From e6ea9d55243a6c3cc8fd7f5e6980c226ab5f7e26 Mon Sep 17 00:00:00 2001
From: bsha6 <54374986+bsha6@users.noreply.github.com>
Date: Thu, 3 Sep 2026 23:58:47 -0400
Subject: [PATCH 1/3] docs: Add operating guidance for agents

---
 AGENTS.md | 29 ++++++++++++++++++++++++++++-
 1 file changed, 28 insertions(+), 1 deletion(-)

diff --git a/AGENTS.md b/AGENTS.md
index 43c994c..8cf7b80 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -1 +1,28 @@
-@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. 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 the clean,
+  detached 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 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.

From 4a653149c2d17400ea3386428c1a8b65581a6b5e Mon Sep 17 00:00:00 2001
From: bsha6 <54374986+bsha6@users.noreply.github.com>
Date: Fri, 4 Sep 2026 00:10:02 -0400
Subject: [PATCH 2/3] docs: Clarify command contracts for agents

---
 AGENTS.md | 11 ++++++-----
 1 file changed, 6 insertions(+), 5 deletions(-)

diff --git a/AGENTS.md b/AGENTS.md
index 8cf7b80..e823400 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -7,18 +7,19 @@ benchmark that produced that evidence. Start with [README.md](README.md); see
 ## Operating the gate
 
 - **Validate provenance first.** Run `agent-release-gate doctor` and stop if it
-  exits nonzero. Do not evaluate evidence against an unproven integration.
+  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 the clean,
-  detached checkout at `../ClawProBench` on commit
+  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 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.
+- **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

From 4a9411ea3f5c63449018e121bc4cea964a35fcf5 Mon Sep 17 00:00:00 2001
From: bsha6 <54374986+bsha6@users.noreply.github.com>
Date: Fri, 4 Sep 2026 00:11:57 -0400
Subject: [PATCH 3/3] docs: Fix agent guidance grammar

---
 AGENTS.md | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/AGENTS.md b/AGENTS.md
index e823400..b9ba8e8 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -10,7 +10,7 @@ benchmark that produced that evidence. Start with [README.md](README.md); see
   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 the clean,
+  clean, or modify its checkout. The default integration expects a clean
   checkout at `../ClawProBench` on commit
   `c4b8395854fe0752eef435b44f140366efd44d8e`, without `ironclaw/` or
   `nanoclaw/` checked out.