diff --git a/docs/BENCHMARKING.md b/docs/BENCHMARKING.md
new file mode 100644
index 0000000..12e9fde
--- /dev/null
+++ b/docs/BENCHMARKING.md
@@ -0,0 +1,44 @@
+# Benchmarking
+
+How the performance numbers attached to this project are measured, so that a
+number in a changelog entry or a release note can be reproduced and trusted.
+
+## The suite
+
+The benchmarks live in `bench_test.go`, next to the code they measure:
+
+| Benchmark | What it measures |
+|---|---|
+| `BenchmarkParse` | `Parse` over a representative configuration document |
+| `BenchmarkMarshal` | `Marshal` of the tree `Parse` produced from the same document |
+| `BenchmarkStrictDecode` | `Decode` into a struct under `DisallowUnknownFields` |
+| `BenchmarkParseLong` | `Parse` over a generated document with about 2000 array-of-tables entries |
+
+## Running
+
+```sh
+just bench
+```
+
+The recipe runs the suite with `-benchmem -count=5`. Every benchmark uses
+`b.Loop`, so setup runs outside the timed region, and `ReportAllocs` records
+allocations per operation. The parse and marshal benchmarks set `SetBytes`, so
+their results read as input bytes per second.
+
+## Method
+
+- An idle machine only: a loaded box times whatever else is running, and the
+  fastest sample can land on the wrong function.
+- An A/B comparison runs both variants inside one process, in one binary;
+  separate processes of identical binaries differ by more than the effect
+  being measured.
+- The five counts are compared through their medians, allocations and bytes
+  per operation alongside the times. Differences within 1 to 2 percent are
+  noise; only a difference beyond that is a result.
+- When timing is hopeless, the allocation and byte counts are the result.
+
+## Reports
+
+The repository stores no benchmark reports. A performance claim in
+`CHANGELOG.md` is measured with the method above on the change that makes it,
+and the number travels with the claim.
diff --git a/docs/DEVELOPMENT.md b/docs/DEVELOPMENT.md
index 097030e..dbdd1b1 100644
--- a/docs/DEVELOPMENT.md
+++ b/docs/DEVELOPMENT.md
@@ -77,7 +77,9 @@ just bench
 ```
 
 Benchmark on an idle machine, and compare only runs made in one process against
-each other. The recipe sweeps `./...` five times with `-benchmem`.
+each other. The recipe sweeps `./...` five times with `-benchmem`. The binding
+measurement method, and what counts as a result, is in
+[docs/BENCHMARKING.md](BENCHMARKING.md).
 
 ## Debugging the build