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/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)