diff --git a/AGENTS.md b/AGENTS.md
index 2390ebb..10b37df 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -97,6 +97,47 @@ same each time: group by the thing that varies, then look.
   version of the truth. It was a different and wrong claim. Retract it rather
   than soften it.
 
+### 3b. Never report a test count that a filter could have hidden a failure in.
+
+Do not commit or push with a failing test. If one fails, it is fixed or reverted
+in the same commit — never "fixed in the next one".
+
+The same rule governs the *number* you report for the suite. Do not derive it
+with a filter that drops the failing binaries:
+
+```sh
+# WRONG. Sums only the binaries that passed. A red build still prints a number,
+# and the number looks fine.
+cargo test -q --workspace | grep "^test result: ok" | awk '{s += $4} END {print s}'
+```
+
+A commit was pushed with a broken test because of exactly this. The filtered sum
+reported a healthy count, because `^test result: ok` matches the six green test
+binaries and ignores the seventh, which was red. The giveaway was that the count
+came back lower than the previous run's, and the difference was mistaken for
+"the disc tests were skipped" rather than "one of them failed". It was read as
+an environment difference and moved on.
+
+Use the exit status, or count the lines that are not green:
+
+```sh
+cargo test --workspace            # then read every "test result:" line
+cargo test --workspace && echo ok # the exit status is the real answer
+```
+
+Concretely:
+
+- Run the full suite with the disc present when the change touches disc-decoding
+  code, and say so in the commit body. A skipped test and a passing one both
+  report success.
+- If the reported count moves between runs, find out why before writing anything
+  down. A drop means either an environment difference or a failure, and the two
+  look identical from the summary line.
+- Never quote a number you computed rather than read.
+
+This is the same defect as rules 3 and 3a, one level up: a quantity that should
+have had one definition had two, and nothing compared them.
+
 ### 4. Controls, and heuristic counts that are not identifications.
 
 Concretely: