diff --git a/AGENTS.md b/AGENTS.md
index 0f978b6c77..8a4ec563c7 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -33,6 +33,10 @@ prose in this repo, including this file.
   succeed on a candidate the shared rules would refuse.
 - Do not measure before confirming the code path can fire. An A/B whose B-arm cannot execute returns
   two green runs masquerading as evidence.
+- A green check is evidence only once you have seen it red. Prove a new regression test against the
+  pre-fix code (revert, run, quote the failing number), a moved test against its gates (planted type
+  error, discovered-count delta), a structural gate against a planted violation. Three vacuous
+  regression tests shipped in one day before this rule; review caught all three.
 - Typed signals over message sniffing: key on structured details (`details.timeoutMs`, reason codes),
   never on error text. Remaining sniffs are owned debt with in-code rationale — do not copy them.
 - Snapshot output is the token budget. Never add per-node bytes to the tree; response-level metadata
diff --git a/docs/agents/testing.md b/docs/agents/testing.md
index 9c4e24b5ec..a5406b7d7d 100644
--- a/docs/agents/testing.md
+++ b/docs/agents/testing.md
@@ -55,6 +55,17 @@ it through the barrel.
 
 Keep tests behavioral. Do not assert shapes or cases TypeScript already proves.
 
+A test added as a regression pin must be shown to fail without the change it pins — vacuity is the
+default failure mode, not the exception, because the adversarial input you imagine is rarely the one
+the old code was slow or wrong on (edge runs that the old regex handled in one pass; invariants the
+old implementation already satisfied; entry points whose trimming defuses the exploit before it
+reaches the flagged pattern). The proof is mechanical: revert the production change locally, watch
+the test fail, note the failing number, restore. Same rule at other layers: after relocating tests,
+prove the runner discovers them (file/test counts must move) and the typechecker reaches them (plant
+a type error, watch it surface, remove it); after adding an ownership/structural gate, plant a
+violation and watch it name the invariant. Quote the red run in the PR — a reviewer who cannot see
+the red has to re-derive it.
+
 Test through public interfaces where practical, and do not add unrelated production exports solely
 to make a test easier — widening the public surface for a test is a product change, and the exports
 outlive the test that motivated them. If a seam is genuinely missing, add it as a real one rather