_clean.S
similarity index 98%
rename from fips202/native/aarch64/keccak_f1600_x1_v84a_asm_clean.S
rename to fips202/native/aarch64/src/keccak_f1600_x1_v84a_asm_clean.S
index 0613a8a75..c60e4cad4 100644
--- a/fips202/native/aarch64/keccak_f1600_x1_v84a_asm_clean.S
+++ b/fips202/native/aarch64/src/keccak_f1600_x1_v84a_asm_clean.S
@@ -37,8 +37,9 @@
 // during load and store, so that the caller need not do this.
 //
 
-#include "config.h"
-#if defined(MLKEM_USE_NATIVE_AARCH64)
+#include "common.h"
+#if defined(MLKEM_NATIVE_FIPS202_BACKEND_AARCH64_DEFAULT) || \
+    defined(MLKEM_NATIVE_FIPS202_BACKEND_AARCH64_A55)
 
 // Needed to provide ASM_LOAD directive
 #include "common.i"
diff --git a/fips202/native/aarch64/keccak_f1600_x2_v84a_asm_clean.S b/fips202/native/aarch64/src/keccak_f1600_x2_v84a_asm_clean.S
similarity index 98%
rename from fips202/native/aarch64/keccak_f1600_x2_v84a_asm_clean.S
rename to fips202/native/aarch64/src/keccak_f1600_x2_v84a_asm_clean.S
index f74856a5f..45304aa25 100644
--- a/fips202/native/aarch64/keccak_f1600_x2_v84a_asm_clean.S
+++ b/fips202/native/aarch64/src/keccak_f1600_x2_v84a_asm_clean.S
@@ -37,8 +37,9 @@
 // during load and store, so that the caller need not do this.
 //
 
-#include "config.h"
-#if defined(MLKEM_USE_NATIVE_AARCH64)
+#include "common.h"
+#if defined(MLKEM_NATIVE_FIPS202_BACKEND_AARCH64_DEFAULT) || \
+    defined(MLKEM_NATIVE_FIPS202_BACKEND_AARCH64_A55)
 
 // Needed to provide ASM_LOAD directive
 #include "common.i"
diff --git a/fips202/native/aarch64/keccak_f1600_x2_v8a_v84a_asm_hybrid.S b/fips202/native/aarch64/src/keccak_f1600_x2_v8a_v84a_asm_hybrid.S
similarity index 98%
rename from fips202/native/aarch64/keccak_f1600_x2_v8a_v84a_asm_hybrid.S
rename to fips202/native/aarch64/src/keccak_f1600_x2_v8a_v84a_asm_hybrid.S
index b7cc6f261..e4fecb11d 100644
--- a/fips202/native/aarch64/keccak_f1600_x2_v8a_v84a_asm_hybrid.S
+++ b/fips202/native/aarch64/src/keccak_f1600_x2_v8a_v84a_asm_hybrid.S
@@ -37,12 +37,12 @@
 // during load and store, so that the caller need not do this.
 //
 
-#include "config.h"
-#if defined(MLKEM_USE_NATIVE_AARCH64)
+#include "common.h"
+#if defined(MLKEM_NATIVE_FIPS202_BACKEND_AARCH64_DEFAULT) || \
+    defined(MLKEM_NATIVE_FIPS202_BACKEND_AARCH64_A55)
 
 // Needed to provide ASM_LOAD directive
 #include "common.i"
-#include "namespace.h"
 
 #if defined(__ARM_FEATURE_SHA3)
 
diff --git a/fips202/native/aarch64/keccak_f1600_x4_scalar_v84a_asm_hybrid_opt.S b/fips202/native/aarch64/src/keccak_f1600_x4_scalar_v84a_asm_hybrid_opt.S
similarity index 99%
rename from fips202/native/aarch64/keccak_f1600_x4_scalar_v84a_asm_hybrid_opt.S
rename to fips202/native/aarch64/src/keccak_f1600_x4_scalar_v84a_asm_hybrid_opt.S
index 19a93b9e8..abf1af497 100644
--- a/fips202/native/aarch64/keccak_f1600_x4_scalar_v84a_asm_hybrid_opt.S
+++ b/fips202/native/aarch64/src/keccak_f1600_x4_scalar_v84a_asm_hybrid_opt.S
@@ -24,12 +24,12 @@
  *
  */
 
-#include "config.h"
-#if defined(MLKEM_USE_NATIVE_AARCH64)
+#include "common.h"
+#if defined(MLKEM_NATIVE_FIPS202_BACKEND_AARCH64_DEFAULT) || \
+    defined(MLKEM_NATIVE_FIPS202_BACKEND_AARCH64_A55)
 
 // Needed to provide ASM_LOAD directive
 #include "common.i"
-#include "namespace.h"
 
 #if defined(__ARM_FEATURE_SHA3)
 
diff --git a/fips202/native/aarch64/keccak_f1600_x4_v8a_scalar_hybrid_asm_opt.S b/fips202/native/aarch64/src/keccak_f1600_x4_v8a_scalar_hybrid_asm_opt.S
similarity index 99%
rename from fips202/native/aarch64/keccak_f1600_x4_v8a_scalar_hybrid_asm_opt.S
rename to fips202/native/aarch64/src/keccak_f1600_x4_v8a_scalar_hybrid_asm_opt.S
index a32375f07..f001d3eaf 100644
--- a/fips202/native/aarch64/keccak_f1600_x4_v8a_scalar_hybrid_asm_opt.S
+++ b/fips202/native/aarch64/src/keccak_f1600_x4_v8a_scalar_hybrid_asm_opt.S
@@ -24,12 +24,12 @@
  *
  */
 
-#include "config.h"
-#if defined(MLKEM_USE_NATIVE_AARCH64)
+#include "common.h"
+#if defined(MLKEM_NATIVE_FIPS202_BACKEND_AARCH64_DEFAULT) || \
+    defined(MLKEM_NATIVE_FIPS202_BACKEND_AARCH64_A55)
 
 // Needed to provide ASM_LOAD directive
 #include "common.i"
-#include "namespace.h"
 
 #define KECCAK_F1600_ROUNDS 24
 
diff --git a/fips202/native/aarch64/keccak_f1600_x4_v8a_v84a_scalar_hybrid_asm_opt.S b/fips202/native/aarch64/src/keccak_f1600_x4_v8a_v84a_scalar_hybrid_asm_opt.S
similarity index 99%
rename from fips202/native/aarch64/keccak_f1600_x4_v8a_v84a_scalar_hybrid_asm_opt.S
rename to fips202/native/aarch64/src/keccak_f1600_x4_v8a_v84a_scalar_hybrid_asm_opt.S
index 889ba2983..67a0adccd 100644
--- a/fips202/native/aarch64/keccak_f1600_x4_v8a_v84a_scalar_hybrid_asm_opt.S
+++ b/fips202/native/aarch64/src/keccak_f1600_x4_v8a_v84a_scalar_hybrid_asm_opt.S
@@ -24,8 +24,9 @@
  *
  */
 
-#include "config.h"
-#if defined(MLKEM_USE_NATIVE_AARCH64)
+#include "common.h"
+#if defined(MLKEM_NATIVE_FIPS202_BACKEND_AARCH64_DEFAULT) || \
+    defined(MLKEM_NATIVE_FIPS202_BACKEND_AARCH64_A55)
 
 // Needed to provide ASM_LOAD directive
 #include "common.i"
diff --git a/fips202/native/api.h b/fips202/native/api.h
new file mode 100644
index 000000000..685b8bfc5
--- /dev/null
+++ b/fips202/native/api.h
@@ -0,0 +1,33 @@
+/*
+ * Copyright (c) 2024 The mlkem-native project authors
+ * SPDX-License-Identifier: Apache-2.0
+ */
+#ifndef FIPS202_NATIVE_H
+#define FIPS202_NATIVE_H
+
+#include <stdint.h>
+#include "common.h"
+
+/*
+ * FIPS202 native interface
+ */
+
+/*
+ * Those functions are meant to be trivial wrappers around
+ *  the chosen native implementation. The are static inline
+ * to avoid unnecessary calls.
+ * The macro before each declaration controls whether a native
+ * implementation is present.
+ */
+
+#if defined(MLKEM_USE_FIPS202_X1_NATIVE)
+static INLINE void keccak_f1600_x1_native(uint64_t *state);
+#endif
+#if defined(MLKEM_USE_FIPS202_X2_NATIVE)
+static INLINE void keccak_f1600_x2_native(uint64_t *state);
+#endif
+#if defined(MLKEM_USE_FIPS202_X4_NATIVE)
+static INLINE void keccak_f1600_x4_native(uint64_t *state);
+#endif
+
+#endif /* FIPS202_NATIVE_H */
diff --git a/fips202/native/default.h b/fips202/native/default.h
new file mode 100644
index 000000000..41c068910
--- /dev/null
+++ b/fips202/native/default.h
@@ -0,0 +1,22 @@
+/*
+ * Copyright (c) 2024 The mlkem-native project authors
+ * SPDX-License-Identifier: Apache-2.0
+ */
+
+#ifndef MLKEM_NATIVE_FIPS202_BACKEND_DEFAULT_H
+#define MLKEM_NATIVE_FIPS202_BACKEND_DEFAULT_H
+
+/*
+ * Default FIPS202 backend
+ */
+#include "sys.h"
+
+#if defined(SYS_AARCH64)
+#include "aarch64/default.h"
+#endif
+
+#if defined(SYS_X86_64) && defined(SYS_X86_64_AVX2)
+#include "x86_64/xkcp.h"
+#endif
+
+#endif /* MLKEM_NATIVE_FIPS202_BACKEND_DEFAULT_H */
diff --git a/fips202/native/fips202_native.h b/fips202/native/fips202_native.h
deleted file mode 100644
index ccddc32a9..000000000
--- a/fips202/native/fips202_native.h
+++ /dev/null
@@ -1,81 +0,0 @@
-/*
- * Copyright (c) 2024 The mlkem-native project authors
- * SPDX-License-Identifier: Apache-2.0
- */
-#ifndef FIPS202_NATIVE_H
-#define FIPS202_NATIVE_H
-
-#include <stdint.h>
-#include "config.h"
-#include "params.h"
-
-#if defined(MLKEM_USE_NATIVE)
-
-/*
- * FIPS202 native profile
- *
- * The profile decides which implementation(s) of FIPS202 to use.
- *
- * If you don't change anything, the default profile will be used. This profile
- * picks implementations based on characteristics of your system visible to
- * the compiler.
- *
- * The default logic is not perfect, and you may want to pick a specific
- * profile for your target. There are three ways to do so, in descending
- * order of convenience to the user:
- * 1. Pick one of the profiles shipped with this repository.
- * 2. Provide your own profile and register it via FIPS202_ASM_PROFILE
- *    (which must be the profile's path relative to this directoru).
- * 3. Set FIPS202_NATIVE_MANUAL and use an adhoc profile specified via CFLAGS.
- */
-
-/* Option 2: Manually written profile */
-#if defined(FIPS202_NATIVE_PROFILE)
-
-#define STRINGIFY_(x) #x
-#define STRINGIFY(x) STRINGIFY_(x)
-#include STRINGIFY(FIPS202_NATIVE_PROFILE)
-
-/* Option 1: Choose from shipped list of profiles */
-#elif !defined(FIPS202_NATIVE_MANUAL)
-
-#ifdef SYS_AARCH64
-/* Pick exactly one profile from the following list */
-#include "aarch64/profiles/default.h"
-/* #include "aarch64/profiles/cortex_a55.h" */
-#endif
-
-#if defined(SYS_X86_64) && defined(SYS_X86_64_AVX2)
-#include "x86_64/profiles/xkcp.h"
-#endif
-
-#else /* !FIPS202_NATIVE_PROFILE && FIPS202_NATIVE_MANUAL */
-
-/* Option 3: Build your own profile here, or via CFLAGS */
-
-#endif /* !FIPS202_NATIVE_PROFILE && !FIPS202_NATIVE_MANUAL */
-
-/*
- * FIPS202 native interface
- */
-
-/*
- * Those functions are meant to be trivial wrappers around
- *  the chosen native implementation. The are static inline
- * to avoid unnecessary calls.
- * The macro before each declaration controls whether a native
- * implementation is present.
- */
-
-#if defined(MLKEM_USE_FIPS202_X1_NATIVE)
-static INLINE void keccak_f1600_x1_native(uint64_t *state);
-#endif
-#if defined(MLKEM_USE_FIPS202_X2_NATIVE)
-static INLINE void keccak_f1600_x2_native(uint64_t *state);
-#endif
-#if defined(MLKEM_USE_FIPS202_X4_NATIVE)
-static INLINE void keccak_f1600_x4_native(uint64_t *state);
-#endif
-
-#endif /* MLKEM_USE_NATIVE */
-#endif /* FIPS202_NATIVE_H */
diff --git a/fips202/native/x86_64/fips202_native_x86_64.h b/fips202/native/x86_64/fips202_native_x86_64.h
deleted file mode 100644
index cbc0bb97d..000000000
--- a/fips202/native/x86_64/fips202_native_x86_64.h
+++ /dev/null
@@ -1,20 +0,0 @@
-/*
- * Copyright (c) 2024 The mlkem-native project authors
- * SPDX-License-Identifier: Apache-2.0
- */
-#ifndef FIPS202_X86_64_NATIVE_H
-#define FIPS202_X86_64_NATIVE_H
-
-#include "config.h"
-
-#ifdef MLKEM_USE_NATIVE_X86_64
-
-#include <immintrin.h>
-#include <stddef.h>
-#include <stdint.h>
-
-#include "xkcp/KeccakP-1600-times4-SnP.h"
-
-#endif /* MLKEM_USE_NATIVE_X86_64 */
-
-#endif /* FIPS202_X86_64_NATIVE_H */
diff --git a/fips202/native/x86_64/xkcp/KeccakP-1600-times4-SIMD256.c b/fips202/native/x86_64/src/KeccakP-1600-times4-SIMD256.c
similarity index 99%
rename from fips202/native/x86_64/xkcp/KeccakP-1600-times4-SIMD256.c
rename to fips202/native/x86_64/src/KeccakP-1600-times4-SIMD256.c
index 9a36f17aa..d56823ab3 100644
--- a/fips202/native/x86_64/xkcp/KeccakP-1600-times4-SIMD256.c
+++ b/fips202/native/x86_64/src/KeccakP-1600-times4-SIMD256.c
@@ -24,8 +24,8 @@ and related or neighboring rights to the source code in this file.
  *   Keccak states in memory.
  */
 
-#include "config.h"
-#if defined(MLKEM_USE_NATIVE_X86_64) && defined(SYS_X86_64_AVX2)
+#include "common.h"
+#if defined(MLKEM_NATIVE_FIPS202_BACKEND_X86_64_XKCP)
 
 #include <emmintrin.h>
 #include <immintrin.h>
@@ -445,10 +445,9 @@ void KeccakP1600times4_PermuteAll_24rounds(void *states)
 }
 
 #else
-#include "params.h"
 
 /* Dummy constant to keep compiler happy despite empty CU */
-#define empty_cu_avx2_keccakx4 MLKEM_NAMESPACE(empty_cu_avx2_keccakx4)
+#define empty_cu_avx2_keccakx4 FIPS202_NAMESPACE(empty_cu_avx2_keccakx4)
 int empty_cu_avx2_keccakx4;
 
-#endif /* MLKEM_USE_NATIVE_X86_64 && SYS_X86_64_AVX2 */
+#endif /* MLKEM_NATIVE_FIPS202_BACKEND_X86_64_XKCP */
diff --git a/fips202/native/x86_64/xkcp/KeccakP-1600-times4-SnP.h b/fips202/native/x86_64/src/KeccakP-1600-times4-SnP.h
similarity index 97%
rename from fips202/native/x86_64/xkcp/KeccakP-1600-times4-SnP.h
rename to fips202/native/x86_64/src/KeccakP-1600-times4-SnP.h
index 6d1951787..0738c9488 100644
--- a/fips202/native/x86_64/xkcp/KeccakP-1600-times4-SnP.h
+++ b/fips202/native/x86_64/src/KeccakP-1600-times4-SnP.h
@@ -26,7 +26,7 @@ and related or neighboring rights to the source code in this file.
 
 #include <stddef.h>
 #include "KeccakP-SIMD256-config.h"
-#include "namespace.h"
+#include "common.h"
 
 #define KeccakP1600times4_statesAlignment 32
 
diff --git a/fips202/native/x86_64/xkcp/KeccakP-1600-unrolling.macros b/fips202/native/x86_64/src/KeccakP-1600-unrolling.macros
similarity index 100%
rename from fips202/native/x86_64/xkcp/KeccakP-1600-unrolling.macros
rename to fips202/native/x86_64/src/KeccakP-1600-unrolling.macros
diff --git a/fips202/native/x86_64/xkcp/KeccakP-SIMD256-config.h b/fips202/native/x86_64/src/KeccakP-SIMD256-config.h
similarity index 100%
rename from fips202/native/x86_64/xkcp/KeccakP-SIMD256-config.h
rename to fips202/native/x86_64/src/KeccakP-SIMD256-config.h
diff --git a/fips202/native/x86_64/xkcp/KeccakP-align.h b/fips202/native/x86_64/src/KeccakP-align.h
similarity index 100%
rename from fips202/native/x86_64/xkcp/KeccakP-align.h
rename to fips202/native/x86_64/src/KeccakP-align.h
diff --git a/fips202/native/x86_64/xkcp/KeccakP-brg_endian.h b/fips202/native/x86_64/src/KeccakP-brg_endian.h
similarity index 100%
rename from fips202/native/x86_64/xkcp/KeccakP-brg_endian.h
rename to fips202/native/x86_64/src/KeccakP-brg_endian.h
diff --git a/fips202/native/x86_64/profiles/xkcp.h b/fips202/native/x86_64/src/xkcp_impl.h
similarity index 60%
rename from fips202/native/x86_64/profiles/xkcp.h
rename to fips202/native/x86_64/src/xkcp_impl.h
index 4f320e994..c0ad76850 100644
--- a/fips202/native/x86_64/profiles/xkcp.h
+++ b/fips202/native/x86_64/src/xkcp_impl.h
@@ -5,14 +5,12 @@
 
 /* Default FIPS202 assembly profile for AArch64 systems */
 
-#ifdef FIPS202_NATIVE_PROFILE_H
+#ifdef MLKEM_NATIVE_FIPS202_PROFILE_IMPL_H
 #error Only one FIPS202 assembly profile can be defined -- did you include multiple profiles?
 #else
-#define FIPS202_NATIVE_PROFILE_H
+#define MLKEM_NATIVE_FIPS202_PROFILE_IMPL_H
 
-#include "../fips202_native_x86_64.h"
-
-#if defined(MLKEM_USE_NATIVE_X86_64) && defined(SYS_X86_64_AVX2)
+#include "KeccakP-1600-times4-SnP.h"
 
 #define MLKEM_USE_FIPS202_X4_NATIVE
 static INLINE void keccak_f1600_x4_native(uint64_t *state)
@@ -20,6 +18,4 @@ static INLINE void keccak_f1600_x4_native(uint64_t *state)
   KeccakP1600times4_PermuteAll_24rounds(state);
 }
 
-#endif /* MLKEM_USE_NATIVE_X86_64 && SYS_X86_64_AVX2 */
-
-#endif /* FIPS202_NATIVE_PROFILE_H */
+#endif /* MLKEM_NATIVE_FIPS202_PROFILE_IMPL_H */
diff --git a/fips202/native/x86_64/xkcp.h b/fips202/native/x86_64/xkcp.h
new file mode 100644
index 000000000..1cfcee9fd
--- /dev/null
+++ b/fips202/native/x86_64/xkcp.h
@@ -0,0 +1,24 @@
+/*
+ * Copyright (c) 2024 The mlkem-native project authors
+ * SPDX-License-Identifier: Apache-2.0
+ */
+
+/* Default FIPS202 assembly profile for AArch64 systems */
+
+#ifdef MLKEM_NATIVE_FIPS202_PROFILE_H
+#error Only one FIPS202 assembly profile can be defined -- did you include multiple profiles?
+#else
+#define MLKEM_NATIVE_FIPS202_PROFILE_H
+
+/* Identifier for this backend so that source and assembly files
+ * in the build can be appropriately guarded. */
+#define MLKEM_NATIVE_FIPS202_BACKEND_X86_64_XKCP
+
+#define MLKEM_NATIVE_FIPS202_BACKEND_NAME X86_64_XKCP
+
+/* Filename of the C backend implementation.
+ * This is not inlined here because this header is included in assembly
+ * files as well. */
+#define MLKEM_NATIVE_FIPS202_BACKEND_IMPL "x86_64/src/xkcp_impl.h"
+
+#endif /* MLKEM_NATIVE_FIPS202_PROFILE_H */
diff --git a/mk/crypto.mk b/mk/crypto.mk
index faf0eb047..9cb04227d 100644
--- a/mk/crypto.mk
+++ b/mk/crypto.mk
@@ -2,7 +2,7 @@
 CPPFLAGS += -Ifips202 -Ifips202/native
 FIPS202_SRCS = $(wildcard fips202/*.c)
 ifeq ($(OPT),1)
-	FIPS202_SRCS += $(wildcard fips202/native/aarch64/*.S) $(wildcard fips202/native/x86_64/xkcp/*.c)
+	FIPS202_SRCS += $(wildcard fips202/native/aarch64/src/*.S) $(wildcard fips202/native/x86_64/src/*.c)
 endif
 
 $(BUILD_DIR)/libmlkem.a: $(call OBJS, $(FIPS202_SRCS))
diff --git a/mk/schemes.mk b/mk/schemes.mk
index deabcdd86..ede00a635 100644
--- a/mk/schemes.mk
+++ b/mk/schemes.mk
@@ -1,7 +1,7 @@
 # SPDX-License-Identifier: Apache-2.0
 SOURCES += $(wildcard mlkem/*.c) $(wildcard mlkem/debug/*.c)
 ifeq ($(OPT),1)
-	SOURCES += $(wildcard mlkem/native/aarch64/*.[csS]) $(wildcard mlkem/native/x86_64/*.[csS])
+	SOURCES += $(wildcard mlkem/native/aarch64/src/*.[csS]) $(wildcard mlkem/native/x86_64/src/*.[csS])
 	CPPFLAGS += -DMLKEM_USE_NATIVE
 endif
 
diff --git a/mlkem/api.h b/mlkem/api.h
deleted file mode 100644
index 94597323f..000000000
--- a/mlkem/api.h
+++ /dev/null
@@ -1,62 +0,0 @@
-/*
- * Copyright (c) 2024 The mlkem-native project authors
- * SPDX-License-Identifier: Apache-2.0
- */
-#ifndef API_H
-#define API_H
-
-#include <stdint.h>
-
-#define PQCP_MLKEM_NATIVE_MLKEM512_SECRETKEYBYTES 1632
-#define PQCP_MLKEM_NATIVE_MLKEM512_PUBLICKEYBYTES 800
-#define PQCP_MLKEM_NATIVE_MLKEM512_CIPHERTEXTBYTES 768
-#define PQCP_MLKEM_NATIVE_MLKEM512_KEYPAIRCOINBYTES 64
-#define PQCP_MLKEM_NATIVE_MLKEM512_ENCCOINBYTES 32
-#define PQCP_MLKEM_NATIVE_MLKEM512_BYTES 32
-
-int PQCP_MLKEM_NATIVE_MLKEM512_keypair_derand(uint8_t *pk, uint8_t *sk,
-                                              const uint8_t *coins);
-int PQCP_MLKEM_NATIVE_MLKEM512_keypair(uint8_t *pk, uint8_t *sk);
-int PQCP_MLKEM_NATIVE_MLKEM512_enc_derand(uint8_t *ct, uint8_t *ss,
-                                          const uint8_t *pk,
-                                          const uint8_t *coins);
-int PQCP_MLKEM_NATIVE_MLKEM512_enc(uint8_t *ct, uint8_t *ss, const uint8_t *pk);
-int PQCP_MLKEM_NATIVE_MLKEM512_dec(uint8_t *ss, const uint8_t *ct,
-                                   const uint8_t *sk);
-
-#define PQCP_MLKEM_NATIVE_MLKEM768_SECRETKEYBYTES 2400
-#define PQCP_MLKEM_NATIVE_MLKEM768_PUBLICKEYBYTES 1184
-#define PQCP_MLKEM_NATIVE_MLKEM768_CIPHERTEXTBYTES 1088
-#define PQCP_MLKEM_NATIVE_MLKEM768_KEYPAIRCOINBYTES 64
-#define PQCP_MLKEM_NATIVE_MLKEM768_ENCCOINBYTES 32
-#define PQCP_MLKEM_NATIVE_MLKEM768_BYTES 32
-
-int PQCP_MLKEM_NATIVE_MLKEM768_keypair_derand(uint8_t *pk, uint8_t *sk,
-                                              const uint8_t *coins);
-int PQCP_MLKEM_NATIVE_MLKEM768_keypair(uint8_t *pk, uint8_t *sk);
-int PQCP_MLKEM_NATIVE_MLKEM768_enc_derand(uint8_t *ct, uint8_t *ss,
-                                          const uint8_t *pk,
-                                          const uint8_t *coins);
-int PQCP_MLKEM_NATIVE_MLKEM768_enc(uint8_t *ct, uint8_t *ss, const uint8_t *pk);
-int PQCP_MLKEM_NATIVE_MLKEM768_dec(uint8_t *ss, const uint8_t *ct,
-                                   const uint8_t *sk);
-
-#define PQCP_MLKEM_NATIVE_MLKEM1024_SECRETKEYBYTES 3168
-#define PQCP_MLKEM_NATIVE_MLKEM1024_PUBLICKEYBYTES 1568
-#define PQCP_MLKEM_NATIVE_MLKEM1024_CIPHERTEXTBYTES 1568
-#define PQCP_MLKEM_NATIVE_MLKEM1024_KEYPAIRCOINBYTES 64
-#define PQCP_MLKEM_NATIVE_MLKEM1024_ENCCOINBYTES 32
-#define PQCP_MLKEM_NATIVE_MLKEM1024_BYTES 32
-
-int PQCP_MLKEM_NATIVE_MLKEM1024_keypair_derand(uint8_t *pk, uint8_t *sk,
-                                               const uint8_t *coins);
-int PQCP_MLKEM_NATIVE_MLKEM1024_keypair(uint8_t *pk, uint8_t *sk);
-int PQCP_MLKEM_NATIVE_MLKEM1024_enc_derand(uint8_t *ct, uint8_t *ss,
-                                           const uint8_t *pk,
-                                           const uint8_t *coins);
-int PQCP_MLKEM_NATIVE_MLKEM1024_enc(uint8_t *ct, uint8_t *ss,
-                                    const uint8_t *pk);
-int PQCP_MLKEM_NATIVE_MLKEM1024_dec(uint8_t *ss, const uint8_t *ct,
-                                    const uint8_t *sk);
-
-#endif
diff --git a/mlkem/arith_backend.h b/mlkem/arith_backend.h
new file mode 100644
index 000000000..7fb3df84d
--- /dev/null
+++ b/mlkem/arith_backend.h
@@ -0,0 +1,18 @@
+/*
+ * Copyright (c) 2024 The mlkem-native project authors
+ * SPDX-License-Identifier: Apache-2.0
+ */
+
+#ifdef MLKEM_NATIVE_ARITH_IMPL_H
+#error Only one ARITH assembly profile can be defined -- did you include multiple profiles?
+#else
+#define MLKEM_NATIVE_ARITH_IMPL_H
+
+/* Include to enforce consistency of API and implementation */
+#include "native/api.h"
+
+#if defined(MLKEM_NATIVE_ARITH_BACKEND_IMPL)
+#include MLKEM_NATIVE_ARITH_BACKEND_IMPL
+#endif
+
+#endif /* MLKEM_NATIVE_ARITH_IMPL_H */
diff --git a/mlkem/cbd.c b/mlkem/cbd.c
index 073f3c81d..2e0fac38a 100644
--- a/mlkem/cbd.c
+++ b/mlkem/cbd.c
@@ -4,7 +4,6 @@
  */
 #include "cbd.h"
 #include <stdint.h>
-#include "params.h"
 
 /*************************************************
  * Name:        load32_littleendian
diff --git a/mlkem/cbd.h b/mlkem/cbd.h
index 4dc8635bb..31c9649e3 100644
--- a/mlkem/cbd.h
+++ b/mlkem/cbd.h
@@ -6,7 +6,7 @@
 #define CBD_H
 
 #include <stdint.h>
-#include "params.h"
+#include "common.h"
 #include "poly.h"
 
 #define poly_cbd_eta1 MLKEM_NAMESPACE(poly_cbd_eta1)
diff --git a/mlkem/common.h b/mlkem/common.h
index 94c29ed92..0b379d9bd 100644
--- a/mlkem/common.h
+++ b/mlkem/common.h
@@ -2,64 +2,21 @@
  * Copyright (c) 2024 The mlkem-native project authors
  * SPDX-License-Identifier: Apache-2.0
  */
-#ifndef COMMON_H
-#define COMMON_H
+#ifndef MLKEM_NATIVE_COMMON_H
+#define MLKEM_NATIVE_COMMON_H
 
+#include "config.h"
+#include "params.h"
+#include "sys.h"
 
-/*
- * C90 does not have the inline compiler directive yet.
- * We don't use it in C90 builds.
- * However, in that case the compiler warns about some inline functions in
- * header files not being used in every compilation unit that includes that
- * header. To work around it we silence that warning in that case using
- * __attribute__((unused)).
- */
-
-/* Do not use inline for C90 builds*/
-#if !defined(inline)
-#if defined(__STDC_VERSION__) && __STDC_VERSION__ >= 199901L
-#define INLINE inline
-#define ALWAYS_INLINE __attribute__((always_inline))
-#elif defined(_MSC_VER)
-#define INLINE __inline
-#define ALWAYS_INLINE __forceinline
-#else
-#define INLINE __attribute__((unused))
-#define ALWAYS_INLINE
-#endif
-
-#else
-#define INLINE inline
-#define ALWAYS_INLINE __attribute__((always_inline))
-#endif
-
-
-/*
- * C90 does not have the restrict compiler directive yet.
- * We don't use it in C90 builds.
- */
-#if !defined(restrict)
-#if defined(__STDC_VERSION__) && __STDC_VERSION__ >= 199901L
-#define RESTRICT restrict
-#else
-#define RESTRICT
-#endif
-
-#else
-
-#define RESTRICT restrict
-#endif
-
-#define DEFAULT_ALIGN 32
-#if defined(_WIN32)
-#define ALIGN __declspec(align(DEFAULT_ALIGN))
-#define asm __asm
-#else
-#define asm __asm__
-#define ALIGN __attribute__((aligned(DEFAULT_ALIGN)))
+/* Include backend metadata */
+#if defined(MLKEM_USE_NATIVE)
+#include MLKEM_NATIVE_ARITH_BACKEND
+#include MLKEM_NATIVE_FIPS202_BACKEND
 #endif
 
-#define MLKEM_CONCAT_(left, right) left##right
-#define MLKEM_CONCAT(left, right) MLKEM_CONCAT_(left, right)
+/* This must come after the inclusion of the backend metadata
+ * since the backend choice may be part of the namespace. */
+#include "namespace.h"
 
-#endif
+#endif /* MLKEM_NATIVE_COMMON_H */
diff --git a/mlkem/config.h b/mlkem/config.h
new file mode 100644
index 000000000..198b08977
--- /dev/null
+++ b/mlkem/config.h
@@ -0,0 +1,103 @@
+/*
+ * Copyright (c) 2024 The mlkem-native project authors
+ * SPDX-License-Identifier: Apache-2.0
+ */
+
+#ifndef MLKEM_NATIVE_CONFIG_H
+#define MLKEM_NATIVE_CONFIG_H
+
+/******************************************************************************
+ * Name:        MLKEM_K
+ *
+ * Description: Determines the security level for ML-KEM
+ *              - MLKEM_K=2 corresponds to ML-KEM-512
+ *              - MLKEM_K=3 corresponds to ML-KEM-768
+ *              - MLKEM_K=4 corresponds to ML-KEM-1024
+ *
+ *              This can also be set using CFLAGS.
+ *
+ *****************************************************************************/
+#ifndef MLKEM_K
+#define MLKEM_K 3 /* Change this for different security strengths */
+#endif
+
+/******************************************************************************
+ * Name:        MLKEM_NAMESPACE
+ *              _MLKEM_NAMESPACE
+ *
+ * Description: The macros to use to namespace global symbols
+ *              from mlkem/.
+ *****************************************************************************/
+#define MLKEM_NAMESPACE(sym) MLKEM_DEFAULT_NAMESPACE(sym)
+#define _MLKEM_NAMESPACE(sym) _MLKEM_DEFAULT_NAMESPACE(sym)
+
+/******************************************************************************
+ * Name:        FIPS202_NAMESPACE
+ *              _FIPS202_NAMESPACE
+ *
+ * Description: The macros to use to namespace global symbols
+ *              from fips202/.
+ *****************************************************************************/
+#define FIPS202_NAMESPACE(sym) FIPS202_DEFAULT_NAMESPACE(sym)
+#define _FIPS202_NAMESPACE(sym) _FIPS202_DEFAULT_NAMESPACE(sym)
+
+/******************************************************************************
+ * Name:        MLKEM_USE_NATIVE
+ *
+ * Description: Determines whether a native backend should
+ *              be used, if available.
+ *
+ *              This can also be set using CFLAGS.
+ *
+ *****************************************************************************/
+#if !defined(MLKEM_USE_NATIVE)
+/* #define MLKEM_USE_NATIVE */
+#endif
+
+/******************************************************************************
+ * Name:        MLKEM_NATIVE_ARITH_BACKEND
+ *
+ * Description: The arithmetic backend to use.
+ *
+ *              This must be the filename of an arithmetic
+ *              backend. The backend is expected to define
+ *
+ *              - MLKEM_NATIVE_ARITH_BACKEND_NAME
+ *
+ *                The name of the backend as used in the default namespace.
+ *
+ *              - MLKEM_NATIVE_ARITH_BACKEND_IMPL
+ *
+ *                The filename of the implementation of the arithmetic backend.
+ *
+ *              See the existing backends for more information.
+ *
+ *****************************************************************************/
+#if !defined(MLKEM_NATIVE_ARITH_BACKEND)
+#define MLKEM_NATIVE_ARITH_BACKEND "native/default.h"
+#endif /* MLKEM_NATIVE_ARITH_BACKEND */
+
+/******************************************************************************
+ * Name:        MLKEM_NATIVE_FIPS202_BACKEND
+ *
+ * Description: The FIPS-202 backend to use.
+ *
+ *              This must be the filename of an FIPS-202
+ *              backend. The backend is expected to define
+ *
+ *              - MLKEM_NATIVE_FIPS202_BACKEND_NAME
+ *
+ *                The name of the backend as used in the default namespace.
+ *
+ *              - MLKEM_NATIVE_FIPS202_BACKEND_IMPL
+ *
+ *                The filename of the implementation of the FIPS-202 backend.
+ *
+ *              See the existing backends for more information.
+ *
+ *****************************************************************************/
+#if !defined(MLKEM_NATIVE_FIPS202_BACKEND)
+#define MLKEM_NATIVE_FIPS202_BACKEND "../fips202/native/default.h"
+#endif /* MLKEM_NATIVE_FIPS202_BACKEND */
+
+#endif /* MLkEM_NATIVE_CONFIG_H */
diff --git a/mlkem/debug/debug.c b/mlkem/debug/debug.c
index 0dc2e8822..690977908 100644
--- a/mlkem/debug/debug.c
+++ b/mlkem/debug/debug.c
@@ -52,7 +52,6 @@ void mlkem_debug_print_error(const char *file, int line, const char *msg)
 }
 
 #else /* MLKEM_DEBUG */
-#include "params.h"
 
 #define empty_cu_debug MLKEM_NAMESPACE(empty_cu_debug)
 int empty_cu_debug;
diff --git a/mlkem/debug/debug.h b/mlkem/debug/debug.h
index 65208771d..5838ae4bf 100644
--- a/mlkem/debug/debug.h
+++ b/mlkem/debug/debug.h
@@ -5,6 +5,8 @@
 #ifndef MLKEM_DEBUG_H
 #define MLKEM_DEBUG_H
 
+#include "common.h"
+
 #if defined(MLKEM_DEBUG)
 #include <stdint.h>
 #include <stdio.h>
@@ -153,6 +155,9 @@ void mlkem_debug_print_error(const char *file, int line, const char *msg);
                       "polyvec unsigned bound for " #ptr ".vec[i]");       \
   } while (0)
 
+#define MLKEM_CONCAT_(left, right) left##right
+#define MLKEM_CONCAT(left, right) MLKEM_CONCAT_(left, right)
+
 /* Following AWS-LC to define a C99-compliant static assert */
 #define MLKEM_STATIC_ASSERT_DEFINE(cond, msg)                            \
   typedef struct                                                         \
diff --git a/mlkem/indcpa.c b/mlkem/indcpa.c
index bde832c18..ee2f140a0 100644
--- a/mlkem/indcpa.c
+++ b/mlkem/indcpa.c
@@ -10,14 +10,13 @@
 #include "fips202x4.h"
 #include "indcpa.h"
 #include "ntt.h"
-#include "params.h"
 #include "poly.h"
 #include "polyvec.h"
 #include "randombytes.h"
 #include "rej_uniform.h"
 #include "symmetric.h"
 
-#include "arith_native.h"
+#include "arith_backend.h"
 #include "debug/debug.h"
 
 #include "cbmc.h"
diff --git a/mlkem/indcpa.h b/mlkem/indcpa.h
index 0e6485312..7e2a0b247 100644
--- a/mlkem/indcpa.h
+++ b/mlkem/indcpa.h
@@ -7,10 +7,9 @@
 
 #include <stdint.h>
 #include "cbmc.h"
-#include "params.h"
+#include "common.h"
 #include "polyvec.h"
 
-
 #define gen_matrix MLKEM_NAMESPACE(gen_matrix)
 /*************************************************
  * Name:        gen_matrix
diff --git a/mlkem/kem.c b/mlkem/kem.c
index 1abce65e4..03e997af3 100644
--- a/mlkem/kem.c
+++ b/mlkem/kem.c
@@ -7,7 +7,6 @@
 #include <stdint.h>
 #include <string.h>
 #include "indcpa.h"
-#include "params.h"
 #include "randombytes.h"
 #include "symmetric.h"
 #include "verify.h"
diff --git a/mlkem/namespace.h b/mlkem/namespace.h
new file mode 100644
index 000000000..8c409fb0c
--- /dev/null
+++ b/mlkem/namespace.h
@@ -0,0 +1,57 @@
+/*
+ * Copyright (c) 2024 The mlkem-native project authors
+ * SPDX-License-Identifier: Apache-2.0
+ */
+#ifndef MLKEM_NATIVE_NAMESPACE_H
+#define MLKEM_NATIVE_NAMESPACE_H
+
+#if !defined(MLKEM_NATIVE_ARITH_BACKEND_NAME)
+#define MLKEM_NATIVE_ARITH_BACKEND_NAME C
+#endif
+
+/* Don't change parameters below this line */
+#if (MLKEM_K == 2)
+#define MLKEM_PARAM_NAME MLKEM512
+#elif (MLKEM_K == 3)
+#define MLKEM_PARAM_NAME MLKEM768
+#elif (MLKEM_K == 4)
+#define MLKEM_PARAM_NAME MLKEM1024
+#else
+#error "MLKEM_K must be in {2,3,4}"
+#endif
+
+#define ___MLKEM_DEFAULT_NAMESPACE(x1, x2, x3, x4) x1##_##x2##_##x3##_##x4
+#define __MLKEM_DEFAULT_NAMESPACE(x1, x2, x3, x4) \
+  ___MLKEM_DEFAULT_NAMESPACE(x1, x2, x3, x4)
+
+/*
+ * NAMESPACE is PQCP_MLKEM_NATIVE_<PARAM_NAME>_<BACKEND>_
+ * e.g., PQCP_MLKEM_NATIVE_MLKEM512_AARCH64_OPT_
+ */
+#define MLKEM_DEFAULT_NAMESPACE(s)                               \
+  __MLKEM_DEFAULT_NAMESPACE(PQCP_MLKEM_NATIVE, MLKEM_PARAM_NAME, \
+                            MLKEM_NATIVE_ARITH_BACKEND_NAME, s)
+#define _MLKEM_DEFAULT_NAMESPACE(s)                               \
+  __MLKEM_DEFAULT_NAMESPACE(_PQCP_MLKEM_NATIVE, MLKEM_PARAM_NAME, \
+                            MLKEM_NATIVE_ARITH_BACKEND_NAME, s)
+
+#if !defined(MLKEM_NATIVE_FIPS202_BACKEND_NAME)
+#define MLKEM_NATIVE_FIPS202_BACKEND_NAME C
+#endif
+
+#define ___FIPS202_DEFAULT_NAMESPACE(x1, x2, x3) x1##_##x2##_##x3
+#define __FIPS202_DEFAULT_NAMESPACE(x1, x2, x3) \
+  ___FIPS202_DEFAULT_NAMESPACE(x1, x2, x3)
+
+/*
+ * NAMESPACE is PQCP_MLKEM_NATIVE_FIPS202_<BACKEND>_
+ * e.g., PQCP_MLKEM_NATIVE_FIPS202_X86_64_XKCP_
+ */
+#define FIPS202_DEFAULT_NAMESPACE(s)                     \
+  __FIPS202_DEFAULT_NAMESPACE(PQCP_MLKEM_NATIVE_FIPS202, \
+                              MLKEM_NATIVE_FIPS202_BACKEND_NAME, s)
+#define _FIPS202_DEFAULT_NAMESPACE(s)                     \
+  __FIPS202_DEFAULT_NAMESPACE(_PQCP_MLKEM_NATIVE_FIPS202, \
+                              MLKEM_NATIVE_FIPS202_BACKEND_NAME, s)
+
+#endif /* MLKEM_NATIVE_NAMESPACE_H */
diff --git a/mlkem/native/aarch64/README.md b/mlkem/native/aarch64/README.md
index c21e27f36..e499a4a22 100644
--- a/mlkem/native/aarch64/README.md
+++ b/mlkem/native/aarch64/README.md
@@ -16,4 +16,4 @@ This backend comes with two profiles: "clean" and optimized. The "clean" backend
 read and modify; for example, is heavily leverages register aliases and assembly macros. The optimized profile is
 automatically generated from the clean profile via [SLOTHY](https://github.com/slothy-optimizer/slothy). Currently, the
 target architecture is Cortex-A55, but you can easily re-optimize the code for a different microarchitecture supported
-by SLOTHY, by adjusting the parameters in [optimize.sh](optimize.sh).
+by SLOTHY, by adjusting the parameters in [optimize.sh](src/optimize.sh).
diff --git a/mlkem/native/aarch64/clean.h b/mlkem/native/aarch64/clean.h
new file mode 100644
index 000000000..43a401dfc
--- /dev/null
+++ b/mlkem/native/aarch64/clean.h
@@ -0,0 +1,24 @@
+/*
+ * Copyright (c) 2024 The mlkem-native project authors
+ * SPDX-License-Identifier: Apache-2.0
+ */
+
+/* ML-KEM arithmetic native profile for clean assembly */
+
+#ifdef MLKEM_NATIVE_ARITH_PROFILE_H
+#error Only one MLKEM_ARITH assembly profile can be defined -- did you include multiple profiles?
+#else
+#define MLKEM_NATIVE_ARITH_PROFILE_H
+
+/* Identifier for this backend so that source and assembly files
+ * in the build can be appropriately guarded. */
+#define MLKEM_NATIVE_ARITH_BACKEND_AARCH64_CLEAN
+
+#define MLKEM_NATIVE_ARITH_BACKEND_NAME AARCH64_CLEAN
+
+/* Filename of the C backend implementation.
+ * This is not inlined here because this header is included in assembly
+ * files as well. */
+#define MLKEM_NATIVE_ARITH_BACKEND_IMPL "aarch64/src/clean_impl.h"
+
+#endif /* MLKEM_NATIVE_ARITH_PROFILE_H */
diff --git a/mlkem/native/aarch64/opt.h b/mlkem/native/aarch64/opt.h
new file mode 100644
index 000000000..04323c3e7
--- /dev/null
+++ b/mlkem/native/aarch64/opt.h
@@ -0,0 +1,24 @@
+/*
+ * Copyright (c) 2024 The mlkem-native project authors
+ * SPDX-License-Identifier: Apache-2.0
+ */
+
+/* ML-KEM arithmetic native profile for clean assembly */
+
+#ifdef MLKEM_NATIVE_ARITH_PROFILE_H
+#error Only one MLKEM_ARITH assembly profile can be defined -- did you include multiple profiles?
+#else
+#define MLKEM_NATIVE_ARITH_PROFILE_H
+
+/* Identifier for this backend so that source and assembly files
+ * in the build can be appropriately guarded. */
+#define MLKEM_NATIVE_ARITH_BACKEND_AARCH64_OPT
+
+#define MLKEM_NATIVE_ARITH_BACKEND_NAME AARCH64_OPT
+
+/* Filename of the C backend implementation.
+ * This is not inlined here because this header is included in assembly
+ * files as well. */
+#define MLKEM_NATIVE_ARITH_BACKEND_IMPL "aarch64/src/opt_impl.h"
+
+#endif /* MLKEM_NATIVE_ARITH_PROFILE_H */
diff --git a/mlkem/native/aarch64/aarch64_zetas.c b/mlkem/native/aarch64/src/aarch64_zetas.c
similarity index 98%
rename from mlkem/native/aarch64/aarch64_zetas.c
rename to mlkem/native/aarch64/src/aarch64_zetas.c
index f0a0d5546..ecf1b529a 100644
--- a/mlkem/native/aarch64/aarch64_zetas.c
+++ b/mlkem/native/aarch64/src/aarch64_zetas.c
@@ -8,9 +8,13 @@
  *          Do not modify it directly.
  */
 
-#include "arith_native_aarch64.h"
+#include "common.h"
+
+#if defined(MLKEM_NATIVE_ARITH_BACKEND_AARCH64_CLEAN) || \
+    defined(MLKEM_NATIVE_ARITH_BACKEND_AARCH64_OPT)
 
-#ifdef MLKEM_USE_NATIVE_AARCH64
+#include <stdint.h>
+#include "arith_native_aarch64.h"
 
 /*
  * Table of zeta values used in the AArch64 forward NTT
@@ -163,10 +167,10 @@ const int16_t aarch64_zetas_mulcache_twisted_native[] = {
     -11566, 11566,
 };
 
-#else /* MLKEM_USE_NATIVE_AARCH64 */
+#else
 #include "params.h"
 
 /* Dummy declaration for compilers disliking empty compilation units */
 #define empty_cu_aarch64_zetas MLKEM_NAMESPACE(empty_cu_aarch64_zetas)
 int empty_cu_aarch64_zetas;
-#endif /* MLKEM_USE_NATIVE_AARCH64 */
+#endif
diff --git a/mlkem/native/aarch64/arith_native_aarch64.h b/mlkem/native/aarch64/src/arith_native_aarch64.h
similarity index 96%
rename from mlkem/native/aarch64/arith_native_aarch64.h
rename to mlkem/native/aarch64/src/arith_native_aarch64.h
index f39eb9ac8..2f3b0ef4f 100644
--- a/mlkem/native/aarch64/arith_native_aarch64.h
+++ b/mlkem/native/aarch64/src/arith_native_aarch64.h
@@ -6,10 +6,7 @@
 #define MLKEM_AARCH64_NATIVE_H
 
 #include <stdint.h>
-#include "config.h"
-#include "params.h"
-
-#ifdef MLKEM_USE_NATIVE_AARCH64
+#include "common.h"
 
 #define aarch64_ntt_zetas_layer01234 \
   MLKEM_NAMESPACE(aarch64_ntt_zetas_layer01234)
@@ -72,6 +69,9 @@ void poly_mulcache_compute_asm_opt(int16_t *, const int16_t *, const int16_t *,
 #define poly_tobytes_asm_clean MLKEM_NAMESPACE(poly_tobytes_asm_clean)
 void poly_tobytes_asm_clean(uint8_t *r, const int16_t *a);
 
+#define poly_tobytes_asm_opt MLKEM_NAMESPACE(poly_tobytes_asm_opt)
+void poly_tobytes_asm_opt(uint8_t *r, const int16_t *a);
+
 #define polyvec_basemul_acc_montgomery_cached_asm_clean \
   MLKEM_NAMESPACE(polyvec_basemul_acc_montgomery_cached_asm_clean)
 void polyvec_basemul_acc_montgomery_cached_asm_clean(int16_t *r,
@@ -85,6 +85,4 @@ void polyvec_basemul_acc_montgomery_cached_asm_opt(int16_t *r, const int16_t *a,
                                                    const int16_t *b,
                                                    const int16_t *b_cache);
 
-
-#endif /* MLKEM_USE_NATIVE_AARCH64 */
 #endif /* MLKEM_AARCH64_NATIVE_H */
diff --git a/mlkem/native/aarch64/profiles/clean.h b/mlkem/native/aarch64/src/clean_impl.h
similarity index 90%
rename from mlkem/native/aarch64/profiles/clean.h
rename to mlkem/native/aarch64/src/clean_impl.h
index e7f71bd53..0a40199e7 100644
--- a/mlkem/native/aarch64/profiles/clean.h
+++ b/mlkem/native/aarch64/src/clean_impl.h
@@ -5,14 +5,14 @@
 
 /* ML-KEM arithmetic native profile for clean assembly */
 
-#ifdef MLKEM_ARITH_NATIVE_PROFILE_H
+#ifdef MLKEM_NATIVE_ARITH_PROFILE_IMPL_H
 #error Only one MLKEM_ARITH assembly profile can be defined -- did you include multiple profiles?
 #else
-#define MLKEM_ARITH_NATIVE_PROFILE_H
+#define MLKEM_NATIVE_ARITH_PROFILE_IMPL_H
 
-#include "../../arith_native.h"
-#include "../arith_native_aarch64.h"
+#include "arith_native_aarch64.h"
 
+/* Set of primitives that this backend replaces */
 #define MLKEM_USE_NATIVE_NTT
 #define MLKEM_USE_NATIVE_INTT
 #define MLKEM_USE_NATIVE_POLY_REDUCE
@@ -22,7 +22,6 @@
 #define MLKEM_USE_NATIVE_POLY_TOBYTES
 #define MLKEM_USE_NATIVE_REJ_UNIFORM
 
-#define NTT_BOUND_NATIVE (6 * MLKEM_Q)
 static INLINE void ntt_native(poly *data)
 {
   ntt_asm_clean(data->coeffs, aarch64_ntt_zetas_layer01234,
@@ -75,4 +74,4 @@ static INLINE int rej_uniform_native(int16_t *r, unsigned int len,
   return (int)rej_uniform_asm_clean(r, buf, buflen);
 }
 
-#endif /* MLKEM_ARITH_NATIVE_PROFILE_H */
+#endif /* MLKEM_NATIVE_ARITH_PROFILE_IMPL_H */
diff --git a/mlkem/native/aarch64/common.i b/mlkem/native/aarch64/src/common.i
similarity index 100%
rename from mlkem/native/aarch64/common.i
rename to mlkem/native/aarch64/src/common.i
diff --git a/mlkem/native/aarch64/consts.h b/mlkem/native/aarch64/src/consts.h
similarity index 96%
rename from mlkem/native/aarch64/consts.h
rename to mlkem/native/aarch64/src/consts.h
index 0e834bed4..c40947299 100644
--- a/mlkem/native/aarch64/consts.h
+++ b/mlkem/native/aarch64/src/consts.h
@@ -7,7 +7,7 @@
 #define MLKEM_NATIVE_AARCH64_CONSTS
 
 #include <stdint.h>
-#include "params.h"
+#include "common.h"
 
 #define zetas_mulcache_native MLKEM_NAMESPACE(zetas_mulcache_native)
 extern const int16_t zetas_mulcache_native[256];
diff --git a/mlkem/native/aarch64/intt_clean.S b/mlkem/native/aarch64/src/intt_clean.S
similarity index 98%
rename from mlkem/native/aarch64/intt_clean.S
rename to mlkem/native/aarch64/src/intt_clean.S
index 2ac4e0f15..2f05d8cca 100644
--- a/mlkem/native/aarch64/intt_clean.S
+++ b/mlkem/native/aarch64/src/intt_clean.S
@@ -23,8 +23,8 @@
 /// SOFTWARE.
 ///
 
-#include "config.h"
-#if defined(MLKEM_USE_NATIVE_AARCH64)
+#include "common.h"
+#if defined(MLKEM_NATIVE_ARITH_BACKEND_AARCH64_CLEAN)
 
 // Needed to provide ASM_LOAD directive
 #include "common.i"
@@ -371,4 +371,4 @@ layer012_start:
         pop_stack
         ret
 
-#endif /* MLKEM_USE_NATIVE_AARCH64 */
+#endif /* MLKEM_NATIVE_ARITH_BACKEND_AARCH64_CLEAN */
diff --git a/mlkem/native/aarch64/intt_opt.S b/mlkem/native/aarch64/src/intt_opt.S
similarity index 99%
rename from mlkem/native/aarch64/intt_opt.S
rename to mlkem/native/aarch64/src/intt_opt.S
index 48c412f36..fc720e504 100644
--- a/mlkem/native/aarch64/intt_opt.S
+++ b/mlkem/native/aarch64/src/intt_opt.S
@@ -23,8 +23,8 @@
 /// SOFTWARE.
 ///
 
-#include "config.h"
-#if defined(MLKEM_USE_NATIVE_AARCH64)
+#include "common.h"
+#if defined(MLKEM_NATIVE_ARITH_BACKEND_AARCH64_OPT)
 
 // Needed to provide ASM_LOAD directive
 #include "common.i"
@@ -1027,4 +1027,4 @@ layer012_start:
         pop_stack
         ret
 
-#endif /* MLKEM_USE_NATIVE_AARCH64 */
+#endif /* MLKEM_NATIVE_ARITH_BACKEND_AARCH64_OPT */
diff --git a/mlkem/native/aarch64/ntt_clean.S b/mlkem/native/aarch64/src/ntt_clean.S
similarity index 98%
rename from mlkem/native/aarch64/ntt_clean.S
rename to mlkem/native/aarch64/src/ntt_clean.S
index 4243b9431..ad9086e27 100644
--- a/mlkem/native/aarch64/ntt_clean.S
+++ b/mlkem/native/aarch64/src/ntt_clean.S
@@ -24,8 +24,8 @@
 /// SOFTWARE.
 ///
 
-#include "config.h"
-#if defined(MLKEM_USE_NATIVE_AARCH64)
+#include "common.h"
+#if defined(MLKEM_NATIVE_ARITH_BACKEND_AARCH64_CLEAN)
 
 // Needed to provide ASM_LOAD directive
 #include "common.i"
@@ -295,4 +295,4 @@ layer3456_start:
         pop_stack
         ret
 
-#endif /* MLKEM_USE_NATIVE_AARCH64 */
+#endif /* MLKEM_NATIVE_ARITH_BACKEND_AARCH64_CLEAN */
diff --git a/mlkem/native/aarch64/ntt_opt.S b/mlkem/native/aarch64/src/ntt_opt.S
similarity index 99%
rename from mlkem/native/aarch64/ntt_opt.S
rename to mlkem/native/aarch64/src/ntt_opt.S
index 71779afe9..f4cbc928f 100644
--- a/mlkem/native/aarch64/ntt_opt.S
+++ b/mlkem/native/aarch64/src/ntt_opt.S
@@ -24,8 +24,8 @@
 /// SOFTWARE.
 ///
 
-#include "config.h"
-#if defined(MLKEM_USE_NATIVE_AARCH64)
+#include "common.h"
+#if defined(MLKEM_NATIVE_ARITH_BACKEND_AARCH64_OPT)
 
 // Needed to provide ASM_LOAD directive
 #include "common.i"
@@ -892,4 +892,4 @@ layer3456_start:
         pop_stack
         ret
 
-#endif /* MLKEM_USE_NATIVE_AARCH64 */
+#endif /* MLKEM_NATIVE_ARITH_BACKEND_AARCH64_OPT */
diff --git a/mlkem/native/aarch64/profiles/opt.h b/mlkem/native/aarch64/src/opt_impl.h
similarity index 89%
rename from mlkem/native/aarch64/profiles/opt.h
rename to mlkem/native/aarch64/src/opt_impl.h
index 2414700bf..bc9af0dee 100644
--- a/mlkem/native/aarch64/profiles/opt.h
+++ b/mlkem/native/aarch64/src/opt_impl.h
@@ -5,14 +5,14 @@
 
 /* ML-KEM arithmetic native profile for clean assembly */
 
-#ifdef MLKEM_ARITH_NATIVE_PROFILE_H
+#ifdef MLKEM_NATIVE_ARITH_PROFILE_IMPL_H
 #error Only one MLKEM_ARITH assembly profile can be defined -- did you include multiple profiles?
 #else
-#define MLKEM_ARITH_NATIVE_PROFILE_H
+#define MLKEM_NATIVE_ARITH_PROFILE_IMPL_H
 
-#include "../../arith_native.h"
-#include "../arith_native_aarch64.h"
+#include "arith_native_aarch64.h"
 
+/* Set of primitives that this backend replaces */
 #define MLKEM_USE_NATIVE_NTT
 #define MLKEM_USE_NATIVE_INTT
 #define MLKEM_USE_NATIVE_POLY_REDUCE
@@ -62,7 +62,7 @@ static INLINE void polyvec_basemul_acc_montgomery_cached_native(
 static INLINE void poly_tobytes_native(uint8_t r[MLKEM_POLYBYTES],
                                        const poly *a)
 {
-  poly_tobytes_asm_clean(r, a->coeffs);
+  poly_tobytes_asm_opt(r, a->coeffs);
 }
 
 static INLINE int rej_uniform_native(int16_t *r, unsigned int len,
@@ -75,4 +75,4 @@ static INLINE int rej_uniform_native(int16_t *r, unsigned int len,
   return (int)rej_uniform_asm_clean(r, buf, buflen);
 }
 
-#endif /* MLKEM_ARITH_NATIVE_PROFILE_H */
+#endif /* MLKEM_NATIVE_ARITH_PROFILE_IMPL_H */
diff --git a/mlkem/native/aarch64/optimize.sh b/mlkem/native/aarch64/src/optimize.sh
similarity index 100%
rename from mlkem/native/aarch64/optimize.sh
rename to mlkem/native/aarch64/src/optimize.sh
diff --git a/mlkem/native/aarch64/poly_clean.S b/mlkem/native/aarch64/src/poly_clean.S
similarity index 98%
rename from mlkem/native/aarch64/poly_clean.S
rename to mlkem/native/aarch64/src/poly_clean.S
index 86e4e494a..3e1bc5cf4 100644
--- a/mlkem/native/aarch64/poly_clean.S
+++ b/mlkem/native/aarch64/src/poly_clean.S
@@ -3,8 +3,8 @@
  * SPDX-License-Identifier: Apache-2.0
  */
 
-#include "config.h"
-#if defined(MLKEM_USE_NATIVE_AARCH64)
+#include "common.h"
+#if defined(MLKEM_NATIVE_ARITH_BACKEND_AARCH64_CLEAN)
 
 // Needed to provide ASM_LOAD directive
 #include "common.i"
@@ -285,4 +285,4 @@ poly_tomont_asm_loop:
 
         ret
 
-#endif /* MLKEM_USE_NATIVE_AARCH64 */
+#endif /* MLKEM_NATIVE_ARITH_BACKEND_AARCH64_CLEAN */
diff --git a/mlkem/native/aarch64/poly_opt.S b/mlkem/native/aarch64/src/poly_opt.S
similarity index 99%
rename from mlkem/native/aarch64/poly_opt.S
rename to mlkem/native/aarch64/src/poly_opt.S
index d86595059..df3b21008 100644
--- a/mlkem/native/aarch64/poly_opt.S
+++ b/mlkem/native/aarch64/src/poly_opt.S
@@ -3,8 +3,8 @@
  * SPDX-License-Identifier: Apache-2.0
  */
 
-#include "config.h"
-#if defined(MLKEM_USE_NATIVE_AARCH64)
+#include "common.h"
+#if defined(MLKEM_NATIVE_ARITH_BACKEND_AARCH64_OPT)
 
 // Needed to provide ASM_LOAD directive
 #include "common.i"
@@ -649,4 +649,4 @@ poly_tomont_asm_loop:
 
         ret
 
-#endif /* MLKEM_USE_NATIVE_AARCH64 */
+#endif /* MLKEM_NATIVE_ARITH_BACKEND_AARCH64_OPT */
diff --git a/mlkem/native/aarch64/polyvec_clean.S b/mlkem/native/aarch64/src/polyvec_clean.S
similarity index 98%
rename from mlkem/native/aarch64/polyvec_clean.S
rename to mlkem/native/aarch64/src/polyvec_clean.S
index 59f57d77a..bfd1d2b8a 100644
--- a/mlkem/native/aarch64/polyvec_clean.S
+++ b/mlkem/native/aarch64/src/polyvec_clean.S
@@ -9,8 +9,8 @@
 // https://eprint.iacr.org/2021/986
 // https://github.com/neon-ntt/neon-ntt
 
-#include "config.h"
-#if defined(MLKEM_USE_NATIVE_AARCH64)
+#include "common.h"
+#if defined(MLKEM_NATIVE_ARITH_BACKEND_AARCH64_CLEAN)
 
 #include "params.h"
 
@@ -279,4 +279,4 @@ k4_loop_start:
         ret
 #endif /* MLKEM_K == 4 */
 
-#endif /* MLKEM_USE_NATIVE_AARCH64 */
+#endif /* MLKEM_NATIVE_ARITH_BACKEND_AARCH64_CLEAN */
diff --git a/mlkem/native/aarch64/polyvec_opt.S b/mlkem/native/aarch64/src/polyvec_opt.S
similarity index 99%
rename from mlkem/native/aarch64/polyvec_opt.S
rename to mlkem/native/aarch64/src/polyvec_opt.S
index 9ae035639..07dc98efd 100644
--- a/mlkem/native/aarch64/polyvec_opt.S
+++ b/mlkem/native/aarch64/src/polyvec_opt.S
@@ -6,12 +6,11 @@
 // AArch64 re-implementation of the asymmetric base multiplication from:
 
 // Neon NTT: Faster Dilithium, Kyber, and Saber on Cortex-A72 and Apple M1
-// https:   // eprint.iacr.org/2021/986
-// https:   // github.com/neon-ntt/neon-ntt
-
-#include "config.h"
-#if defined(MLKEM_USE_NATIVE_AARCH64)
+// https://eprint.iacr.org/2021/986
+// https://github.com/neon-ntt/neon-ntt
 
+#include "common.h"
+#if defined(MLKEM_NATIVE_ARITH_BACKEND_AARCH64_OPT)
 #include "params.h"
 
 // Needed to provide ASM_LOAD directive
@@ -1122,4 +1121,4 @@ k4_loop_start:
         ret
 #endif /* MLKEM_K == 4 */
 
-#endif /* MLKEM_USE_NATIVE_AARCH64 */
+#endif /* MLKEM_NATIVE_ARITH_BACKEND_AARCH64_OPT */
diff --git a/mlkem/native/aarch64/rej_uniform_asm_clean.S b/mlkem/native/aarch64/src/rej_uniform_asm_clean.S
similarity index 99%
rename from mlkem/native/aarch64/rej_uniform_asm_clean.S
rename to mlkem/native/aarch64/src/rej_uniform_asm_clean.S
index 00332ec65..c51e53188 100644
--- a/mlkem/native/aarch64/rej_uniform_asm_clean.S
+++ b/mlkem/native/aarch64/src/rej_uniform_asm_clean.S
@@ -18,9 +18,9 @@
  *
  * Returns number of sampled 16-bit integers (at most MLKEM_N).
  **************************************************/
-#include "config.h"
-#include "params.h"
-#if defined(MLKEM_USE_NATIVE_AARCH64)
+#include "common.h"
+#if defined(MLKEM_NATIVE_ARITH_BACKEND_AARCH64_CLEAN) || \
+    defined(MLKEM_NATIVE_ARITH_BACKEND_AARCH64_OPT)
 
 // Needed to provide ASM_LOAD directive
 #include "common.i"
@@ -605,4 +605,5 @@ return:
     pop_stack
     ret
 
-#endif /* MLKEM_USE_NATIVE_AARCH64 */
+#endif /* defined(MLKEM_NATIVE_ARITH_BACKEND_AARCH64_CLEAN) ||
+          defined(MLKEM_NATIVE_ARITH_BACKEND_AARCH64_OPT) */
diff --git a/mlkem/native/arith_native.h b/mlkem/native/api.h
similarity index 98%
rename from mlkem/native/arith_native.h
rename to mlkem/native/api.h
index b7e921323..5a2d92b1d 100644
--- a/mlkem/native/arith_native.h
+++ b/mlkem/native/api.h
@@ -6,14 +6,12 @@
 #define MLKEM_ARITH_NATIVE_H
 
 #include <stdint.h>
-#include "config.h"
-#include "params.h"
-
-#if defined(MLKEM_USE_NATIVE)
 
+#include "cbmc.h"
 #include "poly.h"
 #include "polyvec.h"
-#include "profile.h"
+
+#include "common.h"
 
 /*
  * MLKEM native arithmetic interface
@@ -246,5 +244,4 @@ static INLINE int rej_uniform_native(int16_t *r, unsigned int len,
                                      const uint8_t *buf, unsigned int buflen);
 #endif /* MLKEM_USE_NATIVE_REJ_UNIFORM */
 
-#endif /* MLKEM_USE_NATIVE */
 #endif /* MLKEM_ARITH_NATIVE_H */
diff --git a/mlkem/native/default.h b/mlkem/native/default.h
new file mode 100644
index 000000000..d1e41c52e
--- /dev/null
+++ b/mlkem/native/default.h
@@ -0,0 +1,32 @@
+/*
+ * Copyright (c) 2024 The mlkem-native project authors
+ * SPDX-License-Identifier: Apache-2.0
+ */
+#ifndef MLKEM_NATIVE_ARITH_BACKEND_DEFAULT_H
+#define MLKEM_NATIVE_ARITH_BACKEND_DEFAULT_H
+
+/*
+ * Default arithmetic backend
+ */
+#include "sys.h"
+
+#ifdef SYS_AARCH64
+/*
+ * For AArch64, we currently we have one clean and one opt profile.
+ * We default to the opt profile.
+ *
+ * In the future, this may branch further depending on the microarchitecture.
+ */
+#include "aarch64/opt.h"
+#endif /* SYS_AARCH64 */
+
+#ifdef SYS_X86_64_AVX2
+/*
+ * For now, there's only one x86_64 profile, based on
+ * the AVX2 code from the Kyber repository.
+ * https://github.com/pq-crystals/kyber
+ */
+#include "x86_64/default.h"
+#endif /* SYS_X86_64 */
+
+#endif /* MLKEM_NATIVE_ARITH_BACKEND_DEFAULT_H */
diff --git a/mlkem/native/profile.h b/mlkem/native/profile.h
deleted file mode 100644
index ab9d93aa1..000000000
--- a/mlkem/native/profile.h
+++ /dev/null
@@ -1,73 +0,0 @@
-/*
- * Copyright (c) 2024 The mlkem-native project authors
- * SPDX-License-Identifier: Apache-2.0
- */
-#ifndef MLKEM_ARITH_NATIVE_PROFILE_CHOICE_H
-#define MLKEM_ARITH_NATIVE_PROFILE_CHOICE_H
-
-#include <stdint.h>
-#include "config.h"
-#include "params.h"
-
-#if defined(MLKEM_USE_NATIVE)
-#include "poly.h"
-#include "polyvec.h"
-
-/*
- * MLKEM native arithmetic profile
- *
- * The profile decides which implementation(s) of the arithmetic backend to use.
- *
- * If you don't change anything, the default profile will be used. This profile
- * picks implementations based on characteristics of your system visible to
- * the compiler.
- *
- * If you want to pick a specific profile for your target, there are three ways
- * to do so, in descending order of convenience to the user:
- * 1. Pick one of the profiles shipped with this repository.
- * 2. Provide your own profile and register it via MLKEM_ARITH_NATIVE_PROFILE
- *    (which must be the profile's path relative to this directoru).
- * 3. Set MLKEM_ARITH_NATIVE_MANUAL and use an adhoc profile specified via
- * CFLAGS.
- */
-
-/* Option 2: Manually written profile */
-#if defined(MLKEM_ARITH_NATIVE_PROFILE)
-
-#define STRINGIFY_(x) #x
-#define STRINGIFY(x) STRINGIFY_(x)
-#include STRINGIFY(MLKEM_ARITH_NATIVE_PROFILE)
-
-/* Option 1: Choose from shipped list of profiles */
-#elif !defined(MLKEM_ARITH_NATIVE_MANUAL)
-
-#ifdef SYS_AARCH64
-/*
- * For now, we only have clean and opt profiles.
- * In the future, this is likely to branch further depending
- * on the microarchitecture.
- */
-#if defined(MLKEM_USE_NATIVE_AARCH64_CLEAN)
-#include "aarch64/profiles/clean.h"
-#else /* MLKEM_USE_NATIVE_AARCH64_CLEAN */
-#include "aarch64/profiles/opt.h"
-#endif /* !MLKEM_USE_NATIVE_AARCH64_CLEAN */
-#endif /* SYS_AARCH64 */
-
-#ifdef SYS_X86_64_AVX2
-/*
- * For now, there's only one x86_64 profile, which is essentially
- * the AVX2 code from the Kyber repository
- * https://github.com/pq-crystals/kyber
- */
-#include "x86_64/profiles/default.h"
-#endif /* SYS_X86_64 */
-
-#else /* !MLKEM_ARITH_NATIVE_PROFILE && MLKEM_ARITH_NATIVE_MANUAL */
-
-/* Option 3: Build your own profile here, or via CFLAGS */
-
-#endif /* !MLKEM_ARITH_NATIVE_PROFILE && !MLKEM_ARITH_NATIVE_MANUAL */
-
-#endif /* MLKEM_USE_NATIVE */
-#endif /* MLKEM_ARITH_NATIVE_PROFILE_CHOICE_H */
diff --git a/mlkem/native/x86_64/default.h b/mlkem/native/x86_64/default.h
new file mode 100644
index 000000000..592e8996d
--- /dev/null
+++ b/mlkem/native/x86_64/default.h
@@ -0,0 +1,24 @@
+/*
+ * Copyright (c) 2024 The mlkem-native project authors
+ * SPDX-License-Identifier: Apache-2.0
+ */
+
+/* ML-KEM arithmetic native profile for clean assembly */
+
+#ifdef MLKEM_NATIVE_ARITH_PROFILE_H
+#error Only one MLKEM_ARITH assembly profile can be defined -- did you include multiple profiles?
+#else
+#define MLKEM_NATIVE_ARITH_PROFILE_H
+
+/* Identifier for this backend so that source and assembly files
+ * in the build can be appropriately guarded. */
+#define MLKEM_NATIVE_ARITH_BACKEND_X86_64_DEFAULT
+
+#define MLKEM_NATIVE_ARITH_BACKEND_NAME X86_64_DEFAULT
+
+/* Filename of the C backend implementation.
+ * This is not inlined here because this header is included in assembly
+ * files as well. */
+#define MLKEM_NATIVE_ARITH_BACKEND_IMPL "x86_64/src/default_impl.h"
+
+#endif /* MLKEM_NATIVE_ARITH_PROFILE_H */
diff --git a/mlkem/native/x86_64/align.h b/mlkem/native/x86_64/src/align.h
similarity index 100%
rename from mlkem/native/x86_64/align.h
rename to mlkem/native/x86_64/src/align.h
diff --git a/mlkem/native/x86_64/arith_native_x86_64.h b/mlkem/native/x86_64/src/arith_native_x86_64.h
similarity index 91%
rename from mlkem/native/x86_64/arith_native_x86_64.h
rename to mlkem/native/x86_64/src/arith_native_x86_64.h
index b346e1eb2..4b78c004a 100644
--- a/mlkem/native/x86_64/arith_native_x86_64.h
+++ b/mlkem/native/x86_64/src/arith_native_x86_64.h
@@ -5,16 +5,13 @@
 #ifndef MLKEM_X86_64_NATIVE_H
 #define MLKEM_X86_64_NATIVE_H
 
-#include <stdint.h>
-#include "config.h"
-#include "fips202.h"
-#include "params.h"
-#include "polyvec.h"
-
-#if defined(MLKEM_USE_NATIVE_X86_64) && defined(SYS_X86_64_AVX2)
+#include "common.h"
 
 #include <immintrin.h>
 #include <stdint.h>
+#include "consts.h"
+#include "fips202.h"
+#include "polyvec.h"
 
 #define REJ_UNIFORM_AVX_NBLOCKS \
   ((12 * MLKEM_N / 8 * (1 << 12) / MLKEM_Q + SHAKE128_RATE) / SHAKE128_RATE)
@@ -57,6 +54,4 @@ void nttfrombytes_avx2(__m256i *r, const uint8_t *a, const __m256i *qdata);
 #define tomont_avx2 MLKEM_NAMESPACE(tomont_avx2)
 void tomont_avx2(__m256i *r, const __m256i *qdata);
 
-#endif /* MLKEM_USE_NATIVE_X86_64 && SYS_X86_64_AVX2 */
-
 #endif /* MLKEM_X86_64_NATIVE_H */
diff --git a/mlkem/native/x86_64/basemul.S b/mlkem/native/x86_64/src/basemul.S
similarity index 96%
rename from mlkem/native/x86_64/basemul.S
rename to mlkem/native/x86_64/src/basemul.S
index 797b0f46a..503fbeb51 100644
--- a/mlkem/native/x86_64/basemul.S
+++ b/mlkem/native/x86_64/src/basemul.S
@@ -6,9 +6,8 @@
 // Implementation from Kyber reference repository
 // https://github.com/pq-crystals/kyber/blob/main/avx2
 
-#include "config.h"
-
-#if defined(MLKEM_USE_NATIVE_X86_64) && defined(SYS_X86_64_AVX2)
+#include "common.h"
+#if defined(MLKEM_NATIVE_ARITH_BACKEND_X86_64_DEFAULT)
 
 #include "consts.h"
 #include "params.h"
@@ -135,4 +134,4 @@ schoolbook	3
 mov		%r8,%rsp
 ret
 
-#endif /* MLKEM_USE_NATIVE_X86_64 && SYS_X86_64_AVX2 */
+#endif /* MLKEM_NATIVE_ARITH_BACKEND_X86_64_DEFAULT */
diff --git a/mlkem/native/x86_64/basemul.c b/mlkem/native/x86_64/src/basemul.c
similarity index 90%
rename from mlkem/native/x86_64/basemul.c
rename to mlkem/native/x86_64/src/basemul.c
index c9dba813f..3f1653ed3 100644
--- a/mlkem/native/x86_64/basemul.c
+++ b/mlkem/native/x86_64/src/basemul.c
@@ -3,15 +3,16 @@
  * SPDX-License-Identifier: Apache-2.0
  */
 
-#include "config.h"
+#include "common.h"
 
-#if defined(MLKEM_USE_NATIVE_X86_64) && defined(SYS_X86_64_AVX2)
+#if defined(MLKEM_NATIVE_ARITH_BACKEND_X86_64_DEFAULT)
 
-#include "arith_native_x86_64.h"
 #include "consts.h"
 #include "poly.h"
 #include "polyvec.h"
 
+#include "arith_native_x86_64.h"
+
 static void poly_basemul_montgomery_avx2(poly *r, const poly *a, const poly *b)
 {
   basemul_avx2((__m256i *)r->coeffs, (const __m256i *)a->coeffs,
@@ -57,12 +58,11 @@ void polyvec_basemul_acc_montgomery_cached_avx2(poly *r, const polyvec *a,
   }
 }
 
-#else
-#include "params.h"
+#else /* MLKEM_NATIVE_ARITH_BACKEND_X86_64_DEFAULT */
 
 /* Dummy constant to keep compiler happy despite empty CU */
 
 #define empty_cu_avx2_basemul MLKEM_NAMESPACE(empty_cu_avx2_basemul)
 int empty_cu_avx2_basemul;
 
-#endif /* MLKEM_USE_NATIVE_X86_64 && SYS_X86_64_AVX2 */
+#endif /* MLKEM_NATIVE_ARITH_BACKEND_X86_64_DEFAULT */
diff --git a/mlkem/native/x86_64/consts.c b/mlkem/native/x86_64/src/consts.c
similarity index 92%
rename from mlkem/native/x86_64/consts.c
rename to mlkem/native/x86_64/src/consts.c
index 573eabb64..9b6d0993e 100644
--- a/mlkem/native/x86_64/consts.c
+++ b/mlkem/native/x86_64/src/consts.c
@@ -8,13 +8,12 @@
  * https://github.com/pq-crystals/kyber/blob/main/avx2/consts.c
  */
 
-#include "config.h"
+#include "common.h"
 
-#if defined(MLKEM_USE_NATIVE_X86_64) && defined(SYS_X86_64_AVX2)
+#if defined(MLKEM_NATIVE_ARITH_BACKEND_X86_64_DEFAULT)
 
 #include "align.h"
 #include "consts.h"
-#include "params.h"
 
 #define Q MLKEM_Q
 #define MONT -1044     /* 2^16 mod q */
@@ -86,10 +85,9 @@ const qdata_t qdata = {{
     SHIFT,    SHIFT,    SHIFT,    SHIFT,    SHIFT,    SHIFT,
     SHIFT,    SHIFT,    SHIFT,    SHIFT}};
 
-#else /* MLKEM_USE_NATIVE_X86_64 && SYS_X86_64_AVX2 */
-#include "params.h"
+#else /* MLKEM_NATIVE_ARITH_BACKEND_X86_64_DEFAULT */
 
 /* Dummy declaration for compilers disliking empty compilation units */
 #define empty_cu_consts MLKEM_NAMESPACE(empty_cu_consts)
 int empty_cu_consts;
-#endif /* MLKEM_USE_NATIVE_X86_64 && SYS_X86_64_AVX2 */
+#endif /* MLKEM_NATIVE_ARITH_BACKEND_X86_64_DEFAULT */
diff --git a/mlkem/native/x86_64/consts.h b/mlkem/native/x86_64/src/consts.h
similarity index 97%
rename from mlkem/native/x86_64/consts.h
rename to mlkem/native/x86_64/src/consts.h
index 8ca98bbf8..a9f0c484f 100644
--- a/mlkem/native/x86_64/consts.h
+++ b/mlkem/native/x86_64/src/consts.h
@@ -11,7 +11,7 @@
 #ifndef CONSTS_H
 #define CONSTS_H
 
-#include "params.h"
+#include "common.h"
 
 #define _16XQ 0
 #define _16XQINV 16
diff --git a/mlkem/native/x86_64/profiles/default.h b/mlkem/native/x86_64/src/default_impl.h
similarity index 92%
rename from mlkem/native/x86_64/profiles/default.h
rename to mlkem/native/x86_64/src/default_impl.h
index 55eb0ad01..4f3cd2c9a 100644
--- a/mlkem/native/x86_64/profiles/default.h
+++ b/mlkem/native/x86_64/src/default_impl.h
@@ -5,17 +5,14 @@
 
 /* ML-KEM arithmetic native profile for clean assembly */
 
-#ifdef MLKEM_ARITH_NATIVE_PROFILE_H
+#ifdef MLKEM_NATIVE_ARITH_PROFILE_IMPL_H
 #error Only one MLKEM_ARITH assembly profile can be defined -- did you include multiple profiles?
 #else
-#define MLKEM_ARITH_NATIVE_PROFILE_H
+#define MLKEM_NATIVE_ARITH_PROFILE_IMPL_H
 
 #include <string.h>
 
-#include "../../arith_native.h"
-#include "../arith_native_x86_64.h"
-#include "../consts.h"
-
+#include "arith_native_x86_64.h"
 #include "poly.h"
 
 #define MLKEM_USE_NATIVE_NTT_CUSTOM_ORDER
@@ -96,4 +93,4 @@ static INLINE void poly_frombytes_native(poly *r,
   nttfrombytes_avx2((__m256i *)r->coeffs, a, qdata.vec);
 }
 
-#endif /* MLKEM_ARITH_NATIVE_PROFILE_H */
+#endif /* MLKEM_NATIVE_ARITH_PROFILE_IMPL_H */
diff --git a/mlkem/native/x86_64/fq.S b/mlkem/native/x86_64/src/fq.S
similarity index 91%
rename from mlkem/native/x86_64/fq.S
rename to mlkem/native/x86_64/src/fq.S
index 00fb093c5..50ef190b7 100644
--- a/mlkem/native/x86_64/fq.S
+++ b/mlkem/native/x86_64/src/fq.S
@@ -11,13 +11,12 @@
 //   in [0,1,...,q-1] rather than [0,1,...,q], matching the
 //   semantics of poly_reduce().
 
-#include "config.h"
-#include "params.h"
+#include "common.h"
 
-#if defined(MLKEM_USE_NATIVE_X86_64) && defined(SYS_X86_64_AVX2)
+#if defined(MLKEM_NATIVE_ARITH_BACKEND_X86_64_DEFAULT)
 #include "consts.h"
 
-.include "fq.inc"
+#include "fq.inc"
 
 .text
 reduce128_avx2:
@@ -115,4 +114,4 @@ add		$256,%rdi
 call		tomont128_avx2
 ret
 
-#endif /* MLKEM_USE_NATIVE_X86_64 && SYS_X86_64_AVX2 */
+#endif /* MLKEM_NATIVE_ARITH_BACKEND_X86_64_DEFAULT */
diff --git a/mlkem/native/x86_64/fq.inc b/mlkem/native/x86_64/src/fq.inc
similarity index 100%
rename from mlkem/native/x86_64/fq.inc
rename to mlkem/native/x86_64/src/fq.inc
diff --git a/mlkem/native/x86_64/intt.S b/mlkem/native/x86_64/src/intt.S
similarity index 97%
rename from mlkem/native/x86_64/intt.S
rename to mlkem/native/x86_64/src/intt.S
index 8418cc471..4860985ed 100644
--- a/mlkem/native/x86_64/intt.S
+++ b/mlkem/native/x86_64/src/intt.S
@@ -8,14 +8,15 @@
  *
  * Changes to placement of modular reductions have
  * been made to simplify reasoning of non-overflow */
-#include "config.h"
 
-#if defined(MLKEM_USE_NATIVE_X86_64) && defined(SYS_X86_64_AVX2)
+#include "common.h"
+
+#if defined(MLKEM_NATIVE_ARITH_BACKEND_X86_64_DEFAULT)
 
 #include "consts.h"
 #include "params.h"
-.include "shuffle.inc"
-.include "fq.inc"
+#include "shuffle.inc"
+#include "fq.inc"
 
 /* Compute four GS butterflies between rh{0,1,2,3} and rl{0,1,2,3}.
  * Butterflies 0,1 use root zh0 and twisted root zl0, and butterflies
@@ -252,4 +253,4 @@ intt_level6	0
 intt_level6	1
 ret
 
-#endif /* MLKEM_USE_NATIVE_X86_64 && SYS_X86_64_AVX2 */
+#endif /* MLKEM_NATIVE_ARITH_BACKEND_X86_64_DEFAULT */
diff --git a/mlkem/native/x86_64/ntt.S b/mlkem/native/x86_64/src/ntt.S
similarity index 97%
rename from mlkem/native/x86_64/ntt.S
rename to mlkem/native/x86_64/src/ntt.S
index f159faf4f..a0b6f734c 100644
--- a/mlkem/native/x86_64/ntt.S
+++ b/mlkem/native/x86_64/src/ntt.S
@@ -6,14 +6,13 @@
 // Implementation from Kyber reference repository
 // https://github.com/pq-crystals/kyber/blob/main/avx2
 
-#include "config.h"
-
-#if defined(MLKEM_USE_NATIVE_X86_64) && defined(SYS_X86_64_AVX2)
+#include "common.h"
+#if defined(MLKEM_NATIVE_ARITH_BACKEND_X86_64_DEFAULT)
 
 #include "consts.h"
 #include "params.h"
 
-.include "shuffle.inc"
+#include "shuffle.inc"
 
 /* Compute steps 1,2 / 3 of Montgomery multiplication */
 .macro mul rh0,rh1,rh2,rh3,zl0=15,zl1=15,zh0=2,zh1=2
@@ -219,4 +218,4 @@ levels1t6	1
 
 ret
 
-#endif /* MLKEM_USE_NATIVE_X86_64 && SYS_X86_64_AVX2 */
+#endif /* MLKEM_NATIVE_ARITH_BACKEND_X86_64_DEFAULT */
diff --git a/mlkem/native/x86_64/rej_uniform_avx2.c b/mlkem/native/x86_64/src/rej_uniform_avx2.c
similarity index 98%
rename from mlkem/native/x86_64/rej_uniform_avx2.c
rename to mlkem/native/x86_64/src/rej_uniform_avx2.c
index c6e663f7d..c3c8b8104 100644
--- a/mlkem/native/x86_64/rej_uniform_avx2.c
+++ b/mlkem/native/x86_64/src/rej_uniform_avx2.c
@@ -8,18 +8,15 @@
  * https://github.com/pq-crystals/kyber/blob/main/avx2
  */
 
-#include "config.h"
+#include "common.h"
 
-#if defined(MLKEM_USE_NATIVE_X86_64) && defined(SYS_X86_64_AVX2)
-
-#include "arith_native_x86_64.h"
+#if defined(MLKEM_NATIVE_ARITH_BACKEND_X86_64_DEFAULT)
 
 #include <immintrin.h>
 #include <stdint.h>
 #include <string.h>
+#include "arith_native_x86_64.h"
 #include "consts.h"
-#include "params.h"
-
 /* #define BMI */
 
 #ifndef BMI
@@ -290,10 +287,9 @@ unsigned int rej_uniform_avx2(int16_t *RESTRICT r, const uint8_t *buf)
   return ctr;
 }
 
-#else /* MLKEM_USE_NATIVE_X86_64 && SYS_X86_64_AVX2 */
-#include "params.h"
+#else /* MLKEM_NATIVE_ARITH_BACKEND_X86_64_DEFAULT */
 
 /* Dummy declaration for compilers disliking empty compilation units */
 #define empty_cu_rej_uniform_avx2 MLKEM_NAMESPACE(empty_cu_rej_uniform_avx2)
 int empty_cu_rej_uniform_avx2;
-#endif /* MLKEM_USE_NATIVE_X86_64 && SYS_X86_64_AVX2 */
+#endif /* MLKEM_NATIVE_ARITH_BACKEND_X86_64_DEFAULT */
diff --git a/mlkem/native/x86_64/shuffle.S b/mlkem/native/x86_64/src/shuffle.S
similarity index 96%
rename from mlkem/native/x86_64/shuffle.S
rename to mlkem/native/x86_64/src/shuffle.S
index dd1243121..34f6b30b0 100644
--- a/mlkem/native/x86_64/shuffle.S
+++ b/mlkem/native/x86_64/src/shuffle.S
@@ -6,14 +6,14 @@
 // Implementation from Kyber reference repository
 // https://github.com/pq-crystals/kyber/blob/main/avx2
 
-#include "config.h"
+#include "common.h"
 
-#if defined(MLKEM_USE_NATIVE_X86_64) && defined(SYS_X86_64_AVX2)
+#if defined(MLKEM_NATIVE_ARITH_BACKEND_X86_64_DEFAULT)
 
 #include "consts.h"
 #include "params.h"
-.include "fq.inc"
-.include "shuffle.inc"
+#include "fq.inc"
+#include "shuffle.inc"
 
 .global MLKEM_NAMESPACE(nttpack_avx2)
 MLKEM_NAMESPACE(nttpack_avx2):
@@ -255,4 +255,4 @@ add		$192,%rsi
 call		nttfrombytes128_avx
 ret
 
-#endif /* MLKEM_USE_NATIVE_X86_64 && SYS_X86_64_AVX2 */
+#endif /* MLKEM_NATIVE_ARITH_BACKEND_X86_64_DEFAULT */
diff --git a/mlkem/native/x86_64/shuffle.inc b/mlkem/native/x86_64/src/shuffle.inc
similarity index 100%
rename from mlkem/native/x86_64/shuffle.inc
rename to mlkem/native/x86_64/src/shuffle.inc
diff --git a/mlkem/native/x86_64/x86_64_zetas.i b/mlkem/native/x86_64/src/x86_64_zetas.i
similarity index 100%
rename from mlkem/native/x86_64/x86_64_zetas.i
rename to mlkem/native/x86_64/src/x86_64_zetas.i
diff --git a/mlkem/ntt.c b/mlkem/ntt.c
index 8de1c5446..178e8467c 100644
--- a/mlkem/ntt.c
+++ b/mlkem/ntt.c
@@ -2,14 +2,12 @@
  * Copyright (c) 2024 The mlkem-native project authors
  * SPDX-License-Identifier: Apache-2.0
  */
-#include "ntt.h"
 #include <stdint.h>
-#include "params.h"
-#include "reduce.h"
 
-#include "arith_native.h"
+#include "arith_backend.h"
 #include "debug/debug.h"
 #include "ntt.h"
+#include "reduce.h"
 
 #if !defined(MLKEM_USE_NATIVE_NTT)
 /*
diff --git a/mlkem/ntt.h b/mlkem/ntt.h
index 68c9e6724..efa38ecc9 100644
--- a/mlkem/ntt.h
+++ b/mlkem/ntt.h
@@ -6,9 +6,8 @@
 #define NTT_H
 
 #include <stdint.h>
-#include "arith_native.h"
 #include "cbmc.h"
-#include "params.h"
+#include "common.h"
 #include "poly.h"
 #include "reduce.h"
 
diff --git a/mlkem/params.h b/mlkem/params.h
index 85797f067..586c31d33 100644
--- a/mlkem/params.h
+++ b/mlkem/params.h
@@ -5,47 +5,12 @@
 #ifndef PARAMS_H
 #define PARAMS_H
 
-#include "common.h"
 #include "config.h"
-#include "cpucap.h"
 
-#ifndef MLKEM_K
-#define MLKEM_K 3 /* Change this for different security strengths */
+#if !defined(MLKEM_K)
+#error MLKEM_K is not defined
 #endif
 
-#if defined(MLKEM_USE_NATIVE_X86_64)
-#define MLKEM_NATIVE_BACKEND X86_64
-#elif defined(MLKEM_USE_NATIVE_AARCH64)
-#define MLKEM_NATIVE_BACKEND AARCH64
-#else
-#define MLKEM_NATIVE_BACKEND C
-#endif
-
-/* Don't change parameters below this line */
-#if (MLKEM_K == 2)
-#define MLKEM_PARAM_NAME MLKEM512
-#elif (MLKEM_K == 3)
-#define MLKEM_PARAM_NAME MLKEM768
-#elif (MLKEM_K == 4)
-#define MLKEM_PARAM_NAME MLKEM1024
-#else
-#error "MLKEM_K must be in {2,3,4}"
-#endif
-
-#define ___MLKEM_NAMESPACE(x1, x2, x3, x4) x1##_##x2##_##x3##_##x4
-#define __MLKEM_NAMESPACE(x1, x2, x3, x4) ___MLKEM_NAMESPACE(x1, x2, x3, x4)
-
-/*
- * NAMESPACE is PQCP_MLKEM_NATIVE_<PARAM_NAME>_<BACKEND>_
- * e.g., PQCP_MLKEM_NATIVE_MLKEM512_AARCH64_
- */
-#define MLKEM_NAMESPACE(s)                                                     \
-  __MLKEM_NAMESPACE(PQCP_MLKEM_NATIVE, MLKEM_PARAM_NAME, MLKEM_NATIVE_BACKEND, \
-                    s)
-#define _MLKEM_NAMESPACE(s)                               \
-  __MLKEM_NAMESPACE(_PQCP_MLKEM_NATIVE, MLKEM_PARAM_NAME, \
-                    MLKEM_NATIVE_BACKEND, s)
-
 #define MLKEM_N 256
 #define MLKEM_Q 3329
 #define UINT12_MAX 4095
diff --git a/mlkem/poly.c b/mlkem/poly.c
index fdc0ef4af..db7d64ebf 100644
--- a/mlkem/poly.c
+++ b/mlkem/poly.c
@@ -2,21 +2,20 @@
  * Copyright (c) 2024 The mlkem-native project authors
  * SPDX-License-Identifier: Apache-2.0
  */
-#include "poly.h"
 #include <stdint.h>
 #include <string.h>
+
+#include "arith_backend.h"
 #include "cbd.h"
 #include "cbmc.h"
+#include "debug/debug.h"
 #include "fips202x4.h"
 #include "ntt.h"
-#include "params.h"
+#include "poly.h"
 #include "reduce.h"
 #include "symmetric.h"
 #include "verify.h"
 
-#include "arith_native.h"
-#include "debug/debug.h"
-
 void poly_compress_du(uint8_t r[MLKEM_POLYCOMPRESSEDBYTES_DU], const poly *a)
 {
   int j;
diff --git a/mlkem/poly.h b/mlkem/poly.h
index 6d119539d..19cf7b96b 100644
--- a/mlkem/poly.h
+++ b/mlkem/poly.h
@@ -8,7 +8,7 @@
 #include <stddef.h>
 #include <stdint.h>
 #include "cbmc.h"
-#include "params.h"
+#include "common.h"
 #include "reduce.h"
 #include "verify.h"
 
diff --git a/mlkem/polyvec.c b/mlkem/polyvec.c
index 7f268ee50..72277a626 100644
--- a/mlkem/polyvec.c
+++ b/mlkem/polyvec.c
@@ -4,10 +4,9 @@
  */
 #include "polyvec.h"
 #include <stdint.h>
-#include "arith_native.h"
+#include "arith_backend.h"
 #include "config.h"
 #include "ntt.h"
-#include "params.h"
 #include "poly.h"
 
 #include "debug/debug.h"
diff --git a/mlkem/polyvec.h b/mlkem/polyvec.h
index a4a03bf6d..cd90734fa 100644
--- a/mlkem/polyvec.h
+++ b/mlkem/polyvec.h
@@ -6,7 +6,7 @@
 #define POLYVEC_H
 
 #include <stdint.h>
-#include "params.h"
+#include "common.h"
 #include "poly.h"
 
 typedef struct
diff --git a/mlkem/reduce.h b/mlkem/reduce.h
index 39e85588a..515f706fa 100644
--- a/mlkem/reduce.h
+++ b/mlkem/reduce.h
@@ -7,8 +7,8 @@
 
 #include <stdint.h>
 #include "cbmc.h"
+#include "common.h"
 #include "debug/debug.h"
-#include "params.h"
 
 #define HALF_Q ((MLKEM_Q + 1) / 2) /* 1665 */
 
diff --git a/mlkem/rej_uniform.c b/mlkem/rej_uniform.c
index 4e8a5ce9b..88102b4d7 100644
--- a/mlkem/rej_uniform.c
+++ b/mlkem/rej_uniform.c
@@ -2,10 +2,9 @@
  * Copyright (c) 2024 The mlkem-native project authors
  * SPDX-License-Identifier: Apache-2.0
  */
-#include "params.h"
 
-#include "arith_native.h"
 #include "rej_uniform.h"
+#include "arith_backend.h"
 
 /*************************************************
  * Name:        rej_uniform_scalar
diff --git a/mlkem/rej_uniform.h b/mlkem/rej_uniform.h
index 7b86fa63e..e422f73cf 100644
--- a/mlkem/rej_uniform.h
+++ b/mlkem/rej_uniform.h
@@ -8,7 +8,7 @@
 #include <stdint.h>
 #include <stdlib.h>
 #include "cbmc.h"
-#include "params.h"
+#include "common.h"
 
 #define rej_uniform MLKEM_NAMESPACE(rej_uniform)
 /*************************************************
diff --git a/mlkem/symmetric.h b/mlkem/symmetric.h
index 4eb6aaba6..55ebbbd53 100644
--- a/mlkem/symmetric.h
+++ b/mlkem/symmetric.h
@@ -7,11 +7,9 @@
 
 #include <stddef.h>
 #include <stdint.h>
-#include "params.h"
-
-#include "fips202.h"
-
 #include "cbmc.h"
+#include "common.h"
+#include "fips202.h"
 
 /* Macros denoting FIPS-203 specific Hash functions */
 
diff --git a/mlkem/sys/cpucap.h b/mlkem/sys.h
similarity index 53%
rename from mlkem/sys/cpucap.h
rename to mlkem/sys.h
index 7c3a1cdec..be3070dc2 100644
--- a/mlkem/sys/cpucap.h
+++ b/mlkem/sys.h
@@ -2,9 +2,8 @@
  * Copyright (c) 2024 The mlkem-native project authors
  * SPDX-License-Identifier: Apache-2.0
  */
-
-#ifndef CPUCAP_H
-#define CPUCAP_H
+#ifndef MLKEM_NATIVE_SYS_H
+#define MLKEM_NATIVE_SYS_H
 
 /* Check if we're running on an AArch64 little endian system. _M_ARM64 is set by
  * MSVC. */
@@ -52,4 +51,56 @@
 #error "FORCE_X86_64 is set, but we don't seem to be on an X86_64 system."
 #endif
 
+/*
+ * C90 does not have the inline compiler directive yet.
+ * We don't use it in C90 builds.
+ * However, in that case the compiler warns about some inline functions in
+ * header files not being used in every compilation unit that includes that
+ * header. To work around it we silence that warning in that case using
+ * __attribute__((unused)).
+ */
+
+/* Do not use inline for C90 builds*/
+#if !defined(inline)
+#if defined(__STDC_VERSION__) && __STDC_VERSION__ >= 199901L
+#define INLINE inline
+#define ALWAYS_INLINE __attribute__((always_inline))
+#elif defined(_MSC_VER)
+#define INLINE __inline
+#define ALWAYS_INLINE __forceinline
+#else
+#define INLINE __attribute__((unused))
+#define ALWAYS_INLINE
+#endif
+
+#else
+#define INLINE inline
+#define ALWAYS_INLINE __attribute__((always_inline))
+#endif
+
+/*
+ * C90 does not have the restrict compiler directive yet.
+ * We don't use it in C90 builds.
+ */
+#if !defined(restrict)
+#if defined(__STDC_VERSION__) && __STDC_VERSION__ >= 199901L
+#define RESTRICT restrict
+#else
+#define RESTRICT
+#endif
+
+#else
+
+#define RESTRICT restrict
+#endif
+
+#define DEFAULT_ALIGN 32
+#if defined(_WIN32)
+#define ALIGN __declspec(align(DEFAULT_ALIGN))
+#define asm __asm
+#else
+#define asm __asm__
+#define ALIGN __attribute__((aligned(DEFAULT_ALIGN)))
 #endif
+
+#endif /* MLKEM_NATIVE_SYS_H */
diff --git a/mlkem/sys/config.h b/mlkem/sys/config.h
deleted file mode 100644
index 370a141a6..000000000
--- a/mlkem/sys/config.h
+++ /dev/null
@@ -1,22 +0,0 @@
-/*
- * Copyright (c) 2024 The mlkem-native project authors
- * SPDX-License-Identifier: Apache-2.0
- */
-
-#ifndef CONFIG_H
-#define CONFIG_H
-
-#include "cpucap.h"
-
-#if defined(MLKEM_USE_NATIVE)
-
-#if defined(SYS_AARCH64)
-#define MLKEM_USE_NATIVE_AARCH64
-#endif /* SYS_AARCH64 */
-
-#if defined(SYS_X86_64)
-#define MLKEM_USE_NATIVE_X86_64
-#endif /* SYS_X86_64 */
-
-#endif /* MLKEM_USE_NATIVE */
-#endif /* CONFIG_H */
diff --git a/mlkem/verify.c b/mlkem/verify.c
index e616a8bab..b7078fcc1 100644
--- a/mlkem/verify.c
+++ b/mlkem/verify.c
@@ -13,7 +13,6 @@
 volatile uint64_t ct_opt_blocker_u64 = 0;
 
 #else /* MLKEM_USE_ASM_VALUE_BARRIER */
-#include "params.h"
 
 #define empty_cu_verify MLKEM_NAMESPACE(empty_cu_verify)
 int empty_cu_verify;
diff --git a/scripts/autogenerate_files.py b/scripts/autogenerate_files.py
index a33d3cdfe..70d37d905 100644
--- a/scripts/autogenerate_files.py
+++ b/scripts/autogenerate_files.py
@@ -295,9 +295,13 @@ def gen_aarch64_mulcache_twiddles_twisted():
 def gen_aarch64_fwd_ntt_zeta_file(dry_run=False):
     def gen():
         yield from gen_header()
-        yield '#include "arith_native_aarch64.h"'
+        yield '#include "common.h"'
+        yield ""
+        yield "#if defined(MLKEM_NATIVE_ARITH_BACKEND_AARCH64_CLEAN) || \\"
+        yield "    defined(MLKEM_NATIVE_ARITH_BACKEND_AARCH64_OPT)"
         yield ""
-        yield "#ifdef MLKEM_USE_NATIVE_AARCH64"
+        yield "#include <stdint.h>"
+        yield '#include "arith_native_aarch64.h"'
         yield ""
         yield "/*"
         yield " * Table of zeta values used in the AArch64 forward NTT"
@@ -327,17 +331,17 @@ def gen():
         yield from map(lambda t: str(t) + ",", gen_aarch64_mulcache_twiddles_twisted())
         yield "};"
         yield ""
-        yield "#else /* MLKEM_USE_NATIVE_AARCH64 */"
+        yield "#else"
         yield '#include "params.h"'
         yield ""
         yield "/* Dummy declaration for compilers disliking empty compilation units */"
         yield "#define empty_cu_aarch64_zetas MLKEM_NAMESPACE(empty_cu_aarch64_zetas)"
         yield "int empty_cu_aarch64_zetas;"
-        yield "#endif /* MLKEM_USE_NATIVE_AARCH64 */"
+        yield "#endif"
         yield ""
 
     update_file(
-        "mlkem/native/aarch64/aarch64_zetas.c", "\n".join(gen()), dry_run=dry_run
+        "mlkem/native/aarch64/src/aarch64_zetas.c", "\n".join(gen()), dry_run=dry_run
     )
 
 
@@ -414,7 +418,9 @@ def gen():
         yield from map(lambda t: str(t) + ",", gen_avx2_fwd_ntt_zetas())
         yield ""
 
-    update_file("mlkem/native/x86_64/x86_64_zetas.i", "\n".join(gen()), dry_run=dry_run)
+    update_file(
+        "mlkem/native/x86_64/src/x86_64_zetas.i", "\n".join(gen()), dry_run=dry_run
+    )
 
 
 def _main():
diff --git a/scripts/lib/mlkem_test.py b/scripts/lib/mlkem_test.py
index 796c4953a..73d24122a 100644
--- a/scripts/lib/mlkem_test.py
+++ b/scripts/lib/mlkem_test.py
@@ -98,9 +98,7 @@ def dict2str(dict):
             f"{self.test_type}",
         ] + extra_make_args
 
-        make_envs = (
-            {"CFLAGS": f"{self.cflags}"} if self.cflags is not None else {}
-        ) | (
+        make_envs = ({"CFLAGS": self.cflags} if self.cflags is not None else {}) | (
             {"ARCH_FLAGS": f"{self.arch_flags}"} if self.arch_flags is not None else {}
         )
         extra_make_envs.update(make_envs)
diff --git a/test/bench_components_mlkem.c b/test/bench_components_mlkem.c
index c922cb76d..b3956171c 100644
--- a/test/bench_components_mlkem.c
+++ b/test/bench_components_mlkem.c
@@ -12,7 +12,7 @@
 #include "randombytes.h"
 #include "rej_uniform.h"
 
-#include "../mlkem/native/arith_native.h"
+#include "../mlkem/arith_backend.h"
 #include "fips202.h"
 #include "indcpa.h"
 #include "keccakf1600.h"
@@ -184,7 +184,7 @@ static int bench(void)
   BENCH("gen_matrix", gen_matrix((polyvec *)data0, (uint8_t *)data1, 0))
 
 
-#if defined(MLKEM_USE_NATIVE_AARCH64)
+#if defined(MLKEM_NATIVE_ARITH_BACKEND_AARCH64_CLEAN)
   BENCH("ntt-clean",
         ntt_asm_clean((int16_t *)data0, (int16_t *)data1, (int16_t *)data2));
   BENCH("intt-clean",
@@ -200,7 +200,9 @@ static int bench(void)
         polyvec_basemul_acc_montgomery_cached_asm_clean(
             (int16_t *)data0, (int16_t *)data1, (int16_t *)data2,
             (int16_t *)data3));
+#endif /* MLKEM_NATIVE_ARITH_BACKEND_AARCH64_CLEAN */
 
+#if defined(MLKEM_NATIVE_ARITH_BACKEND_AARCH64_OPT)
   BENCH("ntt-opt",
         ntt_asm_opt((int16_t *)data0, (int16_t *)data1, (int16_t *)data2));
   BENCH("intt-opt",
@@ -214,7 +216,7 @@ static int bench(void)
         polyvec_basemul_acc_montgomery_cached_asm_opt(
             (int16_t *)data0, (int16_t *)data1, (int16_t *)data2,
             (int16_t *)data3));
-#endif
+#endif /* MLKEM_NATIVE_ARITH_BACKEND_AARCH64_OPT */
 
   return 0;
 }