diff --git a/image_analysis/scale_merge/RotationScaleMerge.cpp b/image_analysis/scale_merge/RotationScaleMerge.cpp
index bfde3cd48..d3597ca03 100644
--- a/image_analysis/scale_merge/RotationScaleMerge.cpp
+++ b/image_analysis/scale_merge/RotationScaleMerge.cpp
@@ -4841,9 +4841,8 @@ RotationScaleMerge::Result RotationScaleMerge::Run(bool for_search, bool full_st
     }
     // The total scale each frame's fulls were amplified by - partial scale, incident flux and the fulls'
     // own G (corr = 1/G here) - taken BEFORE the correction surfaces fold per-observation terms into
-    // corr, and before a collapsed frame's fulls are zeroed: the resident fulls only get that zero with
-    // the corrected corr, so a merge on the device before then still sees them, and must weigh them by
-    // their own scale. Read by MergeAndStats for the CC1/2 weight.
+    // corr, and before a collapsed frame's fulls are zeroed (a zeroed corr says nothing about the
+    // frame's scale). Read by MergeAndStats for the CC1/2 weight.
     frame_scale_fulls.assign(n_frames, NAN);
     for (const auto &o : fulls)
         if (std::isfinite(o.corr) && o.corr > 0.0f && !std::isfinite(frame_scale_fulls[o.frame])) {
@@ -4852,6 +4851,22 @@ RotationScaleMerge::Result RotationScaleMerge::Run(bool for_search, bool full_st
         }
     const bool rejected_full_scales =
             scale_fulls && DropCollapsedFullScales(combined_on_gpu && scaled_fulls_on_gpu);
+#ifdef JFJOCH_USE_CUDA
+    // The host fulls' corr is the truth from here on; a merge on the resident (GPU) fulls reads the
+    // device copy, so every change to it is pushed back before the next device merge.
+    auto upload_fulls_corr = [&]() {
+        std::vector<float> fcorr(fulls.size());
+        ParallelChunks(static_cast<int>(fulls.size()), ThreadsForWork(fulls.size(), nthreads),
+                       [&](int lo, int hi) {
+            for (int i = lo; i < hi; ++i) fcorr[i] = fulls[i].corr;
+        });
+        gpu_->SetFullsCorr(fcorr.data());
+    };
+    // Right away, not with the corrections below: the merge just before the surfaces must see the
+    // collapsed frames' fulls gone on the device exactly as the host merge does.
+    if (rejected_full_scales && combined_on_gpu && scaled_fulls_on_gpu)
+        upload_fulls_corr();
+#endif
 
     // --- 4b. Optional correction surfaces (decay = resolution x time; absorption = goniometer-frame
     //         diffracted-beam direction), each an alternating multiplicative fit of the fulls' corr against
@@ -4917,16 +4932,10 @@ RotationScaleMerge::Result RotationScaleMerge::Run(bool for_search, bool full_st
         CountSweepDisposition();
     }
 #ifdef JFJOCH_USE_CUDA
-    // The corrections mutate the host fulls' corr; when the merge runs on the resident (GPU) fulls, push
-    // the corrected corr back to the device so the merge reads it.
-    if ((corrections || rejected_full_scales || rejected_frames) && combined_on_gpu && scaled_fulls_on_gpu) {
-        std::vector<float> fcorr(fulls.size());
-        ParallelChunks(static_cast<int>(fulls.size()), ThreadsForWork(fulls.size(), nthreads),
-                       [&](int lo, int hi) {
-            for (int i = lo; i < hi; ++i) fcorr[i] = fulls[i].corr;
-        });
-        gpu_->SetFullsCorr(fcorr.data());
-    }
+    // The corrections and the frame rejection mutate the host fulls' corr; push it to the device so the
+    // merge reads it. (The collapsed-scale drop was pushed where it happened.)
+    if ((corrections || rejected_frames) && combined_on_gpu && scaled_fulls_on_gpu)
+        upload_fulls_corr();
 #endif
 
     // --- 5. Error model + merge + statistics. ---