diff --git a/crud.py b/crud.py
index f0f7f37..a622d96 100644
--- a/crud.py
+++ b/crud.py
@@ -1563,6 +1563,13 @@ async def apply_reported_state(
     Returns True if the state was applied, False if the event was not newer
     than what is already on file (see _should_apply_state_event).
 
+    The payload is the machine's full bay set, and the machine owns that
+    layout — the bay count is hardware-determined. So positions absent from
+    the payload are DELETED here. Leaving them behind was half the problem a
+    shrinking bay count caused: the stale row stayed, the dashboard showed a
+    mix of old and new, and every later publish was rejected for a position
+    mismatch with no way out but hand-written SQL.
+
     Populates both the operator-believed columns (denomination, count,
     updated_at, updated_by) and the reported columns (state_denomination,
     state_count, state_at, state_event_id), so the UI can show reported
@@ -1584,6 +1591,19 @@ async def apply_reported_state(
     if not _should_apply_state_event(oldest_state_at, event_created_at):
         return False
 
+    # Drop bays the machine no longer reports, before writing the rest. A crash
+    # between the two leaves rows missing rather than stale, and the next
+    # heartbeat re-inserts them — the safe direction.
+    keep = sorted(payload.positions.keys())
+    placeholders = ", ".join(f":p{i}" for i in range(len(keep)))
+    delete_values: dict = {"mid": machine_id}
+    for i, pos in enumerate(keep):
+        delete_values[f"p{i}"] = pos
+    await db.execute(
+        "DELETE FROM spirekeeper.cassette_configs "
+        f"WHERE machine_id = :mid AND position NOT IN ({placeholders})",
+        delete_values,
+    )
 
     now = datetime.now()
     for pos, row in payload.positions.items():