diff --git a/CHANGELOG.md b/CHANGELOG.md
index 10ec97c6..2c3c64c0 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -1,5 +1,9 @@
 # task_predict_modality 0.2.0
 
+## NEW FUNCTIONALITY
+
+* Added `src/utils/exit_codes.py`, so components can mark themselves non-applicable for a dataset (PR #32).
+
 ## MINOR CHANGES
 
 * `file_train_mod1`, `file_train_mod2`, `file_test_mod1`, `file_test_mod2`: Declare `uns["modality"]`, which `process_dataset` already writes and which six methods and the `run_benchmark` workflow already read (PR #29).
@@ -34,6 +38,10 @@
 
 ## BUG FIXES
 
+* `guanlab_dengkw_pm`: Restore the consensus scheme of the original submission -- five reshuffles of the batches into two halves, ten kernel ridge models averaged. The port had replaced it with a single fixed two-way split for ADT pairs and leave-one-batch-out otherwise, so the result depended on the order the batches happened to come in. `--n_repeats` and `--seed` are now arguments; the unused `--distance_method` and `--n_pcs` are gone (PR #32).
+
+* `guanlab_dengkw_pm`: Only map the predictions back through the mod2 SVD when that SVD was actually fitted. When the target had fewer features than `n_mod2`, the method raised `NameError: embedder_mod2`. Unsupported modality pairs now exit 99 (non-applicable) instead of raising a `KeyError` (PR #32).
+
 * `process_dataset`: Fall back to holding out a quarter of the batches when the dataset has no `obs["is_train"]`, rather than silently producing four empty h5ads. `obs["is_train"]` carries the NeurIPS 2021 competition split and stays optional; `obs["cell_type"]` is now declared and required (PR #28).
 
 * Fix the component paths, build paths and `rename_keys` separator in the helper scripts, which prevented `scripts/create_datasets/test_resources.sh` and both `run_test.sh` scripts from running at all (PR #22).
diff --git a/src/methods/guanlab_dengkw_pm/config.vsh.yaml b/src/methods/guanlab_dengkw_pm/config.vsh.yaml
index 925b5591..fb4b9c20 100644
--- a/src/methods/guanlab_dengkw_pm/config.vsh.yaml
+++ b/src/methods/guanlab_dengkw_pm/config.vsh.yaml
@@ -18,18 +18,22 @@ info:
   preferred_normalization: log_cp10k
   competition_submission_id: 170636
 arguments:
-  - name: "--distance_method"
-    type: "string"
-    default: "minkowski"
-    description: The distance metric to use. Possible values include `euclidean` and `minkowski`.
-    choices: [euclidean, minkowski]
-  - name: "--n_pcs"
+  - name: "--n_repeats"
     type: "integer"
-    default: 50
-    description: Number of components to use for dimensionality reduction.
+    default: 5
+    description: |
+      Number of times the batches are reshuffled into two halves. Each half fits one
+      kernel ridge model, so the consensus averages `2 * n_repeats` predictions.
+    info:
+      test_default: 1
+  - name: "--seed"
+    type: "integer"
+    default: 1000
+    description: Seed for the batch reshuffling.
 resources:
   - type: python_script
     path: script.py
+  - path: /src/utils/exit_codes.py
 engines:
   - type: docker
     image: openproblems/base_python:1
diff --git a/src/methods/guanlab_dengkw_pm/script.py b/src/methods/guanlab_dengkw_pm/script.py
index d11b7355..3a0fd029 100644
--- a/src/methods/guanlab_dengkw_pm/script.py
+++ b/src/methods/guanlab_dengkw_pm/script.py
@@ -1,3 +1,4 @@
+import sys
 import anndata as ad
 import numpy as np
 from scipy.sparse import csc_matrix
@@ -10,15 +11,19 @@
     'input_train_mod1': 'resources_test/task_predict_modality/openproblems_neurips2021/bmmc_multiome/normal/train_mod1.h5ad',
     'input_train_mod2': 'resources_test/task_predict_modality/openproblems_neurips2021/bmmc_multiome/normal/train_mod2.h5ad',
     'input_test_mod1': 'resources_test/task_predict_modality/openproblems_neurips2021/bmmc_multiome/normal/test_mod1.h5ad',
-    'output': 'output.h5ad', 
-    'distance_method': 'minkowski', 
-    'n_pcs': 50
+    'output': 'output.h5ad',
+    'n_repeats': 5,
+    'seed': 1000
 }
 meta = {
-    'name': 'guanlab_dengkw_pm'
+    'name': 'guanlab_dengkw_pm',
+    'resources_dir': 'src/utils'
 }
 ## VIASH END
 
+sys.path.append(meta['resources_dir'])
+from exit_codes import exit_non_applicable
+
 
 ## Removed PCA and normalization steps, as they arr already performed with the input data
 print('Reading input files', flush=True)
@@ -49,6 +54,11 @@
     ("ATAC", "GEX"): (100, 70, 10, 0.1)
 }
 print(f"{mod1_type}, {mod2_type}", flush=True)
+if (mod1_type, mod2_type) not in n_comp_dict:
+    exit_non_applicable(
+        f"Guanlab-dengkw was tuned on the NeurIPS 2021 modality pairs and has no "
+        f"hyperparameters for {mod1_type} -> {mod2_type}"
+    )
 n_mod1, n_mod2, scale, alpha = n_comp_dict[(mod1_type, mod2_type)]
 print(f"{n_mod1}, {n_mod2}, {scale}, {alpha}", flush=True)
 
@@ -68,6 +78,8 @@
     embedder_mod2 = TruncatedSVD(n_components=n_mod2)
     train_gs = embedder_mod2.fit_transform(input_train_mod2.layers["normalized"]).astype(np.float32)
 else:
+    # no dimensionality reduction, so the KRR predicts the mod2 features directly
+    embedder_mod2 = None
     train_gs = input_train_mod2.to_df(layer="normalized").values.astype(np.float32)
 
 del input_train
@@ -86,36 +98,36 @@
 del test_matrix
 
 print('Running KRR model ...', flush=True)
-if batch_len == 1:
-    # just in case there is only one batch
-    batch_subsets = [batches]
-elif mod1_type == "ADT" or mod2_type == "ADT":
-    # two fold consensus predictions
-    batch_subsets = [
-        batches[:batch_len//2],
-        batches[batch_len//2:]
-    ]
-else:
-    # leave-one-batch-out consensus predictions
-    batch_subsets = [
-        batches[:i] + batches[i+1:]
-        for i in range(batch_len)
-    ]
-
+# consensus over n_repeats reshuffles of the batches into two halves, as in the
+# original submission
 y_pred = np.zeros((input_test_mod1.n_obs, input_train_mod2.n_vars), dtype=np.float32)
-for batch in batch_subsets:
-    print(batch, flush=True)
-    kernel = RBF(length_scale = scale)
-    krr = KernelRidge(alpha=alpha, kernel=kernel)
-    print('Fitting KRR ... ', flush=True)
-    krr.fit(
-        train_norm[input_train_mod1.obs.batch.isin(batch)], 
-        train_gs[input_train_mod2.obs.batch.isin(batch)]
-    )
-    y_pred += (krr.predict(test_norm) @ embedder_mod2.components_)
+np.random.seed(par['seed'])
+n_models = 0
+
+for _ in range(par['n_repeats']):
+    np.random.shuffle(batches)
+    for batch in [batches[:batch_len//2], batches[batch_len//2:]]:
+        # with a single batch one half comes out empty
+        if not batch:
+            batch = [batches[0]]
+
+        print(batch, flush=True)
+        kernel = RBF(length_scale = scale)
+        krr = KernelRidge(alpha=alpha, kernel=kernel)
+        print('Fitting KRR ... ', flush=True)
+        krr.fit(
+            train_norm[input_train_mod1.obs.batch.isin(batch)],
+            train_gs[input_train_mod2.obs.batch.isin(batch)]
+        )
+        y_pred_batch = krr.predict(test_norm)
+        if embedder_mod2 is not None:
+            # map the predicted components back to the mod2 feature space
+            y_pred_batch = y_pred_batch @ embedder_mod2.components_
+        y_pred += y_pred_batch
+        n_models += 1
 
 np.clip(y_pred, a_min=0, a_max=None, out=y_pred)
-y_pred /= len(batch_subsets)
+y_pred /= n_models
 
 # Store as sparse matrix to be efficient. 
 # Note that this might require different classifiers/embedders before-hand. 
diff --git a/src/utils/exit_codes.py b/src/utils/exit_codes.py
new file mode 100644
index 00000000..03f270c3
--- /dev/null
+++ b/src/utils/exit_codes.py
@@ -0,0 +1,6 @@
+import sys
+
+
+def exit_non_applicable(msg):
+    print(f"NON-APPLICABLE ERROR: {msg}", flush=True)
+    sys.exit(99)