diff --git a/bench_test.go b/bench_test.go
new file mode 100644
index 0000000..9c5c520
--- /dev/null
+++ b/bench_test.go
@@ -0,0 +1,124 @@
+package acp
+
+import (
+	"context"
+	"io"
+	"os"
+	"path/filepath"
+	"testing"
+)
+
+// benchSource supplies one targetless item, which is the shape a hash-only run uses.
+type benchSource struct {
+	path string
+	done bool
+}
+
+func (s *benchSource) Next(context.Context) ([]Item, error) {
+	if s.done {
+		return nil, io.EOF
+	}
+	s.done = true
+	return []Item{&benchItem{source: s}}, nil
+}
+
+type benchItem struct {
+	source   *benchSource
+	failures int
+}
+
+func (i *benchItem) Source() string    { return i.source.path }
+func (i *benchItem) Targets() []string { return nil }
+func (i *benchItem) Completed(*Result) {}
+func (i *benchItem) Failed(error)      { i.failures++ }
+
+func benchFile(b *testing.B, size int64) string {
+	b.Helper()
+
+	path := filepath.Join(b.TempDir(), "content.bin")
+	file, err := os.Create(path)
+	if err != nil {
+		b.Fatal(err)
+	}
+	defer file.Close()
+
+	block := make([]byte, 1<<20)
+	for index := range block {
+		block[index] = byte(index)
+	}
+	for written := int64(0); written < size; written += int64(len(block)) {
+		if _, err := file.Write(block); err != nil {
+			b.Fatal(err)
+		}
+	}
+	if err := file.Close(); err != nil {
+		b.Fatal(err)
+	}
+	return path
+}
+
+// BenchmarkReadBuffered and BenchmarkReadMapped measure the same hash-only read through
+// each source read mode, which is the comparison behind reading every source buffered.
+func BenchmarkReadBuffered(b *testing.B) {
+	path := benchFile(b, 64<<20)
+	b.SetBytes(64 << 20)
+	b.ResetTimer()
+
+	for i := 0; i < b.N; i++ {
+		source := &benchSource{path: path}
+		if err := Run(context.Background(), source, WithHashPolicy(HashRead), SetFromDevice(WithReadMode(ReadBuffered))); err != nil {
+			b.Fatal(err)
+		}
+	}
+}
+
+func BenchmarkReadMapped(b *testing.B) {
+	path := benchFile(b, 64<<20)
+	b.SetBytes(64 << 20)
+	b.ResetTimer()
+
+	for i := 0; i < b.N; i++ {
+		source := &benchSource{path: path}
+		if err := Run(context.Background(), source, WithHashPolicy(HashRead), SetFromDevice(WithReadMode(ReadMapped))); err != nil {
+			b.Fatal(err)
+		}
+	}
+}
+
+// BenchmarkRefreshSignatureUnchanged and BenchmarkRefreshSignatureChanged measure the
+// cache publication path: an unchanged stored hash is compared and left alone, a changed
+// one is written back. The difference is the physical write a refresh would otherwise cost
+// on every scanned file.
+func BenchmarkRefreshSignatureUnchanged(b *testing.B) {
+	path := benchFile(b, 4096)
+	// Seed the cache once so the measured runs compare against a stored value.
+	if err := Run(context.Background(), &benchSource{path: path}, WithHashPolicy(HashReadRefresh)); err != nil {
+		b.Fatal(err)
+	}
+	b.ResetTimer()
+
+	for i := 0; i < b.N; i++ {
+		if err := Run(context.Background(), &benchSource{path: path}, WithHashPolicy(HashReadRefresh)); err != nil {
+			b.Fatal(err)
+		}
+	}
+}
+
+func BenchmarkRefreshSignatureChanged(b *testing.B) {
+	path := benchFile(b, 4096)
+	b.ResetTimer()
+
+	for i := 0; i < b.N; i++ {
+		// A different mtime makes the stored entry stale, so the run rewrites it.
+		info, err := os.Stat(path)
+		if err != nil {
+			b.Fatal(err)
+		}
+		if err := os.Chtimes(path, info.ModTime(), info.ModTime().Add(1)); err != nil {
+			b.Fatal(err)
+		}
+		if err := Run(context.Background(), &benchSource{path: path}, WithHashPolicy(HashReadRefresh)); err != nil {
+			b.Fatal(err)
+		}
+	}
+}