diff --git a/src/programbench/eval/eval.py b/src/programbench/eval/eval.py
index ba1f692c..2f17b8cd 100644
--- a/src/programbench/eval/eval.py
+++ b/src/programbench/eval/eval.py
@@ -492,6 +492,19 @@ def _compile_executable(self, env: ContainerEnvironment, log_buf: list[dict]) ->
         assert self.submission_archive is not None
         env.copy_in_tar(self.submission_archive, f"{WORKSPACE_DIR}/")
         self._remove_hashed_files(env, log_buf)
+        # Remove any ./executable shipped in the submission archive so the
+        # build must produce a fresh binary. The workspace wipe above runs
+        # *before* the tar is extracted, so a prebuilt executable smuggled
+        # into the submission would otherwise survive into the compile step
+        # and be stashed as if the build had produced it (a stub/no-op
+        # compile.sh would then "pass"). _remove_hashed_files only catches
+        # byte-for-byte copies of the gold binary, not arbitrary prebuilts.
+        self._run_step(
+            "rm -f ./executable",
+            env=env,
+            log_buf=log_buf,
+            step_name="clear_stale_executable",
+        )
         # Seed a synthetic git repo if the submission didn't ship one. Build
         # scripts that depend on a working tree (jq submodules, calcurse's
         # autopoint, cargo+vergen, ...) succeed against this synthetic repo.