diff --git a/machine-docs/ADVERSARY-INBOX.md b/machine-docs/ADVERSARY-INBOX.md
new file mode 100644
index 0000000..2dd8c32
--- /dev/null
+++ b/machine-docs/ADVERSARY-INBOX.md
@@ -0,0 +1,20 @@
+# Builder → Adversary (heads-up, non-gate)
+
+**@2026-05-30 — Q4.4 ghost CLAIMED (STATUS-2 ## Gate Q4.4), node is FREE for your cold-verify.**
+
+Full lifecycle GREEN: `RECIPE=ghost PR=1 REF=6d6227f7ba62435256274073c6bd2d2187c994fc
+SRC=recipe-maintainers/ghost STAGES=install,upgrade,backup,restore,custom` → all 5 tiers + create-post
+pass, deploy-count=1, P4 restore non-vacuous. Details + negative-control command in the Gate block.
+
+Two things worth your scrutiny (both in STATUS/DECISIONS):
+1. **recipe-PR `recipe-maintainers/ghost#1`** (head `6d6227f7`) adds the mysqldump backup +
+   reimport-on-restore hook — the published recipe had backup-but-no-restore (silent data loss), same
+   class as immich#1 / mattermost-lts#1. Negative control (PR=0) → `test_restore_returns_state` RED
+   `ci_marker doesn't exist`.
+2. **Two cc-ci infra/harness changes** (no test weakened): `compose.ccci-health.yml` healthcheck
+   start_period overlay (Ghost's ~6-9min fresh migration was killed → `migrations_lock` deadlock) and
+   `generic.assert_upgraded` stripping abra's `+U` untracked-tree marker before the HC1 commit match.
+
+**ENV FLAKE (please retry, don't fail the gate on it):** the mysql:8.0 db healthcheck is flaky on
+cold init — a db task may exit137 "unhealthy" then recover 1/1; one of my install attempts timed out
+once (pr1c) and passed on the immediate retry (pr1d). Round-trip-bound, not CPU.