diff --git a/tools/sizes.py b/tools/sizes.py
index 184e460..520b136 100755
--- a/tools/sizes.py
+++ b/tools/sizes.py
@@ -84,6 +84,20 @@ def collect() -> dict[str, list[tuple[str, int, int]]]:
         for match in SIZE_TEST.finditer((tree / "CTestTestfile.cmake").read_text()):
             found.setdefault(chip, []).append((match["name"], match["elf"], int(match["limit"])))
     sizes = measure([elf for rows in found.values() for _, elf, _ in rows], tool)
+    # A chip's generated and reflect trees must answer with the same bytes
+    # (the identity invariant), so the same target measuring two sizes means
+    # a stale tree — or an identity breach. Either is a finding; picking one
+    # silently is how a gate reports another build's numbers as today's.
+    for chip, rows in found.items():
+        seen: dict[str, tuple[int, str]] = {}
+        for name, elf, _ in rows:
+            if elf not in sizes:
+                continue
+            if name in seen and seen[name][0] != sizes[elf]:
+                sys.exit(f"{chip} {name}: {seen[name][0]} B in {seen[name][1]} but "
+                         f"{sizes[elf]} B in {elf} — a stale tree (rebuild or remove it) "
+                         f"or a cross-mode identity breach")
+            seen.setdefault(name, (sizes[elf], elf))
     measured = {
         chip: sorted(((name, sizes[elf], limit) for name, elf, limit in rows if elf in sizes),
                      key=lambda row: -row[1])