diff --git a/reg_field.h b/reg_field.h
index b1146a9..f1628aa 100755
--- a/reg_field.h
+++ b/reg_field.h
@@ -674,7 +674,12 @@ void matvec_absolute_rls(
     offset_t nall   = nbatch + ndim;
     offset_t osc    = stride_out[nall];
     offset_t isc    = stride_inp[nall];
-    offset_t wsc    = stride_wgt[nall];
+    // RLS: wgt has wc=1 (a single weight broadcast across all `nc` field
+    // channels), so the per-channel stride must be forced to 0 -- using the
+    // tensor's real (size-1) last-dim stride here indexes wgt[stride*c] for
+    // c>0, walking off the wc=1 buffer into the next voxel's weight (or OOB
+    // at the last voxel). See fastfields-cpu-lib#62.
+    offset_t wsc    = 0;
     offset_t nc     = size[nall];
     offset_t numel  = prod(size, nall);  // no outer loop across channels
 
@@ -721,7 +726,12 @@ void diag_absolute_rls(
 
     offset_t nall   = nbatch + ndim;
     offset_t osc    = stride_out[nall];
-    offset_t wsc    = stride_wgt[nall];
+    // RLS: wgt has wc=1 (a single weight broadcast across all `nc` field
+    // channels), so the per-channel stride must be forced to 0 -- using the
+    // tensor's real (size-1) last-dim stride here indexes wgt[stride*c] for
+    // c>0, walking off the wc=1 buffer into the next voxel's weight (or OOB
+    // at the last voxel). See fastfields-cpu-lib#62.
+    offset_t wsc    = 0;
     offset_t nc     = size[nall];
     offset_t numel  = prod(size, nall);    // no outer loop across channels
 
@@ -775,7 +785,12 @@ void relax_absolute_rls_(
     offset_t osc    = stride_sol[nall];
     offset_t hsc    = stride_hes[nall];
     offset_t gsc    = stride_grd[nall];
-    offset_t wsc    = stride_wgt[nall];
+    // RLS: wgt has wc=1 (a single weight broadcast across all `nc` field
+    // channels), so the per-channel stride must be forced to 0 -- using the
+    // tensor's real (size-1) last-dim stride here indexes wgt[stride*c] for
+    // c>0, walking off the wc=1 buffer into the next voxel's weight (or OOB
+    // at the last voxel). See fastfields-cpu-lib#62.
+    offset_t wsc    = 0;
     offset_t nc    = size[nall];
     offset_t numel  = prod(size, nall);    // no outer loop across channels
 
@@ -1043,7 +1058,12 @@ void matvec_membrane_rls(
     offset_t nall   = nbatch + ndim;
     offset_t osc    = stride_out[nall];
     offset_t isc    = stride_inp[nall];
-    offset_t wsc    = stride_wgt[nall];
+    // RLS: wgt has wc=1 (a single weight broadcast across all `nc` field
+    // channels), so the per-channel stride must be forced to 0 -- using the
+    // tensor's real (size-1) last-dim stride here indexes wgt[stride*c] for
+    // c>0, walking off the wc=1 buffer into the next voxel's weight (or OOB
+    // at the last voxel). See fastfields-cpu-lib#62.
+    offset_t wsc    = 0;
     offset_t nc     = size[nall];
     offset_t numel  = prod(size, nall);  // no outer loop across channels
 
@@ -1096,7 +1116,12 @@ void diag_membrane_rls(
     reduce_t voxel_size[ndim];    fillfrom<ndim>(voxel_size, _voxel_size);
     offset_t nall   = nbatch + ndim;
     offset_t osc    = stride_out[nall];
-    offset_t wsc    = stride_wgt[nall];
+    // RLS: wgt has wc=1 (a single weight broadcast across all `nc` field
+    // channels), so the per-channel stride must be forced to 0 -- using the
+    // tensor's real (size-1) last-dim stride here indexes wgt[stride*c] for
+    // c>0, walking off the wc=1 buffer into the next voxel's weight (or OOB
+    // at the last voxel). See fastfields-cpu-lib#62.
+    offset_t wsc    = 0;
     offset_t nc     = size[nall];
     offset_t numel  = prod(size, nall);    // no outer loop across channels
 
@@ -1156,7 +1181,12 @@ void relax_membrane_rls_(
     offset_t osc    = stride_sol[nall];
     offset_t hsc    = stride_hes[nall];
     offset_t gsc    = stride_grd[nall];
-    offset_t wsc    = stride_wgt[nall];
+    // RLS: wgt has wc=1 (a single weight broadcast across all `nc` field
+    // channels), so the per-channel stride must be forced to 0 -- using the
+    // tensor's real (size-1) last-dim stride here indexes wgt[stride*c] for
+    // c>0, walking off the wc=1 buffer into the next voxel's weight (or OOB
+    // at the last voxel). See fastfields-cpu-lib#62.
+    offset_t wsc    = 0;
     offset_t nc     = size[nall];
     offset_t numel  = prod(size, nall);    // no outer loop across channels
 
@@ -1452,7 +1482,12 @@ void matvec_bending_rls(
     offset_t nall   = nbatch + ndim;
     offset_t osc    = stride_out[nall];
     offset_t isc    = stride_inp[nall];
-    offset_t wsc    = stride_wgt[nall];
+    // RLS: wgt has wc=1 (a single weight broadcast across all `nc` field
+    // channels), so the per-channel stride must be forced to 0 -- using the
+    // tensor's real (size-1) last-dim stride here indexes wgt[stride*c] for
+    // c>0, walking off the wc=1 buffer into the next voxel's weight (or OOB
+    // at the last voxel). See fastfields-cpu-lib#62.
+    offset_t wsc    = 0;
     offset_t nc     = size[nall];
     offset_t numel  = prod(size, nall);  // no outer loop across channels
 
@@ -1506,7 +1541,12 @@ void diag_bending_rls(
     reduce_t voxel_size[ndim];    fillfrom<ndim>(voxel_size, _voxel_size);
     offset_t nall   = nbatch + ndim;
     offset_t osc    = stride_out[nall];
-    offset_t wsc    = stride_wgt[nall];
+    // RLS: wgt has wc=1 (a single weight broadcast across all `nc` field
+    // channels), so the per-channel stride must be forced to 0 -- using the
+    // tensor's real (size-1) last-dim stride here indexes wgt[stride*c] for
+    // c>0, walking off the wc=1 buffer into the next voxel's weight (or OOB
+    // at the last voxel). See fastfields-cpu-lib#62.
+    offset_t wsc    = 0;
     offset_t nc     = size[nall];
     offset_t numel  = prod(size, nall);    // no outer loop across channels
 
@@ -1567,7 +1607,12 @@ void relax_bending_rls_(
     offset_t osc    = stride_sol[nall];
     offset_t hsc    = stride_hes[nall];
     offset_t gsc    = stride_grd[nall];
-    offset_t wsc    = stride_wgt[nall];
+    // RLS: wgt has wc=1 (a single weight broadcast across all `nc` field
+    // channels), so the per-channel stride must be forced to 0 -- using the
+    // tensor's real (size-1) last-dim stride here indexes wgt[stride*c] for
+    // c>0, walking off the wc=1 buffer into the next voxel's weight (or OOB
+    // at the last voxel). See fastfields-cpu-lib#62.
+    offset_t wsc    = 0;
     offset_t nc     = size[nall];
     offset_t numel  = prod(size, nall);    // no outer loop across channels