diff --git a/.github/workflows/submit_rail_knn_4tasks.yaml b/.github/workflows/submit_rail_knn_4tasks.yaml
new file mode 100644
index 0000000..b99a399
--- /dev/null
+++ b/.github/workflows/submit_rail_knn_4tasks.yaml
@@ -0,0 +1,38 @@
+---
+# This workflow will install Python dependencies and run tests
+
+name: Unit test and code coverage
+
+on:
+  push:
+    branches: [main]
+  pull_request:
+    branches: [main]
+
+jobs:
+  build:
+
+    runs-on: ubuntu-latest
+    strategy:
+      matrix:
+        python-version: ['3.13']
+        submission: ['rail_knn_4tasks']
+
+    steps:
+      - uses: actions/checkout@v3
+      - name: Set up Python ${{ matrix.python-version }}
+        uses: actions/setup-python@v4
+        with:
+          python-version: ${{ matrix.python-version }}
+      - name: Install dependencies
+        run: |
+          sudo apt-get update
+          sudo apt install libbz2-dev
+          python -m pip install --upgrade pip
+          pip install wheel
+          pip install .
+          pip install .[dev]
+          if [ -f requirements_${{ matrix.submission }}.txt ]; then pip install -r requirements_${{ matrix.submission }}.txt; fi
+      - name: Run unit tests with pytest
+        run: |
+          python -m pytest tests/test_${{ matrix.submission }}.py
diff --git a/requirements_rail_knn_4tasks.txt b/requirements_rail_knn_4tasks.txt
new file mode 100644
index 0000000..f649f3b
--- /dev/null
+++ b/requirements_rail_knn_4tasks.txt
@@ -0,0 +1,4 @@
+pz-rail-base>=2.0.0
+pz-rail-sklearn>=2.0.0
+qp-prob>=1.0.0
+tables_io>=1.0.0
diff --git a/scripts/build_rail_knn_4tasks_submission.py b/scripts/build_rail_knn_4tasks_submission.py
new file mode 100644
index 0000000..3ad0cf7
--- /dev/null
+++ b/scripts/build_rail_knn_4tasks_submission.py
@@ -0,0 +1,109 @@
+#!/usr/bin/env python3
+"""Build precomputed rail_knn_4tasks submission tarball."""
+import math
+import os
+import tarfile
+
+import numpy as np
+import tables_io
+from rail.core.data import TableHandle
+from rail.estimation.algos.k_nearneigh import KNearNeighEstimator, KNearNeighInformer
+from rail.utils import catalog_utils
+
+PUBLIC_AREA = os.environ.get("PUBLIC_AREA", "tests/public")
+OUT_DIR = os.environ.get("OUT_DIR", "build_submission/rail_knn_4tasks")
+TASKSETS = [1, 2, 3, 4]
+SIMS = ["cardinal", "flagship"]
+SCENARIOS = ["1yr", "10yr"]
+_CHUNK = 150_000
+_ZMAX = 3.0
+_NZ = 151
+
+
+def clean_training_file(train_file: str) -> str:
+    data = tables_io.read(train_file)
+    bad_mask = np.isnan(data["redshift"])
+    if not bad_mask.any():
+        return train_file
+    cleaned_path = train_file.replace(".hdf5", "_cleaned.hdf5")
+    cleaned_data = {key: val[~bad_mask] for key, val in data.items()}
+    tables_io.write(cleaned_data, cleaned_path)
+    return cleaned_path
+
+
+def make_informer() -> KNearNeighInformer:
+    return KNearNeighInformer.make_stage(
+        name="inform",
+        hdf5_groupname="",
+        zmax=_ZMAX,
+        nzbins=_NZ,
+        chunk_size=_CHUNK,
+        nondetect_val=math.nan,
+        trainfrac=0.2,
+        nneigh_min=3,
+        nneigh_max=5,
+        ngrid_sigma=6,
+    )
+
+
+def make_estimator(model) -> KNearNeighEstimator:
+    return KNearNeighEstimator.make_stage(
+        name="estimate",
+        model=model,
+        hdf5_groupname="",
+        output_mode="return",
+        nzbins=_NZ,
+        zmax=_ZMAX,
+        chunk_size=_CHUNK,
+        nondetect_val=math.nan,
+    )
+
+
+def main() -> None:
+    os.makedirs(OUT_DIR, exist_ok=True)
+    catalog_utils.clear()
+    catalog_utils.load_yaml("tests/catalogs.yaml")
+    catalog_utils.apply("cardinal_roman_rubin")
+
+    for taskset in TASKSETS:
+        for sim in SIMS:
+            for scenario in SCENARIOS:
+                train_file = (
+                    f"{PUBLIC_AREA}/pz_challenge_taskset_{taskset}_{sim}_"
+                    f"training_{scenario}.hdf5"
+                )
+                test_file = (
+                    f"{PUBLIC_AREA}/pz_challenge_taskset_{taskset}_{sim}_"
+                    f"test_{scenario}.hdf5"
+                )
+                model_path = (
+                    f"{OUT_DIR}/pz_challenge_taskset_{taskset}_{sim}_"
+                    f"pz_model_{scenario}.pkl"
+                )
+                estimate_path = (
+                    f"{OUT_DIR}/pz_challenge_taskset_{taskset}_{sim}_"
+                    f"pz_estimate_{scenario}.hdf5"
+                )
+
+                cleaned_train_file = clean_training_file(train_file)
+                train_data = TableHandle("train", path=cleaned_train_file)
+                test_data = TableHandle("test", path=test_file)
+
+                model = make_informer().inform(train_data)
+                model.path = model_path
+                model.write()
+
+                pz_out = make_estimator(model).estimate(test_data)
+                pz_out.data.ancil["object_id"] = np.asarray(test_data()["object_id"])
+                pz_out.path = estimate_path
+                pz_out.write()
+                print(f"Wrote {model_path} and {estimate_path}")
+
+    tarball = "rail_knn_4tasks_submission.tgz"
+    with tarfile.open(tarball, "w:gz") as tar:
+        tar.add(OUT_DIR, arcname="rail_knn_4tasks")
+    print(f"Created {tarball}")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/tests/test_rail_knn_4tasks.py b/tests/test_rail_knn_4tasks.py
new file mode 100644
index 0000000..6aa23a3
--- /dev/null
+++ b/tests/test_rail_knn_4tasks.py
@@ -0,0 +1,260 @@
+import math
+import os
+from pathlib import Path
+
+import numpy as np
+import pytest
+import tables_io
+from rail.core.data import TableHandle
+from rail.estimation.algos.k_nearneigh import KNearNeighEstimator, KNearNeighInformer
+from rail.utils import catalog_utils
+
+from pz_data_challenge import submit_utils
+from pz_data_challenge.taskset_1 import run_taskset_1
+from pz_data_challenge.taskset_2 import run_taskset_2
+from pz_data_challenge.taskset_3 import run_taskset_3
+from pz_data_challenge.taskset_4 import run_taskset_4
+
+SUBMISSION_NAME: str = "rail_knn_4tasks"
+SUBMISSION_URL: str = (
+    "https://github.com/Lhior/pz_data_challenge/releases/download/"
+    "pzdc-rail-knn-4tasks-v1/rail_knn_4tasks_submission.tgz"
+)
+
+SUBMIT_DIR: str = f"submissions/{SUBMISSION_NAME}"
+PUBLIC_AREA: str = "tests/public"
+
+CATALOG_TAG = "cardinal_roman_rubin"
+_CHUNK = 150_000
+_ZMAX = 3.0
+_NZ = 151
+
+
+def _attach_object_ids(pz_out, test_data: TableHandle) -> None:
+    pz_out.data.ancil["object_id"] = np.asarray(test_data()["object_id"])
+
+
+def _clean_training_file(train_file: str | Path) -> str:
+    path = str(train_file)
+    data = tables_io.read(path)
+    bad_mask = np.isnan(data["redshift"])
+    if not bad_mask.any():
+        return path
+    cleaned_path = path.replace(".hdf5", "_cleaned.hdf5")
+    cleaned_data = {key: val[~bad_mask] for key, val in data.items()}
+    tables_io.write(cleaned_data, cleaned_path)
+    return cleaned_path
+
+
+def _make_knn_informer() -> KNearNeighInformer:
+    return KNearNeighInformer.make_stage(
+        name="inform",
+        hdf5_groupname="",
+        zmax=_ZMAX,
+        nzbins=_NZ,
+        chunk_size=_CHUNK,
+        nondetect_val=math.nan,
+        trainfrac=0.2,
+        nneigh_min=3,
+        nneigh_max=5,
+        ngrid_sigma=6,
+    )
+
+
+def _make_knn_estimator(model) -> KNearNeighEstimator:
+    return KNearNeighEstimator.make_stage(
+        name="estimate",
+        model=model,
+        hdf5_groupname="",
+        output_mode="return",
+        nzbins=_NZ,
+        zmax=_ZMAX,
+        chunk_size=_CHUNK,
+        nondetect_val=math.nan,
+    )
+
+
+@pytest.fixture(name="setup_submit_area", scope="module")
+def setup_submit_area(request: pytest.FixtureRequest) -> int:
+    if not os.path.exists(SUBMIT_DIR):
+        if not SUBMISSION_URL:
+            raise ValueError(
+                f"SUBMISSION_URL in tests/test_{SUBMISSION_NAME}.py has not been set"
+            )
+        submit_utils.download_and_extract_tar(SUBMISSION_URL, "submissions")
+
+    def teardown_submit_area() -> None:
+        if not os.environ.get("NO_TEARDOWN"):
+            os.system(f"\\rm -rf {SUBMIT_DIR}")
+
+    try:
+        os.makedirs(os.path.join(SUBMIT_DIR, "outputs_2"))
+    except Exception:
+        pass
+
+    try:
+        os.makedirs(os.path.join(SUBMIT_DIR, "outputs_3"))
+    except Exception:
+        pass
+
+    request.addfinalizer(teardown_submit_area)
+
+    catalog_utils.clear()
+    catalog_utils.load_yaml("tests/catalogs.yaml")
+    catalog_utils.apply(CATALOG_TAG)
+
+    return 0
+
+
+def run_taskset_x_estimation_only(
+    model_file: str | Path,
+    test_file: str | Path,
+    output_file: str | Path,
+) -> None:
+    test_data = TableHandle("test", path=str(test_file))
+    estimator = _make_knn_estimator(str(model_file))
+    pz_out = estimator.estimate(test_data)
+    _attach_object_ids(pz_out, test_data)
+    pz_out.path = output_file
+    pz_out.write()
+
+
+def run_taskset_x_training_and_estimation(
+    train_file: str | Path,
+    test_file: str | Path,
+    output_file: str | Path,
+) -> None:
+    cleaned_train_file = _clean_training_file(train_file)
+    train_data = TableHandle("train", path=cleaned_train_file)
+    test_data = TableHandle("test", path=str(test_file))
+    informer = _make_knn_informer()
+    model = informer.inform(train_data)
+    estimator = _make_knn_estimator(model)
+    pz_out = estimator.estimate(test_data)
+    _attach_object_ids(pz_out, test_data)
+    pz_out.path = output_file
+    pz_out.write()
+
+
+def run_taskset_1_estimation_only(
+    model_file: str | Path,
+    test_file: str | Path,
+    output_file: str | Path,
+) -> None:
+    run_taskset_x_estimation_only(model_file, test_file, output_file)
+
+
+def run_taskset_1_training_and_estimation(
+    train_file: str | Path,
+    test_file: str | Path,
+    output_file: str | Path,
+) -> None:
+    run_taskset_x_training_and_estimation(train_file, test_file, output_file)
+
+
+def run_taskset_2_estimation_only(
+    model_file: str | Path,
+    test_file: str | Path,
+    output_file: str | Path,
+) -> None:
+    run_taskset_x_estimation_only(model_file, test_file, output_file)
+
+
+def run_taskset_2_training_and_estimation(
+    train_file: str | Path,
+    test_file: str | Path,
+    output_file: str | Path,
+) -> None:
+    run_taskset_x_training_and_estimation(train_file, test_file, output_file)
+
+
+def run_taskset_3_estimation_only(
+    model_file: str | Path,
+    test_file: str | Path,
+    output_file: str | Path,
+) -> None:
+    run_taskset_x_estimation_only(model_file, test_file, output_file)
+
+
+def run_taskset_3_training_and_estimation(
+    train_file: str | Path,
+    test_file: str | Path,
+    output_file: str | Path,
+) -> None:
+    run_taskset_x_training_and_estimation(train_file, test_file, output_file)
+
+
+def run_taskset_4_estimation_only(
+    model_file: str | Path,
+    test_file: str | Path,
+    output_file: str | Path,
+) -> None:
+    run_taskset_x_estimation_only(model_file, test_file, output_file)
+
+
+def run_taskset_4_training_and_estimation(
+    train_file: str | Path,
+    test_file: str | Path,
+    output_file: str | Path,
+) -> None:
+    run_taskset_x_training_and_estimation(train_file, test_file, output_file)
+
+
+def test_example_taskset_1(
+    setup_public_area: int,
+    setup_submit_area: int,
+) -> None:
+    assert setup_public_area == 0
+    assert setup_submit_area == 0
+
+    run_taskset_1(
+        PUBLIC_AREA,
+        SUBMISSION_NAME,
+        run_taskset_1_estimation_only,
+        run_taskset_1_training_and_estimation,
+    )
+
+
+def test_example_taskset_2(
+    setup_public_area: int,
+    setup_submit_area: int,
+) -> None:
+    assert setup_public_area == 0
+    assert setup_submit_area == 0
+
+    run_taskset_2(
+        PUBLIC_AREA,
+        SUBMISSION_NAME,
+        run_taskset_2_estimation_only,
+        run_taskset_2_training_and_estimation,
+    )
+
+
+def test_example_taskset_3(
+    setup_public_area: int,
+    setup_submit_area: int,
+) -> None:
+    assert setup_public_area == 0
+    assert setup_submit_area == 0
+
+    run_taskset_3(
+        PUBLIC_AREA,
+        SUBMISSION_NAME,
+        run_taskset_3_estimation_only,
+        run_taskset_3_training_and_estimation,
+    )
+
+
+def test_example_taskset_4(
+    setup_public_area: int,
+    setup_submit_area: int,
+) -> None:
+    assert setup_public_area == 0
+    assert setup_submit_area == 0
+
+    run_taskset_4(
+        PUBLIC_AREA,
+        SUBMISSION_NAME,
+        run_taskset_4_estimation_only,
+        run_taskset_4_training_and_estimation,
+    )