diff --git a/scripts/tools/benchmark_gemv.py b/csrc/bench/benchmark_gemv.py
similarity index 100%
rename from scripts/tools/benchmark_gemv.py
rename to csrc/bench/benchmark_gemv.py
diff --git a/scripts/tools/benchmark_gemv_common.py b/csrc/bench/benchmark_gemv_common.py
similarity index 100%
rename from scripts/tools/benchmark_gemv_common.py
rename to csrc/bench/benchmark_gemv_common.py
diff --git a/scripts/tools/benchmark_swiglu.py b/csrc/bench/benchmark_swiglu.py
similarity index 100%
rename from scripts/tools/benchmark_swiglu.py
rename to csrc/bench/benchmark_swiglu.py
diff --git a/docs/developer/cuda_kernels.md b/docs/developer/cuda_kernels.md
index 7b5f7ed..7704938 100644
--- a/docs/developer/cuda_kernels.md
+++ b/docs/developer/cuda_kernels.md
@@ -84,7 +84,7 @@ OPT 1.3B M=1 is +4.54%. Qwen2 and LLaMA 3 70B M=1, and all three new
 families at M=8, remain exact PyTorch fallbacks.
 
 These are synthetic projection-chain measurements, not whole-model throughput
-claims. Reproduce them with `scripts/tools/benchmark_gemv_common.py`.
+claims. Reproduce them with `csrc/bench/benchmark_gemv_common.py`.
 
 The AstrAI 1B A→B→B→A results use the real `InferenceEngine`, including scheduler,
 sampling, and CUDA Graph. M=8 stays on PyTorch because its remaining
diff --git a/docs/developer/decode_linear_benchmark.md b/docs/developer/decode_linear_benchmark.md
index 7119a9c..7af4746 100644
--- a/docs/developer/decode_linear_benchmark.md
+++ b/docs/developer/decode_linear_benchmark.md
@@ -1,6 +1,6 @@
 # Decode linear shape benchmark
 
-`scripts/tools/benchmark_gemv.py` records the `F.linear` baseline used to decide
+`csrc/bench/benchmark_gemv.py` records the `F.linear` baseline used to decide
 whether a BF16 GEMV or small-M kernel should enter automatic inference dispatch.
 It does not change model execution or select a custom kernel.
 
@@ -10,7 +10,7 @@ modes. Results include device-event latency samples, p50/p90/p99, estimated
 effective IO bandwidth, and CUDA kernel launches per call.
 
 ```bash
-CUDA_VISIBLE_DEVICES=0 python scripts/tools/benchmark_gemv.py \
+CUDA_VISIBLE_DEVICES=0 python csrc/bench/benchmark_gemv.py \
   --output results/decode_linear.json \
   --markdown-output results/decode_linear.md
 ```
@@ -25,7 +25,7 @@ For direct A/B coverage of the custom kernel and guarded dispatcher across
 traditional LLaMA and GPT-NeoX decode shapes, use:
 
 ```bash
-CUDA_VISIBLE_DEVICES=0 PYTHONPATH=. python scripts/tools/benchmark_gemv_common.py \
+CUDA_VISIBLE_DEVICES=0 PYTHONPATH=. python csrc/bench/benchmark_gemv_common.py \
   --suite all --family traditional --m 2 4 \
   --output results/gemv_common.json
 ```
diff --git a/docs/developer/swiglu_benchmark.md b/docs/developer/swiglu_benchmark.md
index 575ad81..199fc8e 100644
--- a/docs/developer/swiglu_benchmark.md
+++ b/docs/developer/swiglu_benchmark.md
@@ -1,12 +1,12 @@
 # Fused SwiGLU benchmark
 
-`scripts/tools/benchmark_swiglu.py` compares the directly callable fused BF16
+`csrc/bench/benchmark_swiglu.py` compares the directly callable fused BF16
 SwiGLU primitive with both `F.linear` and the existing two-GEMV chain. It covers
 the native AstrAI 1B MLP plus LLaMA 2 7B/13B, LLaMA 3 8B, and GPT-NeoX 20B
 up/gate shapes at M=1/2/4/8 in eager and CUDA Graph modes.
 
 ```bash
-CUDA_VISIBLE_DEVICES=0 python scripts/tools/benchmark_swiglu.py \
+CUDA_VISIBLE_DEVICES=0 python csrc/bench/benchmark_swiglu.py \
   --output results/swiglu.json \
   --markdown-output results/swiglu.md \
   --m-values 1,2,4,8 --mode both \