diff --git a/scripts/create_resources/combine/process_datasets_allen_nebius.sh b/scripts/create_resources/combine/process_datasets_allen_nebius.sh
new file mode 100644
index 000000000..ec29fdb2b
--- /dev/null
+++ b/scripts/create_resources/combine/process_datasets_allen_nebius.sh
@@ -0,0 +1,108 @@
+#!/bin/bash
+
+# Process ONLY the Allen Brain Cell Atlas (ABCA) MERFISH datasets (combine step) — FOUR
+# whole-brain sections, ALL paired with the SAME mouse-brain SC reference (like the LTX/MPII
+# combine scripts, and unlike Kuppe which uses condition-specific references):
+#   mouse1_coronal, mouse2_coronal, mouse3_sagittal, mouse4_sagittal
+#     <-> allen_brain_cell_atlas/2023_yao_mouse_brain_scrnaseq_10xv2 (ABCA 2023 Yao 10xv2 atlas)
+#
+# Reads each spatial input (process_allen_brain_cell_atlas_merfish loader output) and the SC
+# reference from the local /scratch raw folder, and writes the combined datasets to the
+# /scratch datasets folder (same layout the other *_nebius.sh combine scripts use).
+#
+# Prerequisites (both publish to the same /scratch raw folder used below):
+#   1. scripts/create_resources/spatial/process_allen_brain_cell_atlas_merfish_nebius.sh
+#        -> /scratch/.../raw/allen_brain_cell_atlas_merfish/mouse{1,2,3,4}_*/rep1/dataset.zarr
+#   2. scripts/create_resources/sc/process_allen_brain_cell_atlas_brain_nebius.sh  (log_cp -> hvg
+#      -> pca -> knn, i.e. it must carry a 'normalized' layer)
+#        -> /scratch/.../raw/allen_brain_cell_atlas/2023_yao_mouse_brain_scrnaseq_10xv2/dataset.h5ad
+
+# get the root of the directory
+REPO_ROOT=$(git rev-parse --show-toplevel)
+
+# ensure that the command below is run from the root of the repository
+cd "$REPO_ROOT"
+
+set -e
+
+# process_allen_brain_cell_atlas_merfish_nebius.sh + process_allen_brain_cell_atlas_brain_nebius.sh
+# both publish here.
+raw_dir='/scratch/task_ist_preprocessing/raw'
+publish_dir='/scratch/task_ist_preprocessing/datasets'
+
+# shared SC reference for all four sections (ABCA 2023 Yao whole-brain 10xv2 atlas)
+sc_ref="$raw_dir/allen_brain_cell_atlas/2023_yao_mouse_brain_scrnaseq_10xv2/dataset.h5ad"
+
+launch_batch() {
+  local params_file="$1"
+  local label="$2"
+  tw launch https://github.com/openproblems-bio/task_ist_preprocessing.git \
+    --revision build/main \
+    --pull-latest \
+    --main-script target/nextflow/workflows/process_datasets/main.nf \
+    --workspace 167877437119966 \
+    --compute-env 5hfmdCBxMRd4nHZaJKYEQZ \
+    --params-file "$params_file" \
+    --config src/base/labels_nebius.config \
+    --labels "task_ist_preprocessing,process_datasets,$label"
+}
+
+cat > /tmp/params_allen.yaml << HERE
+param_list:
+
+  - id: "allen_brain_cell_atlas_merfish_combined/mouse1_coronal/rep1"
+    input_sp: "$raw_dir/allen_brain_cell_atlas_merfish/mouse1_coronal/rep1/dataset.zarr"
+    input_sc: "$sc_ref"
+    dataset_id: "allen_brain_cell_atlas_merfish_combined/mouse1_coronal/rep1"
+    dataset_name: "Mouse brain combined ABCA MERFISH mouse1 coronal + 2023 Yao scRNAseq"
+    dataset_url: "https://download.brainimagelibrary.org/29/3c/293cc39ceea87f6d/"
+    dataset_reference: "10.1038/s41586-023-06812-z"
+    dataset_summary: "Allen Brain Cell Atlas whole-brain MERFISH mouse 1 (coronal, ~1100-gene panel) + 2023 Yao mouse-brain scRNAseq"
+    dataset_description: "Brain-wide MERFISH spatial transcriptomics (Zhuang lab, Allen Brain Cell Atlas), mouse 1 coronal section, paired with the ABCA 2023 Yao whole-mouse-brain 10xv2 scRNAseq reference."
+    dataset_organism: "mus_musculus"
+
+  - id: "allen_brain_cell_atlas_merfish_combined/mouse2_coronal/rep1"
+    input_sp: "$raw_dir/allen_brain_cell_atlas_merfish/mouse2_coronal/rep1/dataset.zarr"
+    input_sc: "$sc_ref"
+    dataset_id: "allen_brain_cell_atlas_merfish_combined/mouse2_coronal/rep1"
+    dataset_name: "Mouse brain combined ABCA MERFISH mouse2 coronal + 2023 Yao scRNAseq"
+    dataset_url: "https://download.brainimagelibrary.org/29/3c/293cc39ceea87f6d/"
+    dataset_reference: "10.1038/s41586-023-06812-z"
+    dataset_summary: "Allen Brain Cell Atlas whole-brain MERFISH mouse 2 (coronal, ~1100-gene panel) + 2023 Yao mouse-brain scRNAseq"
+    dataset_description: "Brain-wide MERFISH spatial transcriptomics (Zhuang lab, Allen Brain Cell Atlas), mouse 2 coronal section, paired with the ABCA 2023 Yao whole-mouse-brain 10xv2 scRNAseq reference."
+    dataset_organism: "mus_musculus"
+
+  - id: "allen_brain_cell_atlas_merfish_combined/mouse3_sagittal/rep1"
+    input_sp: "$raw_dir/allen_brain_cell_atlas_merfish/mouse3_sagittal/rep1/dataset.zarr"
+    input_sc: "$sc_ref"
+    dataset_id: "allen_brain_cell_atlas_merfish_combined/mouse3_sagittal/rep1"
+    dataset_name: "Mouse brain combined ABCA MERFISH mouse3 sagittal + 2023 Yao scRNAseq"
+    dataset_url: "https://download.brainimagelibrary.org/29/3c/293cc39ceea87f6d/"
+    dataset_reference: "10.1038/s41586-023-06812-z"
+    dataset_summary: "Allen Brain Cell Atlas whole-brain MERFISH mouse 3 (sagittal, ~1100-gene panel) + 2023 Yao mouse-brain scRNAseq"
+    dataset_description: "Brain-wide MERFISH spatial transcriptomics (Zhuang lab, Allen Brain Cell Atlas), mouse 3 sagittal section, paired with the ABCA 2023 Yao whole-mouse-brain 10xv2 scRNAseq reference."
+    dataset_organism: "mus_musculus"
+
+  - id: "allen_brain_cell_atlas_merfish_combined/mouse4_sagittal/rep1"
+    input_sp: "$raw_dir/allen_brain_cell_atlas_merfish/mouse4_sagittal/rep1/dataset.zarr"
+    input_sc: "$sc_ref"
+    dataset_id: "allen_brain_cell_atlas_merfish_combined/mouse4_sagittal/rep1"
+    dataset_name: "Mouse brain combined ABCA MERFISH mouse4 sagittal + 2023 Yao scRNAseq"
+    dataset_url: "https://download.brainimagelibrary.org/29/3c/293cc39ceea87f6d/"
+    dataset_reference: "10.1038/s41586-023-06812-z"
+    dataset_summary: "Allen Brain Cell Atlas whole-brain MERFISH mouse 4 (sagittal, ~1100-gene panel) + 2023 Yao mouse-brain scRNAseq"
+    dataset_description: "Brain-wide MERFISH spatial transcriptomics (Zhuang lab, Allen Brain Cell Atlas), mouse 4 sagittal section, paired with the ABCA 2023 Yao whole-mouse-brain 10xv2 scRNAseq reference."
+    dataset_organism: "mus_musculus"
+
+# ABCA whole-brain images are enormous (~83k x 102k px) and the tissue sits off-centre,
+# so an image-centred crop can miss it (mouse1: 0 of 42M transcripts). Centre the crop
+# on the transcript density instead (opt-in flag; default is image-centred).
+tissue_centered_crop: true
+
+output_sc: "\$id/output_sc.h5ad"
+output_sp: "\$id/output_sp.zarr"
+output_state: "\$id/state.yaml"
+publish_dir: "$publish_dir"
+HERE
+
+launch_batch /tmp/params_allen.yaml "allen_brain_cell_atlas_merfish"
diff --git a/scripts/run_benchmark/param_sweep/run_test_baysor_nebius.sh b/scripts/run_benchmark/param_sweep/run_test_baysor_nebius.sh
new file mode 100644
index 000000000..dd1d42b4e
--- /dev/null
+++ b/scripts/run_benchmark/param_sweep/run_test_baysor_nebius.sh
@@ -0,0 +1,96 @@
+#!/bin/bash
+
+# Nebius test run: all default methods + baysor at the transcript-assignment stage,
+# with a parameter sweep over baysor's tuning knobs.
+# See src/methods_transcript_assignment/baysor/NOTES.md ("Optimization / tuning").
+#
+# The stage default (basic_transcript_assignment) stays enabled alongside baysor so the run
+# also produces the baseline the sweep is scored against — the workflow allows at most ONE
+# non-default variant at a time, and baysor is that one non-default method here; every OTHER
+# stage stays on its single default.
+#
+# PARAMS-FILE CAVEAT: `method_parameters_yaml` is opened by the WORKFLOW at runtime on the cloud
+# (readYaml -> Nextflow file()), so a local /tmp path won't exist there and /scratch is read-only
+# from the launch host. file() DOES stage http(s):// and this repo is public, so the sweep lives
+# in a COMMITTED file read from GitHub via its raw URL => commit AND PUSH
+# scripts/run_benchmark/param_sweep/baysor_params.yaml to $params_branch before launching.
+# This is independent of --revision (which selects the pipeline CODE).
+
+# get the root of the directory
+REPO_ROOT=$(git rev-parse --show-toplevel)
+
+# ensure that the command below is run from the root of the repository
+cd "$REPO_ROOT"
+
+set -e
+
+resources_test_s3="/scratch/task_ist_preprocessing/resources_test/task_ist_preprocessing/"
+# Results publish to /scratch — created and written by the cloud compute env (read-only here).
+publish_dir="/scratch/results/runs/$(date +%Y-%m-%d_%H-%M-%S)_baysor"
+
+# The sweep lives in a committed file, read from GitHub at runtime. $params_branch defaults to
+# the branch you are on; the file must be pushed there. (Independent of --revision below.)
+params_repo="openproblems-bio/task_ist_preprocessing"
+params_branch="$(git rev-parse --abbrev-ref HEAD)"
+params_url="https://raw.githubusercontent.com/${params_repo}/${params_branch}/scripts/run_benchmark/param_sweep/baysor_params.yaml"
+
+cat > /tmp/params_settings.yaml << HERE
+default_methods:
+  - custom_segmentation
+  - basic_transcript_assignment
+  - basic_count_aggregation
+  - basic_qc_filter
+  - alpha_shapes
+  - normalize_by_volume
+  - tacco
+  - no_correction
+segmentation_methods:
+  - custom_segmentation
+transcript_assignment_methods:
+  - basic_transcript_assignment
+  - baysor
+count_aggregation_methods:
+  - basic_count_aggregation
+qc_filtering_methods:
+  - basic_qc_filter
+volume_calculation_methods:
+  - alpha_shapes
+normalization_methods:
+  - normalize_by_volume
+celltype_annotation_methods:
+  - tacco
+expression_correction_methods:
+  - no_correction
+gene_efficiency_correction_methods:
+  - no_correction
+method_parameters_yaml: $params_url
+HERE
+
+# Write the parameters to file (input_states version, NOTE: enable `-entry_name auto` for this)
+cat > /tmp/params.yaml << HERE
+input_states: $resources_test_s3/**/state.yaml
+rename_keys: 'input_sc:output_sc;input_sp:output_sp'
+save_spatial_data: false
+settings: '$(yq -o json /tmp/params_settings.yaml | jq -c .)'
+output_state: "state.yaml"
+publish_dir: "$publish_dir"
+HERE
+
+# Fail early with a clear message if the params file isn't reachable on GitHub yet.
+if ! curl -fsSL -o /dev/null "$params_url"; then
+  echo "ERROR: params file not reachable at:" >&2
+  echo "  $params_url" >&2
+  echo "Commit and push scripts/run_benchmark/param_sweep/baysor_params.yaml to '$params_branch' first." >&2
+  exit 1
+fi
+
+tw launch https://github.com/openproblems-bio/task_ist_preprocessing.git \
+  --revision build/main \
+  --pull-latest \
+  --main-script target/nextflow/workflows/run_benchmark/main.nf \
+  --workspace 167877437119966 \
+  --compute-env 5hfmdCBxMRd4nHZaJKYEQZ \
+  --params-file /tmp/params.yaml \
+  --entry-name auto \
+  --config src/base/labels_nebius.config \
+  --labels task_ist_preprocessing,test,baysor
diff --git a/scripts/run_benchmark/param_sweep/run_test_binning_nebius.sh b/scripts/run_benchmark/param_sweep/run_test_binning_nebius.sh
index 527e0d3d6..3cab53120 100644
--- a/scripts/run_benchmark/param_sweep/run_test_binning_nebius.sh
+++ b/scripts/run_benchmark/param_sweep/run_test_binning_nebius.sh
@@ -31,7 +31,7 @@ cd "$REPO_ROOT"
 
 set -e
 
-resources_test_s3=s3://openproblems-data/resources_test/task_ist_preprocessing
+resources_test_s3="/scratch/task_ist_preprocessing/resources_test/task_ist_preprocessing/"
 # Results publish to /scratch — created and written by the cloud compute env, so
 # the launcher does NOT create it here (it is read-only from the launch host).
 publish_dir="/scratch/results/runs/$(date +%Y-%m-%d_%H-%M-%S)_binning"
diff --git a/scripts/run_benchmark/param_sweep/run_test_cellpose_nebius.sh b/scripts/run_benchmark/param_sweep/run_test_cellpose_nebius.sh
index e3458e9a6..05d49e0bb 100644
--- a/scripts/run_benchmark/param_sweep/run_test_cellpose_nebius.sh
+++ b/scripts/run_benchmark/param_sweep/run_test_cellpose_nebius.sh
@@ -29,7 +29,7 @@ cd "$REPO_ROOT"
 
 set -e
 
-resources_test_s3=s3://openproblems-data/resources_test/task_ist_preprocessing
+resources_test_s3="/scratch/task_ist_preprocessing/resources_test/task_ist_preprocessing/"
 # Results publish to /scratch — created and written by the cloud compute env, so
 # the launcher does NOT create it here (it is read-only from the launch host).
 publish_dir="/scratch/results/runs/$(date +%Y-%m-%d_%H-%M-%S)_cellpose"
diff --git a/scripts/run_benchmark/param_sweep/run_test_cellposev4_nebius.sh b/scripts/run_benchmark/param_sweep/run_test_cellposev4_nebius.sh
index 5c092195e..b2877beef 100644
--- a/scripts/run_benchmark/param_sweep/run_test_cellposev4_nebius.sh
+++ b/scripts/run_benchmark/param_sweep/run_test_cellposev4_nebius.sh
@@ -15,7 +15,7 @@
 #   publish) is READ-ONLY from the launch host — which is why the binning
 #   method_params block is commented out in run_test_nebius.sh. file() does stage
 #   http(s):// though, and this repo is public, so we keep the sweep in a COMMITTED
-#   file (scripts/run_benchmark/cellposev4_params.yaml) and read it from GitHub via
+#   file (scripts/run_benchmark/param_sweep/cellposev4_params.yaml) and read it from GitHub via
 #   its raw URL. => the params file must be committed AND PUSHED to $params_branch
 #   before launching (edit the file there, not here, to change the sweep).
 
@@ -27,7 +27,7 @@ cd "$REPO_ROOT"
 
 set -e
 
-resources_test_s3=s3://openproblems-data/resources_test/task_ist_preprocessing
+resources_test_s3="/scratch/task_ist_preprocessing/resources_test/task_ist_preprocessing/"
 # Results publish to /scratch — created and written by the cloud compute env, so
 # the launcher does NOT create it here (it is read-only from the launch host).
 publish_dir="/scratch/results/runs/$(date +%Y-%m-%d_%H-%M-%S)_cellposev4"
@@ -37,7 +37,7 @@ publish_dir="/scratch/results/runs/$(date +%Y-%m-%d_%H-%M-%S)_cellposev4"
 # is independent of --revision below, which selects the pipeline CODE to run.)
 params_repo="openproblems-bio/task_ist_preprocessing"
 params_branch="$(git rev-parse --abbrev-ref HEAD)"
-params_url="https://raw.githubusercontent.com/${params_repo}/${params_branch}/scripts/run_benchmark/cellposev4_params.yaml"
+params_url="https://raw.githubusercontent.com/${params_repo}/${params_branch}/scripts/run_benchmark/param_sweep/cellposev4_params.yaml"
 
 cat > /tmp/params_settings.yaml << HERE
 default_methods:
@@ -89,7 +89,7 @@ HERE
 if ! curl -fsSL -o /dev/null "$params_url"; then
   echo "ERROR: params file not reachable at:" >&2
   echo "  $params_url" >&2
-  echo "Commit and push scripts/run_benchmark/cellposev4_params.yaml to '$params_branch' first." >&2
+  echo "Commit and push scripts/run_benchmark/param_sweep/cellposev4_params.yaml to '$params_branch' first." >&2
   exit 1
 fi
 
diff --git a/scripts/run_benchmark/param_sweep/run_test_clustermap_nebius.sh b/scripts/run_benchmark/param_sweep/run_test_clustermap_nebius.sh
new file mode 100644
index 000000000..4183a0b4a
--- /dev/null
+++ b/scripts/run_benchmark/param_sweep/run_test_clustermap_nebius.sh
@@ -0,0 +1,96 @@
+#!/bin/bash
+
+# Nebius test run: all default methods + clustermap at the transcript-assignment stage,
+# with a parameter sweep over clustermap's tuning knobs.
+# See src/methods_transcript_assignment/clustermap/NOTES.md ("Optimization / tuning").
+#
+# The stage default (basic_transcript_assignment) stays enabled alongside clustermap so the run
+# also produces the baseline the sweep is scored against — the workflow allows at most ONE
+# non-default variant at a time, and clustermap is that one non-default method here; every OTHER
+# stage stays on its single default.
+#
+# PARAMS-FILE CAVEAT: `method_parameters_yaml` is opened by the WORKFLOW at runtime on the cloud
+# (readYaml -> Nextflow file()), so a local /tmp path won't exist there and /scratch is read-only
+# from the launch host. file() DOES stage http(s):// and this repo is public, so the sweep lives
+# in a COMMITTED file read from GitHub via its raw URL => commit AND PUSH
+# scripts/run_benchmark/param_sweep/clustermap_params.yaml to $params_branch before launching.
+# This is independent of --revision (which selects the pipeline CODE).
+
+# get the root of the directory
+REPO_ROOT=$(git rev-parse --show-toplevel)
+
+# ensure that the command below is run from the root of the repository
+cd "$REPO_ROOT"
+
+set -e
+
+resources_test_s3="/scratch/task_ist_preprocessing/resources_test/task_ist_preprocessing/"
+# Results publish to /scratch — created and written by the cloud compute env (read-only here).
+publish_dir="/scratch/results/runs/$(date +%Y-%m-%d_%H-%M-%S)_clustermap"
+
+# The sweep lives in a committed file, read from GitHub at runtime. $params_branch defaults to
+# the branch you are on; the file must be pushed there. (Independent of --revision below.)
+params_repo="openproblems-bio/task_ist_preprocessing"
+params_branch="$(git rev-parse --abbrev-ref HEAD)"
+params_url="https://raw.githubusercontent.com/${params_repo}/${params_branch}/scripts/run_benchmark/param_sweep/clustermap_params.yaml"
+
+cat > /tmp/params_settings.yaml << HERE
+default_methods:
+  - custom_segmentation
+  - basic_transcript_assignment
+  - basic_count_aggregation
+  - basic_qc_filter
+  - alpha_shapes
+  - normalize_by_volume
+  - tacco
+  - no_correction
+segmentation_methods:
+  - custom_segmentation
+transcript_assignment_methods:
+  - basic_transcript_assignment
+  - clustermap
+count_aggregation_methods:
+  - basic_count_aggregation
+qc_filtering_methods:
+  - basic_qc_filter
+volume_calculation_methods:
+  - alpha_shapes
+normalization_methods:
+  - normalize_by_volume
+celltype_annotation_methods:
+  - tacco
+expression_correction_methods:
+  - no_correction
+gene_efficiency_correction_methods:
+  - no_correction
+method_parameters_yaml: $params_url
+HERE
+
+# Write the parameters to file (input_states version, NOTE: enable `-entry_name auto` for this)
+cat > /tmp/params.yaml << HERE
+input_states: $resources_test_s3/**/state.yaml
+rename_keys: 'input_sc:output_sc;input_sp:output_sp'
+save_spatial_data: false
+settings: '$(yq -o json /tmp/params_settings.yaml | jq -c .)'
+output_state: "state.yaml"
+publish_dir: "$publish_dir"
+HERE
+
+# Fail early with a clear message if the params file isn't reachable on GitHub yet.
+if ! curl -fsSL -o /dev/null "$params_url"; then
+  echo "ERROR: params file not reachable at:" >&2
+  echo "  $params_url" >&2
+  echo "Commit and push scripts/run_benchmark/param_sweep/clustermap_params.yaml to '$params_branch' first." >&2
+  exit 1
+fi
+
+tw launch https://github.com/openproblems-bio/task_ist_preprocessing.git \
+  --revision build/main \
+  --pull-latest \
+  --main-script target/nextflow/workflows/run_benchmark/main.nf \
+  --workspace 167877437119966 \
+  --compute-env 5hfmdCBxMRd4nHZaJKYEQZ \
+  --params-file /tmp/params.yaml \
+  --entry-name auto \
+  --config src/base/labels_nebius.config \
+  --labels task_ist_preprocessing,test,clustermap
diff --git a/scripts/run_benchmark/param_sweep/run_test_comseg_nebius.sh b/scripts/run_benchmark/param_sweep/run_test_comseg_nebius.sh
new file mode 100644
index 000000000..d31e7d018
--- /dev/null
+++ b/scripts/run_benchmark/param_sweep/run_test_comseg_nebius.sh
@@ -0,0 +1,96 @@
+#!/bin/bash
+
+# Nebius test run: all default methods + comseg at the transcript-assignment stage,
+# with a parameter sweep over comseg's tuning knobs.
+# See src/methods_transcript_assignment/comseg/NOTES.md ("Optimization / tuning").
+#
+# The stage default (basic_transcript_assignment) stays enabled alongside comseg so the run
+# also produces the baseline the sweep is scored against — the workflow allows at most ONE
+# non-default variant at a time, and comseg is that one non-default method here; every OTHER
+# stage stays on its single default.
+#
+# PARAMS-FILE CAVEAT: `method_parameters_yaml` is opened by the WORKFLOW at runtime on the cloud
+# (readYaml -> Nextflow file()), so a local /tmp path won't exist there and /scratch is read-only
+# from the launch host. file() DOES stage http(s):// and this repo is public, so the sweep lives
+# in a COMMITTED file read from GitHub via its raw URL => commit AND PUSH
+# scripts/run_benchmark/param_sweep/comseg_params.yaml to $params_branch before launching.
+# This is independent of --revision (which selects the pipeline CODE).
+
+# get the root of the directory
+REPO_ROOT=$(git rev-parse --show-toplevel)
+
+# ensure that the command below is run from the root of the repository
+cd "$REPO_ROOT"
+
+set -e
+
+resources_test_s3="/scratch/task_ist_preprocessing/resources_test/task_ist_preprocessing/"
+# Results publish to /scratch — created and written by the cloud compute env (read-only here).
+publish_dir="/scratch/results/runs/$(date +%Y-%m-%d_%H-%M-%S)_comseg"
+
+# The sweep lives in a committed file, read from GitHub at runtime. $params_branch defaults to
+# the branch you are on; the file must be pushed there. (Independent of --revision below.)
+params_repo="openproblems-bio/task_ist_preprocessing"
+params_branch="$(git rev-parse --abbrev-ref HEAD)"
+params_url="https://raw.githubusercontent.com/${params_repo}/${params_branch}/scripts/run_benchmark/param_sweep/comseg_params.yaml"
+
+cat > /tmp/params_settings.yaml << HERE
+default_methods:
+  - custom_segmentation
+  - basic_transcript_assignment
+  - basic_count_aggregation
+  - basic_qc_filter
+  - alpha_shapes
+  - normalize_by_volume
+  - tacco
+  - no_correction
+segmentation_methods:
+  - custom_segmentation
+transcript_assignment_methods:
+  - basic_transcript_assignment
+  - comseg
+count_aggregation_methods:
+  - basic_count_aggregation
+qc_filtering_methods:
+  - basic_qc_filter
+volume_calculation_methods:
+  - alpha_shapes
+normalization_methods:
+  - normalize_by_volume
+celltype_annotation_methods:
+  - tacco
+expression_correction_methods:
+  - no_correction
+gene_efficiency_correction_methods:
+  - no_correction
+method_parameters_yaml: $params_url
+HERE
+
+# Write the parameters to file (input_states version, NOTE: enable `-entry_name auto` for this)
+cat > /tmp/params.yaml << HERE
+input_states: $resources_test_s3/**/state.yaml
+rename_keys: 'input_sc:output_sc;input_sp:output_sp'
+save_spatial_data: false
+settings: '$(yq -o json /tmp/params_settings.yaml | jq -c .)'
+output_state: "state.yaml"
+publish_dir: "$publish_dir"
+HERE
+
+# Fail early with a clear message if the params file isn't reachable on GitHub yet.
+if ! curl -fsSL -o /dev/null "$params_url"; then
+  echo "ERROR: params file not reachable at:" >&2
+  echo "  $params_url" >&2
+  echo "Commit and push scripts/run_benchmark/param_sweep/comseg_params.yaml to '$params_branch' first." >&2
+  exit 1
+fi
+
+tw launch https://github.com/openproblems-bio/task_ist_preprocessing.git \
+  --revision build/main \
+  --pull-latest \
+  --main-script target/nextflow/workflows/run_benchmark/main.nf \
+  --workspace 167877437119966 \
+  --compute-env 5hfmdCBxMRd4nHZaJKYEQZ \
+  --params-file /tmp/params.yaml \
+  --entry-name auto \
+  --config src/base/labels_nebius.config \
+  --labels task_ist_preprocessing,test,comseg
diff --git a/scripts/run_benchmark/param_sweep/run_test_fastreseg_nebius.sh b/scripts/run_benchmark/param_sweep/run_test_fastreseg_nebius.sh
new file mode 100644
index 000000000..acaebb3f7
--- /dev/null
+++ b/scripts/run_benchmark/param_sweep/run_test_fastreseg_nebius.sh
@@ -0,0 +1,96 @@
+#!/bin/bash
+
+# Nebius test run: all default methods + fastreseg at the transcript-assignment stage,
+# with a parameter sweep over fastreseg's tuning knobs.
+# See src/methods_transcript_assignment/fastreseg/NOTES.md ("Optimization / tuning").
+#
+# The stage default (basic_transcript_assignment) stays enabled alongside fastreseg so the run
+# also produces the baseline the sweep is scored against — the workflow allows at most ONE
+# non-default variant at a time, and fastreseg is that one non-default method here; every OTHER
+# stage stays on its single default.
+#
+# PARAMS-FILE CAVEAT: `method_parameters_yaml` is opened by the WORKFLOW at runtime on the cloud
+# (readYaml -> Nextflow file()), so a local /tmp path won't exist there and /scratch is read-only
+# from the launch host. file() DOES stage http(s):// and this repo is public, so the sweep lives
+# in a COMMITTED file read from GitHub via its raw URL => commit AND PUSH
+# scripts/run_benchmark/param_sweep/fastreseg_params.yaml to $params_branch before launching.
+# This is independent of --revision (which selects the pipeline CODE).
+
+# get the root of the directory
+REPO_ROOT=$(git rev-parse --show-toplevel)
+
+# ensure that the command below is run from the root of the repository
+cd "$REPO_ROOT"
+
+set -e
+
+resources_test_s3="/scratch/task_ist_preprocessing/resources_test/task_ist_preprocessing/"
+# Results publish to /scratch — created and written by the cloud compute env (read-only here).
+publish_dir="/scratch/results/runs/$(date +%Y-%m-%d_%H-%M-%S)_fastreseg"
+
+# The sweep lives in a committed file, read from GitHub at runtime. $params_branch defaults to
+# the branch you are on; the file must be pushed there. (Independent of --revision below.)
+params_repo="openproblems-bio/task_ist_preprocessing"
+params_branch="$(git rev-parse --abbrev-ref HEAD)"
+params_url="https://raw.githubusercontent.com/${params_repo}/${params_branch}/scripts/run_benchmark/param_sweep/fastreseg_params.yaml"
+
+cat > /tmp/params_settings.yaml << HERE
+default_methods:
+  - custom_segmentation
+  - basic_transcript_assignment
+  - basic_count_aggregation
+  - basic_qc_filter
+  - alpha_shapes
+  - normalize_by_volume
+  - tacco
+  - no_correction
+segmentation_methods:
+  - custom_segmentation
+transcript_assignment_methods:
+  - basic_transcript_assignment
+  - fastreseg
+count_aggregation_methods:
+  - basic_count_aggregation
+qc_filtering_methods:
+  - basic_qc_filter
+volume_calculation_methods:
+  - alpha_shapes
+normalization_methods:
+  - normalize_by_volume
+celltype_annotation_methods:
+  - tacco
+expression_correction_methods:
+  - no_correction
+gene_efficiency_correction_methods:
+  - no_correction
+method_parameters_yaml: $params_url
+HERE
+
+# Write the parameters to file (input_states version, NOTE: enable `-entry_name auto` for this)
+cat > /tmp/params.yaml << HERE
+input_states: $resources_test_s3/**/state.yaml
+rename_keys: 'input_sc:output_sc;input_sp:output_sp'
+save_spatial_data: false
+settings: '$(yq -o json /tmp/params_settings.yaml | jq -c .)'
+output_state: "state.yaml"
+publish_dir: "$publish_dir"
+HERE
+
+# Fail early with a clear message if the params file isn't reachable on GitHub yet.
+if ! curl -fsSL -o /dev/null "$params_url"; then
+  echo "ERROR: params file not reachable at:" >&2
+  echo "  $params_url" >&2
+  echo "Commit and push scripts/run_benchmark/param_sweep/fastreseg_params.yaml to '$params_branch' first." >&2
+  exit 1
+fi
+
+tw launch https://github.com/openproblems-bio/task_ist_preprocessing.git \
+  --revision build/main \
+  --pull-latest \
+  --main-script target/nextflow/workflows/run_benchmark/main.nf \
+  --workspace 167877437119966 \
+  --compute-env 5hfmdCBxMRd4nHZaJKYEQZ \
+  --params-file /tmp/params.yaml \
+  --entry-name auto \
+  --config src/base/labels_nebius.config \
+  --labels task_ist_preprocessing,test,fastreseg
diff --git a/scripts/run_benchmark/param_sweep/run_test_moscot_nebius.sh b/scripts/run_benchmark/param_sweep/run_test_moscot_nebius.sh
index 7319cfb74..27d1a5c6e 100644
--- a/scripts/run_benchmark/param_sweep/run_test_moscot_nebius.sh
+++ b/scripts/run_benchmark/param_sweep/run_test_moscot_nebius.sh
@@ -36,7 +36,7 @@ cd "$REPO_ROOT"
 
 set -e
 
-resources_test_s3=s3://openproblems-data/resources_test/task_ist_preprocessing
+resources_test_s3="/scratch/task_ist_preprocessing/resources_test/task_ist_preprocessing/"
 # Results publish to /scratch — created and written by the cloud compute env, so
 # the launcher does NOT create it here (it is read-only from the launch host).
 publish_dir="/scratch/results/runs/$(date +%Y-%m-%d_%H-%M-%S)_moscot"
diff --git a/scripts/run_benchmark/param_sweep/run_test_pciseq_nebius.sh b/scripts/run_benchmark/param_sweep/run_test_pciseq_nebius.sh
new file mode 100644
index 000000000..be965b33c
--- /dev/null
+++ b/scripts/run_benchmark/param_sweep/run_test_pciseq_nebius.sh
@@ -0,0 +1,96 @@
+#!/bin/bash
+
+# Nebius test run: all default methods + pciseq at the transcript-assignment stage,
+# with a parameter sweep over pciseq's tuning knobs.
+# See src/methods_transcript_assignment/pciseq/NOTES.md ("Optimization / tuning").
+#
+# The stage default (basic_transcript_assignment) stays enabled alongside pciseq so the run
+# also produces the baseline the sweep is scored against — the workflow allows at most ONE
+# non-default variant at a time, and pciseq is that one non-default method here; every OTHER
+# stage stays on its single default.
+#
+# PARAMS-FILE CAVEAT: `method_parameters_yaml` is opened by the WORKFLOW at runtime on the cloud
+# (readYaml -> Nextflow file()), so a local /tmp path won't exist there and /scratch is read-only
+# from the launch host. file() DOES stage http(s):// and this repo is public, so the sweep lives
+# in a COMMITTED file read from GitHub via its raw URL => commit AND PUSH
+# scripts/run_benchmark/param_sweep/pciseq_params.yaml to $params_branch before launching.
+# This is independent of --revision (which selects the pipeline CODE).
+
+# get the root of the directory
+REPO_ROOT=$(git rev-parse --show-toplevel)
+
+# ensure that the command below is run from the root of the repository
+cd "$REPO_ROOT"
+
+set -e
+
+resources_test_s3="/scratch/task_ist_preprocessing/resources_test/task_ist_preprocessing/"
+# Results publish to /scratch — created and written by the cloud compute env (read-only here).
+publish_dir="/scratch/results/runs/$(date +%Y-%m-%d_%H-%M-%S)_pciseq"
+
+# The sweep lives in a committed file, read from GitHub at runtime. $params_branch defaults to
+# the branch you are on; the file must be pushed there. (Independent of --revision below.)
+params_repo="openproblems-bio/task_ist_preprocessing"
+params_branch="$(git rev-parse --abbrev-ref HEAD)"
+params_url="https://raw.githubusercontent.com/${params_repo}/${params_branch}/scripts/run_benchmark/param_sweep/pciseq_params.yaml"
+
+cat > /tmp/params_settings.yaml << HERE
+default_methods:
+  - custom_segmentation
+  - basic_transcript_assignment
+  - basic_count_aggregation
+  - basic_qc_filter
+  - alpha_shapes
+  - normalize_by_volume
+  - tacco
+  - no_correction
+segmentation_methods:
+  - custom_segmentation
+transcript_assignment_methods:
+  - basic_transcript_assignment
+  - pciseq
+count_aggregation_methods:
+  - basic_count_aggregation
+qc_filtering_methods:
+  - basic_qc_filter
+volume_calculation_methods:
+  - alpha_shapes
+normalization_methods:
+  - normalize_by_volume
+celltype_annotation_methods:
+  - tacco
+expression_correction_methods:
+  - no_correction
+gene_efficiency_correction_methods:
+  - no_correction
+method_parameters_yaml: $params_url
+HERE
+
+# Write the parameters to file (input_states version, NOTE: enable `-entry_name auto` for this)
+cat > /tmp/params.yaml << HERE
+input_states: $resources_test_s3/**/state.yaml
+rename_keys: 'input_sc:output_sc;input_sp:output_sp'
+save_spatial_data: false
+settings: '$(yq -o json /tmp/params_settings.yaml | jq -c .)'
+output_state: "state.yaml"
+publish_dir: "$publish_dir"
+HERE
+
+# Fail early with a clear message if the params file isn't reachable on GitHub yet.
+if ! curl -fsSL -o /dev/null "$params_url"; then
+  echo "ERROR: params file not reachable at:" >&2
+  echo "  $params_url" >&2
+  echo "Commit and push scripts/run_benchmark/param_sweep/pciseq_params.yaml to '$params_branch' first." >&2
+  exit 1
+fi
+
+tw launch https://github.com/openproblems-bio/task_ist_preprocessing.git \
+  --revision build/main \
+  --pull-latest \
+  --main-script target/nextflow/workflows/run_benchmark/main.nf \
+  --workspace 167877437119966 \
+  --compute-env 5hfmdCBxMRd4nHZaJKYEQZ \
+  --params-file /tmp/params.yaml \
+  --entry-name auto \
+  --config src/base/labels_nebius.config \
+  --labels task_ist_preprocessing,test,pciseq
diff --git a/scripts/run_benchmark/param_sweep/run_test_proseg_nebius.sh b/scripts/run_benchmark/param_sweep/run_test_proseg_nebius.sh
new file mode 100644
index 000000000..bbcff5c34
--- /dev/null
+++ b/scripts/run_benchmark/param_sweep/run_test_proseg_nebius.sh
@@ -0,0 +1,96 @@
+#!/bin/bash
+
+# Nebius test run: all default methods + proseg at the transcript-assignment stage,
+# with a parameter sweep over proseg's tuning knobs.
+# See src/methods_transcript_assignment/proseg/NOTES.md ("Optimization / tuning").
+#
+# The stage default (basic_transcript_assignment) stays enabled alongside proseg so the run
+# also produces the baseline the sweep is scored against — the workflow allows at most ONE
+# non-default variant at a time, and proseg is that one non-default method here; every OTHER
+# stage stays on its single default.
+#
+# PARAMS-FILE CAVEAT: `method_parameters_yaml` is opened by the WORKFLOW at runtime on the cloud
+# (readYaml -> Nextflow file()), so a local /tmp path won't exist there and /scratch is read-only
+# from the launch host. file() DOES stage http(s):// and this repo is public, so the sweep lives
+# in a COMMITTED file read from GitHub via its raw URL => commit AND PUSH
+# scripts/run_benchmark/param_sweep/proseg_params.yaml to $params_branch before launching.
+# This is independent of --revision (which selects the pipeline CODE).
+
+# get the root of the directory
+REPO_ROOT=$(git rev-parse --show-toplevel)
+
+# ensure that the command below is run from the root of the repository
+cd "$REPO_ROOT"
+
+set -e
+
+resources_test_s3="/scratch/task_ist_preprocessing/resources_test/task_ist_preprocessing/"
+# Results publish to /scratch — created and written by the cloud compute env (read-only here).
+publish_dir="/scratch/results/runs/$(date +%Y-%m-%d_%H-%M-%S)_proseg"
+
+# The sweep lives in a committed file, read from GitHub at runtime. $params_branch defaults to
+# the branch you are on; the file must be pushed there. (Independent of --revision below.)
+params_repo="openproblems-bio/task_ist_preprocessing"
+params_branch="$(git rev-parse --abbrev-ref HEAD)"
+params_url="https://raw.githubusercontent.com/${params_repo}/${params_branch}/scripts/run_benchmark/param_sweep/proseg_params.yaml"
+
+cat > /tmp/params_settings.yaml << HERE
+default_methods:
+  - custom_segmentation
+  - basic_transcript_assignment
+  - basic_count_aggregation
+  - basic_qc_filter
+  - alpha_shapes
+  - normalize_by_volume
+  - tacco
+  - no_correction
+segmentation_methods:
+  - custom_segmentation
+transcript_assignment_methods:
+  - basic_transcript_assignment
+  - proseg
+count_aggregation_methods:
+  - basic_count_aggregation
+qc_filtering_methods:
+  - basic_qc_filter
+volume_calculation_methods:
+  - alpha_shapes
+normalization_methods:
+  - normalize_by_volume
+celltype_annotation_methods:
+  - tacco
+expression_correction_methods:
+  - no_correction
+gene_efficiency_correction_methods:
+  - no_correction
+method_parameters_yaml: $params_url
+HERE
+
+# Write the parameters to file (input_states version, NOTE: enable `-entry_name auto` for this)
+cat > /tmp/params.yaml << HERE
+input_states: $resources_test_s3/**/state.yaml
+rename_keys: 'input_sc:output_sc;input_sp:output_sp'
+save_spatial_data: false
+settings: '$(yq -o json /tmp/params_settings.yaml | jq -c .)'
+output_state: "state.yaml"
+publish_dir: "$publish_dir"
+HERE
+
+# Fail early with a clear message if the params file isn't reachable on GitHub yet.
+if ! curl -fsSL -o /dev/null "$params_url"; then
+  echo "ERROR: params file not reachable at:" >&2
+  echo "  $params_url" >&2
+  echo "Commit and push scripts/run_benchmark/param_sweep/proseg_params.yaml to '$params_branch' first." >&2
+  exit 1
+fi
+
+tw launch https://github.com/openproblems-bio/task_ist_preprocessing.git \
+  --revision build/main \
+  --pull-latest \
+  --main-script target/nextflow/workflows/run_benchmark/main.nf \
+  --workspace 167877437119966 \
+  --compute-env 5hfmdCBxMRd4nHZaJKYEQZ \
+  --params-file /tmp/params.yaml \
+  --entry-name auto \
+  --config src/base/labels_nebius.config \
+  --labels task_ist_preprocessing,test,proseg
diff --git a/scripts/run_benchmark/param_sweep/run_test_rctd_nebius.sh b/scripts/run_benchmark/param_sweep/run_test_rctd_nebius.sh
index 438c3a14e..5144b1d94 100644
--- a/scripts/run_benchmark/param_sweep/run_test_rctd_nebius.sh
+++ b/scripts/run_benchmark/param_sweep/run_test_rctd_nebius.sh
@@ -34,7 +34,7 @@ cd "$REPO_ROOT"
 
 set -e
 
-resources_test_s3=s3://openproblems-data/resources_test/task_ist_preprocessing
+resources_test_s3="/scratch/task_ist_preprocessing/resources_test/task_ist_preprocessing/"
 # Results publish to /scratch — created and written by the cloud compute env, so
 # the launcher does NOT create it here (it is read-only from the launch host).
 publish_dir="/scratch/results/runs/$(date +%Y-%m-%d_%H-%M-%S)_rctd"
diff --git a/scripts/run_benchmark/param_sweep/run_test_resolvi_correction_nebius.sh b/scripts/run_benchmark/param_sweep/run_test_resolvi_correction_nebius.sh
index a96ab96f0..f09969f85 100644
--- a/scripts/run_benchmark/param_sweep/run_test_resolvi_correction_nebius.sh
+++ b/scripts/run_benchmark/param_sweep/run_test_resolvi_correction_nebius.sh
@@ -29,7 +29,7 @@ cd "$REPO_ROOT"
 
 set -e
 
-resources_test_s3=s3://openproblems-data/resources_test/task_ist_preprocessing
+resources_test_s3="/scratch/task_ist_preprocessing/resources_test/task_ist_preprocessing/"
 # Results publish to /scratch — created and written by the cloud compute env, so the launcher
 # does NOT create it here (it is read-only from the launch host).
 publish_dir="/scratch/results/runs/$(date +%Y-%m-%d_%H-%M-%S)_resolvi_correction"
diff --git a/scripts/run_benchmark/param_sweep/run_test_segger_nebius.sh b/scripts/run_benchmark/param_sweep/run_test_segger_nebius.sh
new file mode 100644
index 000000000..6763f1df3
--- /dev/null
+++ b/scripts/run_benchmark/param_sweep/run_test_segger_nebius.sh
@@ -0,0 +1,99 @@
+#!/bin/bash
+
+# Nebius test run: all default methods + segger at the transcript-assignment stage,
+# with a parameter sweep over segger's tuning knobs.
+# See src/methods_transcript_assignment/segger/NOTES.md ("Optimization / tuning").
+# segger is GPU-only: the `gpuh100` label in its config + src/base/labels_nebius.config
+# (runAsUser:0 + /dev/shm volume) pin it to the GPU node group — no GPU-specific change is
+# needed here (same as the existing run_test_segger_nebius.sh).
+#
+# The stage default (basic_transcript_assignment) stays enabled alongside segger so the run
+# also produces the baseline the sweep is scored against — the workflow allows at most ONE
+# non-default variant at a time, and segger is that one non-default method here; every OTHER
+# stage stays on its single default.
+#
+# PARAMS-FILE CAVEAT: `method_parameters_yaml` is opened by the WORKFLOW at runtime on the cloud
+# (readYaml -> Nextflow file()), so a local /tmp path won't exist there and /scratch is read-only
+# from the launch host. file() DOES stage http(s):// and this repo is public, so the sweep lives
+# in a COMMITTED file read from GitHub via its raw URL => commit AND PUSH
+# scripts/run_benchmark/param_sweep/segger_params.yaml to $params_branch before launching.
+# This is independent of --revision (which selects the pipeline CODE).
+
+# get the root of the directory
+REPO_ROOT=$(git rev-parse --show-toplevel)
+
+# ensure that the command below is run from the root of the repository
+cd "$REPO_ROOT"
+
+set -e
+
+resources_test_s3="/scratch/task_ist_preprocessing/resources_test/task_ist_preprocessing/"
+# Results publish to /scratch — created and written by the cloud compute env (read-only here).
+publish_dir="/scratch/results/runs/$(date +%Y-%m-%d_%H-%M-%S)_segger"
+
+# The sweep lives in a committed file, read from GitHub at runtime. $params_branch defaults to
+# the branch you are on; the file must be pushed there. (Independent of --revision below.)
+params_repo="openproblems-bio/task_ist_preprocessing"
+params_branch="$(git rev-parse --abbrev-ref HEAD)"
+params_url="https://raw.githubusercontent.com/${params_repo}/${params_branch}/scripts/run_benchmark/param_sweep/segger_params.yaml"
+
+cat > /tmp/params_settings.yaml << HERE
+default_methods:
+  - custom_segmentation
+  - basic_transcript_assignment
+  - basic_count_aggregation
+  - basic_qc_filter
+  - alpha_shapes
+  - normalize_by_volume
+  - tacco
+  - no_correction
+segmentation_methods:
+  - custom_segmentation
+transcript_assignment_methods:
+  - basic_transcript_assignment
+  - segger
+count_aggregation_methods:
+  - basic_count_aggregation
+qc_filtering_methods:
+  - basic_qc_filter
+volume_calculation_methods:
+  - alpha_shapes
+normalization_methods:
+  - normalize_by_volume
+celltype_annotation_methods:
+  - tacco
+expression_correction_methods:
+  - no_correction
+gene_efficiency_correction_methods:
+  - no_correction
+method_parameters_yaml: $params_url
+HERE
+
+# Write the parameters to file (input_states version, NOTE: enable `-entry_name auto` for this)
+cat > /tmp/params.yaml << HERE
+input_states: $resources_test_s3/**/state.yaml
+rename_keys: 'input_sc:output_sc;input_sp:output_sp'
+save_spatial_data: false
+settings: '$(yq -o json /tmp/params_settings.yaml | jq -c .)'
+output_state: "state.yaml"
+publish_dir: "$publish_dir"
+HERE
+
+# Fail early with a clear message if the params file isn't reachable on GitHub yet.
+if ! curl -fsSL -o /dev/null "$params_url"; then
+  echo "ERROR: params file not reachable at:" >&2
+  echo "  $params_url" >&2
+  echo "Commit and push scripts/run_benchmark/param_sweep/segger_params.yaml to '$params_branch' first." >&2
+  exit 1
+fi
+
+tw launch https://github.com/openproblems-bio/task_ist_preprocessing.git \
+  --revision build/main \
+  --pull-latest \
+  --main-script target/nextflow/workflows/run_benchmark/main.nf \
+  --workspace 167877437119966 \
+  --compute-env 5hfmdCBxMRd4nHZaJKYEQZ \
+  --params-file /tmp/params.yaml \
+  --entry-name auto \
+  --config src/base/labels_nebius.config \
+  --labels task_ist_preprocessing,test,segger
diff --git a/scripts/run_benchmark/param_sweep/run_test_singler_nebius.sh b/scripts/run_benchmark/param_sweep/run_test_singler_nebius.sh
new file mode 100644
index 000000000..db0323fe7
--- /dev/null
+++ b/scripts/run_benchmark/param_sweep/run_test_singler_nebius.sh
@@ -0,0 +1,113 @@
+#!/bin/bash
+
+# Nebius test run: all default methods + SingleR (singler) cell-type annotation,
+# with a parameter sweep over the one exposed optimization lever for SingleR.
+# See src/methods_cell_type_annotation/singler/NOTES.md ("Optimization / tuning").
+#
+# SingleR (singler-py) is a CPU-only reference-correlation labeller — NO GPU, so
+# there is no `gpu` label here (the non-GPU parts mirror run_test_nebius.sh). The
+# compute env is the same one all these test runs use.
+#
+# The annotation stage keeps its default method `tacco` alongside `singler` so the
+# run has a baseline to compare the singler variants against. Every other stage is
+# left on its single default (the swept method must be the only non-default thing).
+#
+# !!! SUBMITTABLE-BUT-NO-OP CAVEAT !!!
+#   The sweep varies `celltype_key`, which ALREADY EXISTS in build/main's config, so
+#   this launches WITHOUT a rebuild. BUT build/main's baked-in script.py hardcodes the
+#   reference label column (ref_labels = ...column("cell_type")) and never reads
+#   par['celltype_key'] -- so on the deployed build/main container all four variants
+#   produce IDENTICAL annotations. The sweep only measures anything after a one-line
+#   wiring fix (...column(par["celltype_key"])) AND a container rebuild. See NOTES.md.
+#
+# PARAMS-FILE CAVEAT (why this reads from GitHub):
+#   `tw launch --params-file` is read client-side, but `method_parameters_yaml`
+#   is a path the WORKFLOW opens at runtime on the cloud (readYaml -> Nextflow
+#   file()). A local /tmp path does not exist there, and /scratch (where results
+#   publish) is READ-ONLY from the launch host. file() does stage http(s):// though,
+#   and this repo is public, so we keep the sweep in a COMMITTED file
+#   (scripts/run_benchmark/param_sweep/singler_params.yaml) and read it from GitHub
+#   via its raw URL. => the params file must be committed AND PUSHED to
+#   $params_branch before launching (edit the file there, not here, to change the
+#   sweep). This is independent of --revision below, which selects the pipeline CODE.
+
+# get the root of the directory
+REPO_ROOT=$(git rev-parse --show-toplevel)
+
+# ensure that the command below is run from the root of the repository
+cd "$REPO_ROOT"
+
+set -e
+
+resources_test_s3="/scratch/task_ist_preprocessing/resources_test/task_ist_preprocessing/"
+# Results publish to /scratch — created and written by the cloud compute env, so
+# the launcher does NOT create it here (it is read-only from the launch host).
+publish_dir="/scratch/results/runs/$(date +%Y-%m-%d_%H-%M-%S)_singler"
+
+# The sweep lives in a committed file, read from GitHub at runtime. $params_branch
+# defaults to the branch you are on; the file must be pushed there on GitHub. (This
+# is independent of --revision below, which selects the pipeline CODE to run.)
+params_repo="openproblems-bio/task_ist_preprocessing"
+params_branch="$(git rev-parse --abbrev-ref HEAD)"
+params_url="https://raw.githubusercontent.com/${params_repo}/${params_branch}/scripts/run_benchmark/param_sweep/singler_params.yaml"
+
+cat > /tmp/params_settings.yaml << HERE
+default_methods:
+  - custom_segmentation
+  - basic_transcript_assignment
+  - basic_count_aggregation
+  - basic_qc_filter
+  - alpha_shapes
+  - normalize_by_volume
+  - tacco
+  - no_correction
+segmentation_methods:
+  - custom_segmentation
+transcript_assignment_methods:
+  - basic_transcript_assignment
+count_aggregation_methods:
+  - basic_count_aggregation
+qc_filtering_methods:
+  - basic_qc_filter
+volume_calculation_methods:
+  - alpha_shapes
+normalization_methods:
+  - normalize_by_volume
+celltype_annotation_methods:
+  - tacco
+  - singler
+expression_correction_methods:
+  - no_correction
+gene_efficiency_correction_methods:
+  - no_correction
+method_parameters_yaml: $params_url
+HERE
+
+# Write the parameters to file (input_states version, NOTE: enable `-entry_name auto` for this)
+cat > /tmp/params.yaml << HERE
+input_states: $resources_test_s3/**/state.yaml
+rename_keys: 'input_sc:output_sc;input_sp:output_sp'
+save_spatial_data: false
+settings: '$(yq -o json /tmp/params_settings.yaml | jq -c .)'
+output_state: "state.yaml"
+publish_dir: "$publish_dir"
+HERE
+
+# Fail early with a clear message if the params file isn't reachable on GitHub yet.
+if ! curl -fsSL -o /dev/null "$params_url"; then
+  echo "ERROR: params file not reachable at:" >&2
+  echo "  $params_url" >&2
+  echo "Commit and push scripts/run_benchmark/param_sweep/singler_params.yaml to '$params_branch' first." >&2
+  exit 1
+fi
+
+tw launch https://github.com/openproblems-bio/task_ist_preprocessing.git \
+  --revision build/main \
+  --pull-latest \
+  --main-script target/nextflow/workflows/run_benchmark/main.nf \
+  --workspace 167877437119966 \
+  --compute-env 5hfmdCBxMRd4nHZaJKYEQZ \
+  --params-file /tmp/params.yaml \
+  --entry-name auto \
+  --config src/base/labels_nebius.config \
+  --labels task_ist_preprocessing,test,singler
diff --git a/scripts/run_benchmark/param_sweep/run_test_split_nebius.sh b/scripts/run_benchmark/param_sweep/run_test_split_nebius.sh
index 0a20da769..bb63ff97b 100644
--- a/scripts/run_benchmark/param_sweep/run_test_split_nebius.sh
+++ b/scripts/run_benchmark/param_sweep/run_test_split_nebius.sh
@@ -37,7 +37,7 @@ cd "$REPO_ROOT"
 
 set -e
 
-resources_test_s3=s3://openproblems-data/resources_test/task_ist_preprocessing
+resources_test_s3="/scratch/task_ist_preprocessing/resources_test/task_ist_preprocessing/"
 # Results publish to /scratch — created and written by the cloud compute env, so
 # the launcher does NOT create it here (it is read-only from the launch host).
 publish_dir="/scratch/results/runs/$(date +%Y-%m-%d_%H-%M-%S)_split"
diff --git a/scripts/run_benchmark/param_sweep/run_test_ssam_nebius.sh b/scripts/run_benchmark/param_sweep/run_test_ssam_nebius.sh
index 7ac988081..6d0517717 100644
--- a/scripts/run_benchmark/param_sweep/run_test_ssam_nebius.sh
+++ b/scripts/run_benchmark/param_sweep/run_test_ssam_nebius.sh
@@ -30,7 +30,7 @@ cd "$REPO_ROOT"
 
 set -e
 
-resources_test_s3=s3://openproblems-data/resources_test/task_ist_preprocessing
+resources_test_s3="/scratch/task_ist_preprocessing/resources_test/task_ist_preprocessing/"
 # Results publish to /scratch -- created and written by the cloud compute env, so
 # the launcher does NOT create it here (it is read-only from the launch host).
 publish_dir="/scratch/results/runs/$(date +%Y-%m-%d_%H-%M-%S)_ssam"
diff --git a/scripts/run_benchmark/param_sweep/run_test_stardist_nebius.sh b/scripts/run_benchmark/param_sweep/run_test_stardist_nebius.sh
index 21d22c6bc..31d0d6765 100644
--- a/scripts/run_benchmark/param_sweep/run_test_stardist_nebius.sh
+++ b/scripts/run_benchmark/param_sweep/run_test_stardist_nebius.sh
@@ -27,7 +27,7 @@ cd "$REPO_ROOT"
 
 set -e
 
-resources_test_s3=s3://openproblems-data/resources_test/task_ist_preprocessing
+resources_test_s3="/scratch/task_ist_preprocessing/resources_test/task_ist_preprocessing/"
 # Results publish to /scratch — created and written by the cloud compute env, so
 # the launcher does NOT create it here (it is read-only from the launch host).
 publish_dir="/scratch/results/runs/$(date +%Y-%m-%d_%H-%M-%S)_stardist"
diff --git a/scripts/run_benchmark/param_sweep/run_test_tangram_nebius.sh b/scripts/run_benchmark/param_sweep/run_test_tangram_nebius.sh
index be7c40618..a9ddf4d95 100644
--- a/scripts/run_benchmark/param_sweep/run_test_tangram_nebius.sh
+++ b/scripts/run_benchmark/param_sweep/run_test_tangram_nebius.sh
@@ -30,7 +30,7 @@ cd "$REPO_ROOT"
 
 set -e
 
-resources_test_s3=s3://openproblems-data/resources_test/task_ist_preprocessing
+resources_test_s3="/scratch/task_ist_preprocessing/resources_test/task_ist_preprocessing/"
 # Results publish to /scratch — created and written by the cloud compute env, so
 # the launcher does NOT create it here (it is read-only from the launch host).
 publish_dir="/scratch/results/runs/$(date +%Y-%m-%d_%H-%M-%S)_tangram"
diff --git a/scripts/run_benchmark/param_sweep/run_test_watershed_nebius.sh b/scripts/run_benchmark/param_sweep/run_test_watershed_nebius.sh
index 657635223..4118a4e0c 100644
--- a/scripts/run_benchmark/param_sweep/run_test_watershed_nebius.sh
+++ b/scripts/run_benchmark/param_sweep/run_test_watershed_nebius.sh
@@ -27,7 +27,7 @@ cd "$REPO_ROOT"
 
 set -e
 
-resources_test_s3=s3://openproblems-data/resources_test/task_ist_preprocessing
+resources_test_s3="/scratch/task_ist_preprocessing/resources_test/task_ist_preprocessing/"
 # Results publish to /scratch — created and written by the cloud compute env, so
 # the launcher does NOT create it here (it is read-only from the launch host).
 publish_dir="/scratch/results/runs/$(date +%Y-%m-%d_%H-%M-%S)_watershed"
diff --git a/src/api/comp_data_preprocessor.yaml b/src/api/comp_data_preprocessor.yaml
index ddc535078..25b26c512 100644
--- a/src/api/comp_data_preprocessor.yaml
+++ b/src/api/comp_data_preprocessor.yaml
@@ -61,3 +61,14 @@ argument_groups:
         type: string
         description: The organism of the sample in the dataset.
         required: false
+  - name: Processing options
+    arguments:
+      - name: "--tissue_centered_crop"
+        type: boolean
+        default: false
+        description: |
+          When cropping an oversized image, centre the crop window on the transcript
+          density (median global x/y) instead of the image centre. Needed for huge
+          whole-section canvases where the tissue sits well off-centre (e.g. ABCA
+          whole-brain MERFISH), for which an image-centred window can miss the tissue
+          entirely. Default false keeps the historical image-centred crop.
diff --git a/src/data_processors/process_dataset/script.py b/src/data_processors/process_dataset/script.py
index 31b3ce1f8..b311e70db 100644
--- a/src/data_processors/process_dataset/script.py
+++ b/src/data_processors/process_dataset/script.py
@@ -31,7 +31,8 @@
   "dataset_reference": "https://www.10xgenomics.com/datasets/fresh-frozen-mouse-brain-replicates-1-standard;10.1038/s41586-023-06812-z",
   "dataset_summary": "Demonstration of gene expression profiling for fresh frozen mouse brain on the Xenium platform using the pre-designed Mouse Brain Gene Expression Panel (v1);A high-resolution scRNAseq atlas of cell types in the whole mouse brain",
   "dataset_description": "Demonstration of gene expression profiling for fresh frozen mouse brain on the Xenium platform using the pre-designed Mouse Brain Gene Expression Panel (v1). Replicate results demonstrate the high reproducibility of data generated by the platform. 10x Genomics obtained tissue from a C57BL/6 mouse from Charles River Laboratories. Three adjacent 10µm sections were placed on the same slide. Tissues were prepared following the demonstrated protocols Xenium In Situ for Fresh Frozen Tissues - Tissue Preparation Guide (CG000579) and Xenium In Situ for Fresh Frozen Tissues - Fixation & Permeabilization (CG000581).;See dataset_reference for more information. Note that we only took the 10xv2 data from the dataset.",
-  "dataset_organism": "mus_musculus"
+  "dataset_organism": "mus_musculus",
+  "tissue_centered_crop": False
 }
 ### VIASH END
 
@@ -50,7 +51,28 @@ def _full_res_image(img):
     return img["scale0"].image
 
 
-def get_crop_coords(sdata, max_n_pixels=20000*20000): #50000*50000):
+def _transcript_center_global(sdata):
+    """Median (x, y) of the transcripts in the GLOBAL coordinate system.
+
+    Used to centre the crop window on the tissue rather than on the image centre.
+    Global coords are derived from the element's affine transform (same math as
+    crop_points_by_global_xy), so it is correct for Scale/Affine/identity alike, and
+    the median is robust to sparse background points scattered at the canvas edges.
+    """
+    import dask
+    from spatialdata.transformations import get_transformation
+    pts = sdata['transcripts']
+    axes = [a for a in ("x", "y", "z") if a in pts.columns]
+    trans = get_transformation(pts, get_all=True)
+    M = trans["global"].to_affine_matrix(input_axes=tuple(axes), output_axes=tuple(axes))
+    ix, iy = axes.index("x"), axes.index("y")
+    gx = sum(M[ix, i] * pts[a] for i, a in enumerate(axes)) + M[ix, len(axes)]
+    gy = sum(M[iy, i] * pts[a] for i, a in enumerate(axes)) + M[iy, len(axes)]
+    cx, cy = dask.compute(gx.quantile(0.5), gy.quantile(0.5))
+    return float(cx), float(cy)
+
+
+def get_crop_coords(sdata, max_n_pixels=20000*20000, tissue_centered=False): #50000*50000):
     """Get the crop coordinates to subset the sdata to max_n_pixels
     
     Arguments
@@ -84,10 +106,24 @@ def get_crop_coords(sdata, max_n_pixels=20000*20000): #50000*50000):
         w_crop = w
         h_crop = int(max_n_pixels / w_crop)
         
-    # Center the crop
-    h_offset = (h - h_crop) // 2
-    w_offset = (w - w_crop) // 2
-        
+    # Position the crop window. By default it is centred on the IMAGE (historical
+    # behaviour). With tissue_centered=True it is centred on the TRANSCRIPT DENSITY
+    # instead — needed for huge whole-section canvases where the tissue sits well
+    # off-centre (e.g. ABCA whole-brain MERFISH, 83361x102170 px: an image-centred
+    # 20000px window caught 0 of 42M transcripts). The window is clamped inside the
+    # image; a degenerate/unavailable transcript centre falls back to the image centre.
+    if tissue_centered:
+        try:
+            cx, cy = _transcript_center_global(sdata)
+            if not (np.isfinite(cx) and np.isfinite(cy)):
+                cx, cy = w / 2, h / 2
+        except Exception:
+            cx, cy = w / 2, h / 2
+    else:
+        cx, cy = w / 2, h / 2
+    w_offset = int(min(max(cx - w_crop / 2, 0), max(w - w_crop, 0)))
+    h_offset = int(min(max(cy - h_crop / 2, 0), max(h - h_crop, 0)))
+
     crop = [[h_offset, h_offset + h_crop], [w_offset, w_offset + w_crop]]
 
     return crop
@@ -319,7 +355,7 @@ def subsample_adata_group_balanced(adata, group_key, n_samples, seed=0):
         _tbl.obs.index.name = None
 
 # Crop datasets that are too large
-crop_coords = get_crop_coords(sdata)
+crop_coords = get_crop_coords(sdata, tissue_centered=par.get("tissue_centered_crop", False))
 if crop_coords is not None:
     (y0, y1), (x0, x1) = crop_coords  # global box (raw image is at identity -> px == global)
 
diff --git a/src/methods_segmentation/stardist/NOTES.md b/src/methods_segmentation/stardist/NOTES.md
index 0121111b1..d41b06d23 100644
--- a/src/methods_segmentation/stardist/NOTES.md
+++ b/src/methods_segmentation/stardist/NOTES.md
@@ -46,22 +46,33 @@ detector**: the script feeds it `image[0]` only (see Tier 0 below).
 3. **Load model** (`:59`) — `StarDist2D.from_pretrained(par['model'])`. Default
    `2D_versatile_fluo`. This also loads the model's optimized `thresholds.json`
    (prob=0.479071, nms=0.3 for that model) as the fallback thresholds.
-4. **Percentile normalizer** (`:64-79`) — a csbdeep `Normalizer` subclass that min-max
+4. **Percentile normalizer** (`:64-77`) — a csbdeep `Normalizer` subclass that min-max
    scales the image to its **1st / 99.8th percentiles** (`normalize_mi_ma`), the
    recommended StarDist preprocessing but with fixed percentile bounds. `block_size`
-   and `offset` (block overlap/context) are derived from the image width so a large
-   panel is processed in tiles.
+   and `context` are derived from the image width so a large panel is processed in
+   tiles; **`min_overlap` is derived from object size, not block size** (see step 6 and
+   the min_overlap gotcha below).
 5. **Build eval-params** (`:85-89`) — collects the newly exposed tunables
    (`prob_thresh, nms_thresh, scale`) from `par`, **dropping any that are `None`**.
    A dropped key ⇒ `predict_instances` uses the model's own optimized value (so the
    no-args call is byte-for-byte the pre-tuning behaviour). Mirrors the cellposev4
    eval-params pattern.
-6. **Segment** (`:92-95`) — `model.predict_instances_big(image[0], axes='YX',
+6. **Segment** (`:92-131`) — `model.predict_instances_big(image[0], axes='YX',
    block_size=…, min_overlap=…, context=…, normalizer=…, **eval_params)`.
    `predict_instances_big` splits the image into `block_size` blocks, calls
    `predict_instances` on each (forwarding `**eval_params` unchanged — it only
    overrides `axes/overlap_label/return_labels/return_predict`), and reassembles the
    labels into global coordinates. `image[0]` = first channel → a single 2D plane.
+   **The stitching invariant is that every predicted object is smaller than
+   `min_overlap`** (an object bigger than the overlap can span a block seam and can't be
+   uniquely assigned → `RuntimeError: ...violates the assumption of being smaller than
+   'min_overlap'`). So `min_overlap` is set from an **object-size** bound
+   (`max_object_diameter`, default **192 px**), *not* from `block_size` (the old
+   `block_size // 5.5` shrank it to 64 px on small panels while real blobs reached
+   ~110 px → crash). `block_size` is then grown if needed to satisfy
+   `min_overlap + 2*context < block_size`, and the call is wrapped in a **retry that
+   doubles `min_overlap` on that specific error** so a rare oversized blob self-heals
+   instead of failing the run.
 7. **Post-process** (`:100-104`) — `convert_to_lower_dtype` downcasts the label array
    to the smallest uint that holds `max label`; wrap as an `xarray.DataArray`,
    `Labels2DModel.parse` with the copied transform, store as
@@ -109,6 +120,7 @@ Base image `openproblems/base_tensorflow_nvidia:1` (GPU-capable TF). The pip pin
 | `--prob_thresh` | double | *(unset → model 0.479071)* | `predict_instances(prob_thresh=)` |
 | `--nms_thresh` | double | *(unset → model 0.3)* | `predict_instances(nms_thresh=)` |
 | `--scale` | double | *(unset → no rescale)* | `predict_instances(scale=)` |
+| `--max_object_diameter` | integer | *(unset → 192 px)* | `predict_instances_big(min_overlap=)` |
 
 `prob_thresh`, `nms_thresh`, `scale` were **added during tuning** (see below). They are
 **optional with no static default on purpose**: leaving them unset makes StarDist use
diff --git a/src/methods_segmentation/stardist/config.vsh.yaml b/src/methods_segmentation/stardist/config.vsh.yaml
index cba9c2272..13e153995 100644
--- a/src/methods_segmentation/stardist/config.vsh.yaml
+++ b/src/methods_segmentation/stardist/config.vsh.yaml
@@ -47,6 +47,17 @@ arguments:
       If unset, no rescaling. Use it to match nucleus size to the model's training size:
       >1 upscales (small nuclei appear larger), <1 downscales. StarDist's analogue of
       Cellpose's diameter.
+  - name: --max_object_diameter
+    type: integer
+    required: false
+    description: >-
+      Upper bound (in original image pixels) on the largest object StarDist may produce.
+      It sets predict_instances_big's `min_overlap` (the guaranteed block-stitching
+      overlap), which MUST exceed the biggest nucleus/blob — otherwise stitching fails
+      with "violates the assumption of being smaller than 'min_overlap'". Defaults to
+      192 px if unset, and the script auto-doubles it on demand, so this is a
+      robustness/geometry knob, NOT a quality knob: it does not change which pixels get
+      segmented. Raise it if you hit that error or your nuclei are unusually large.
 
 resources:
   - type: python_script
diff --git a/src/methods_segmentation/stardist/script.py b/src/methods_segmentation/stardist/script.py
index 4bf502c61..4014f2772 100644
--- a/src/methods_segmentation/stardist/script.py
+++ b/src/methods_segmentation/stardist/script.py
@@ -43,6 +43,7 @@ def convert_to_lower_dtype(arr):
   "prob_thresh": None,
   "nms_thresh": None,
   "scale": None,
+  "max_object_diameter": None,
 }
 
 ## VIASH END
@@ -75,8 +76,6 @@ def do_after(self):
 
 mi, ma = np.percentile(image, [1,99.8])
 normalizer = MyNormalizer(mi, ma)
-block_size = min(image.shape[1] // 3, 4096)
-offset = min(block_size // 5.5, 128)
 
 # Tunable knobs forwarded through predict_instances_big -> predict_instances.
 # A value left as None (i.e. omitted from par) means "use the model's own optimized
@@ -89,10 +88,44 @@ def do_after(self):
 }
 print(f"predict_instances_big overrides: {eval_params}", flush=True)
 
-labels, _ = model.predict_instances_big(
-    image[0,:,:], axes='YX', block_size=block_size, min_overlap=offset,
-    context=offset, normalizer=normalizer, **eval_params  # n_tiles left to block_size
-)
+# predict_instances_big tiles the image and stitches the per-block predictions. Its
+# stitching invariant is that EVERY predicted object is smaller than `min_overlap`;
+# an object spanning a block seam that is larger than the overlap can't be assigned to
+# a single block, which raises "Found object of shape (...), which violates the
+# assumption of being smaller than 'min_overlap'". So `min_overlap` must be
+# OBJECT-SIZE-based, not block-size-based — the old `block_size // 5.5` shrank the
+# overlap below real nuclei/blobs on small panels (min_overlap fell to 64 px while
+# objects reached ~110 px). Objects are measured in ORIGINAL image pixels
+# (predict_instances undoes `scale` internally), so this bound is in original px and is
+# independent of `scale`. `context` is only the receptive-field margin discarded around
+# each block, so deriving it from the image size is fine.
+block_size = min(image.shape[1] // 3, 4096)
+context = int(min(block_size // 5.5, 128))
+min_overlap = int(par.get("max_object_diameter") or 192)  # px; must exceed largest object
+# predict_instances_big asserts: min_overlap + 2*context < block_size.
+block_size = max(block_size, min_overlap + 2 * context + 1)
+
+# Self-heal: if a rare oversized blob (merged nuclei / debris) still exceeds
+# `min_overlap`, double it (and grow block_size to keep the geometry constraint) and
+# retry, rather than failing the whole segmentation.
+while True:
+    try:
+        labels, _ = model.predict_instances_big(
+            image[0, :, :], axes='YX', block_size=block_size,
+            min_overlap=min_overlap, context=context,
+            normalizer=normalizer, **eval_params,  # n_tiles left to block_size
+        )
+        break
+    except RuntimeError as e:
+        if "min_overlap" not in str(e) or min_overlap >= 2048:
+            raise
+        min_overlap *= 2
+        block_size = max(block_size, min_overlap + 2 * context + 1)
+        print(
+            "predict_instances_big: an object exceeded min_overlap; retrying with "
+            f"min_overlap={min_overlap}, block_size={block_size}",
+            flush=True,
+        )
 
 
 
diff --git a/src/methods_transcript_assignment/segger/NOTES.md b/src/methods_transcript_assignment/segger/NOTES.md
index 0ce302627..f3a1e138e 100644
--- a/src/methods_transcript_assignment/segger/NOTES.md
+++ b/src/methods_transcript_assignment/segger/NOTES.md
@@ -218,6 +218,24 @@ Nextflow runner labels: `[hightime, midcpu, highmem, gpuh100]`.
 
 ## Risk points / gotchas
 
+- **Never start a continuation line in `script.py` with `|` — viash's Nextflow codegen
+  deletes it (Groovy `stripMargin`).** viash embeds `script.py` into the Nextflow module
+  (`target/nextflow/.../segger/main.nf`) as a Groovy string and runs it through
+  `.stripMargin()`, whose **default margin delimiter is `|`**: any line matching
+  `^\s*\|` has the leading whitespace **and the `|`** stripped. A wrapped boolean/bitwise
+  expression like
+  `n_oob = int(np.count_nonzero(\n    (y_coords < 0) | (y_coords >= H)\n    | (x_coords < 0) | (x_coords >= W)\n))`
+  becomes `... (y_coords >= H)\n (x_coords < 0) ...` in the generated module — the two
+  parenthesised groups then **juxtapose into a call**, so at run time Python raises
+  `TypeError: 'numpy.ndarray' object is not callable` (the caret points at the *first*
+  group, the "callable"). **This corrupts ONLY the `=nextflow=>` target**, so `viash test` /
+  `viash run` (the `=executable=>` target, no `stripMargin`) pass clean and the bug is
+  invisible until a Nextflow/Tower benchmark run. It also survives across builds: the
+  `build/main` executable can be correct while the `build/main` *nextflow module* is
+  corrupt. **Fix = put the operator at the END of the line** (`... (y_coords >= H) |` /
+  next line `(x_coords < 0) ...`) so no line begins with `|`; or keep it on one line.
+  Grep guard before committing: `grep -nE '^[[:space:]]*\|' script.py` must be empty.
+  (Diagnosed 2026-08-06: reproduced byte-for-byte, incl. the exact caret placement.)
 - **numba must use the NVIDIA (cuda-python) driver binding, or cudf/numba see 0 GPUs.**
   The load-bearing GPU fix: `run_segger` sets **`NUMBA_CUDA_USE_NVIDIA_BINDING=1`** for the
   subprocess. RAPIDS (cudf/cuml/cugraph) builds its CUDA context through the cuda-python
diff --git a/src/methods_transcript_assignment/segger/script.py b/src/methods_transcript_assignment/segger/script.py
index 71e1eb29d..69c0de530 100644
--- a/src/methods_transcript_assignment/segger/script.py
+++ b/src/methods_transcript_assignment/segger/script.py
@@ -173,9 +173,16 @@
 # rounding is all that remains, and it's handled below. NOTE: this does NOT touch the separate
 # empty-`bd`-batch crash (a cell-free INTERIOR tile), which is intra-field, not OOB-driven, and
 # still needs the segger-encoder guard -- see NOTES.md.
+# NB: keep the `|` operators at the END of each line -- a continuation line must NOT start
+# with `|`. viash embeds this script into the Nextflow module as a Groovy string and runs it
+# through `.stripMargin()`, which treats a leading `|` as the margin delimiter and DELETES it.
+# A leading-`|` line like `| (x_coords < 0) ...` becomes ` (x_coords < 0) ...`, so the previous
+# line's `(...)` and this line's `(...)` juxtapose into a call -> the Nextflow run dies with
+# `TypeError: 'numpy.ndarray' object is not callable` while `viash test`/`viash run` (the
+# executable target, no stripMargin) pass. Trailing operators sidestep it entirely.
 n_oob = int(np.count_nonzero(
-    (y_coords < 0) | (y_coords >= label_image.shape[0])
-    | (x_coords < 0) | (x_coords >= label_image.shape[1])
+    (y_coords < 0) | (y_coords >= label_image.shape[0]) |
+    (x_coords < 0) | (x_coords >= label_image.shape[1])
 ))
 y_coords = np.clip(y_coords, 0, label_image.shape[0] - 1)
 x_coords = np.clip(x_coords, 0, label_image.shape[1] - 1)
diff --git a/src/workflows/process_datasets/main.nf b/src/workflows/process_datasets/main.nf
index 2bfc4c3f9..70948ecf8 100644
--- a/src/workflows/process_datasets/main.nf
+++ b/src/workflows/process_datasets/main.nf
@@ -22,7 +22,8 @@ workflow run_wf {
         "dataset_reference",
         "dataset_summary",
         "dataset_description",
-        "dataset_organism"
+        "dataset_organism",
+        "tissue_centered_crop"
       ],
       toState: [
         "output_sc",