diff --git a/mlperf_logging/rcp_checker/rcp_checker.py b/mlperf_logging/rcp_checker/rcp_checker.py
index a593f3a..99b0fa9 100644
--- a/mlperf_logging/rcp_checker/rcp_checker.py
+++ b/mlperf_logging/rcp_checker/rcp_checker.py
@@ -468,10 +468,27 @@ def _set_results_scaling(self, scale_factor, results_dir):
         with open(filepath, "w") as scaling_file:
             scaling_file.write(json_content)
 
-    def _reset_results_scaling(self, results_dir):
+    def _update_results_scaling(self, scale_factor, results_dir):
+        '''Make results_dir/scaling.json hold scale_factor (no file means 1.0).
+        An existing file that already holds scale_factor is left untouched.'''
         filepath = results_dir+'/scaling.json'
+        old_factor = 1.0
         if os.path.exists(filepath):
+            try:
+                with open(filepath) as scaling_file:
+                    old_factor = float(json.load(scaling_file)['scaling_factor'])
+            except (OSError, ValueError, KeyError, TypeError):
+                old_factor = None
+        if old_factor == scale_factor:
+            return
+        if os.path.exists(filepath):
+            logging.warning(" Replacing scaling factor %s in %s with newly computed %.4f",
+                            old_factor, filepath, scale_factor)
+        if scale_factor == 1.0:
             os.remove(filepath)
+        else:
+            self._set_results_scaling(scale_factor, results_dir)
+            logging.info(" Results scaling set to normalization factor of %.4f", scale_factor)
 
     def _eval_submission_record(self, rcp_record, subm_epochs, results_dir):
         '''Compare reference and submission convergence.'''
@@ -483,7 +500,6 @@ def _eval_submission_record(self, rcp_record, subm_epochs, results_dir):
         samples_rejected = 4 if rcp_record["Benchmark"] == 'unet3d' else 1
         mean_subm_epochs = np.mean(subm_epochs[samples_rejected:len(subm_epochs)-samples_rejected])
         norm_factor = self._find_norm_factor(rcp_record, mean_subm_epochs)
-        self._reset_results_scaling(results_dir)
         if mean_subm_epochs >= (rcp_record["RCP Mean"] / rcp_record["Max Speedup"]):
             logging.info(" RCP Record: %s", rcp_record)
             logging.info(" Submission mean epochs: %.4f", mean_subm_epochs)
@@ -491,9 +507,9 @@ def _eval_submission_record(self, rcp_record, subm_epochs, results_dir):
                 mesg = " Submission mean epochs faster than RCP mean but within max speedup range. Score should be normalized by factor of {} / {} = {}"
                 mesg = mesg.format(rcp_record["RCP Mean"], mean_subm_epochs, norm_factor)
                 logging.info(mesg)
-                if results_dir != '':
-                    self._set_results_scaling(norm_factor, results_dir)
-                    logging.info(" Results scaling set to normalization factor of %.4f", norm_factor)
+            # norm_factor is 1.0 when the submission is no faster than the RCP mean.
+            if results_dir != '':
+                self._update_results_scaling(norm_factor, results_dir)
             return True, norm_factor
         else:
             logging.info(" RCP Record: %s", rcp_record)
diff --git a/mlperf_logging/result_summarizer/result_summarizer.py b/mlperf_logging/result_summarizer/result_summarizer.py
index 1e1b828..17d21c0 100644
--- a/mlperf_logging/result_summarizer/result_summarizer.py
+++ b/mlperf_logging/result_summarizer/result_summarizer.py
@@ -510,12 +510,9 @@ def _compute_strong_scaling_scores(desc, system_folder, usage, ruleset, division
         has_power = _has_power(benchmark_folder)
         benchmark = _benchmark_alias(folder_parts[-1])
         system = folder_parts[-3] if usage == 'hpc' else folder_parts[-2]
-        # Compute base perf/power scores
-        score, power_score, rcp_scaling_factor = _compute_strong_score_standalone(
-            benchmark, system, has_power, benchmark_folder, usage, ruleset, desc
-        )
 
-        # RCP/GBS/Epochs additions for closed division
+        # RCP/GBS/Epochs additions for closed division. The RCP check rewrites
+        # scaling.json, so it must run before the scores below read it.
         benchmark_gbs = None
         benchmark_epochs = None
         benchmark_rcp = None
@@ -576,6 +573,11 @@ def _compute_strong_scaling_scores(desc, system_folder, usage, ruleset, division
                 )
                 traceback.print_exc()
 
+        # Compute base perf/power scores
+        score, power_score, rcp_scaling_factor = _compute_strong_score_standalone(
+            benchmark, system, has_power, benchmark_folder, usage, ruleset, desc
+        )
+
         # Map into metric-suffixed keys for schema
         detailed_bechmark_scores[f"{benchmark}:rcp_scaling_factor"] = float(
             rcp_scaling_factor