diff --git a/AGENTS.md b/AGENTS.md
index 43c994c..b9ba8e8 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -1 +1,29 @@
-@AGENTS.md
+# Agent Release Gate guidance
+
+This repository evaluates completed benchmark evidence. It does not run the
+benchmark that produced that evidence. Start with [README.md](README.md); see
+[docs/architecture.md](docs/architecture.md) for the trust boundary.
+
+## Operating the gate
+
+- **Validate provenance first.** Run `agent-release-gate doctor` and stop if it
+  exits nonzero. Exit `0` confirms provenance only; it is not release approval.
+  Do not evaluate evidence against an unproven integration.
+- **Keep ClawProBench read-only.** Never install, import, execute, fetch, reset,
+  clean, or modify its checkout. The default integration expects a clean
+  checkout at `../ClawProBench` on commit
+  `c4b8395854fe0752eef435b44f140366efd44d8e`, without `ironclaw/` or
+  `nanoclaw/` checked out.
+- **Treat reports as immutable evidence.** Do not fabricate, repair, or rewrite
+  a benchmark report to make a gate pass. Invalid evidence is an evaluation
+  error, not a release result.
+- **Interpret `evaluate` exit codes exactly.** Exit `0` is `go`, exit `1` is a
+  valid `no_go`, and exit `2` means the evaluation itself is invalid. Do not
+  collapse exit `1` and exit `2` into the same outcome.
+- **Read the decision artifact.** Report every blocker from the generated JSON;
+  do not infer success merely because an output file already exists.
+- **Keep output separate.** Never place a decision inside the benchmark
+  checkout or overwrite the report, policy, or integration manifest.
+- **Resolve configuration deliberately.** Run from the source checkout when
+  using the included default policy and integration manifest. Otherwise pass
+  explicit `--policy` and `--integration` paths.