diff --git a/.github/workflows/submit_conclave.yaml b/.github/workflows/submit_conclave.yaml
new file mode 100644
index 0000000..a8a2010
--- /dev/null
+++ b/.github/workflows/submit_conclave.yaml
@@ -0,0 +1,43 @@
+---
+# 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: ['conclave']
+
+    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_${{ matrix.submission }}.txt; fi
+      - name: Run unit tests with pytest
+        # conclave trains 3 estimators; a full-scale retrain (100k x4) exceeds the CI budget.
+        # Subsample the train path for CI (validity smoke); estimation runs on the full test set,
+        # and the shipped full-scale models/estimates (release tgz) carry the real performance.
+        env:
+          PZDC_CI_MAX_TRAIN: "4000"
+        run: |
+          python -m pytest tests/test_${{ matrix.submission }}.py
diff --git a/requirements_conclave.txt b/requirements_conclave.txt
new file mode 100644
index 0000000..bf9df1c
--- /dev/null
+++ b/requirements_conclave.txt
@@ -0,0 +1,6 @@
+# conclave TS1 submission requirements.
+# The method lives in the pip-installable `conclave` package; installing it pulls the
+# RAIL estimator stack (pz-rail-{base,flexzboost,pzflow,gpz-v1}), qp-prob, tables_io,
+# numpy, scipy transitively via conclave's own pyproject dependencies.
+# Pin to the tagged release that matches the shipped pretrained models / estimates.
+conclave @ git+https://github.com/rhw/conclave@ts1-v1
diff --git a/tests/test_conclave.py b/tests/test_conclave.py
new file mode 100644
index 0000000..292689f
--- /dev/null
+++ b/tests/test_conclave.py
@@ -0,0 +1,117 @@
+"""CI test for the `conclave` TS1 submission (LSST-DESC PZ Data Challenge).
+
+conclave = a committee ensemble of PZFlow + GPz + FlexZBoost combined with convex-QP
+optimal weights and a global-PIT recalibration. The method lives in the pip-installable
+`conclave` package (github.com/rhw/conclave, pinned in requirements_conclave.txt); this
+test file is the thin submission entry point that the upstream harness discovers and runs.
+
+Structure mirrors the accepted reference submissions (e.g. test_rail_knn_test.py): it defines
+the four run_taskset_* entry points and calls run_taskset_1/2 over SIMS x SCENARIOS. The two
+functions delegate to conclave.submission, which internally applies the estimators' catalog tag
+and freezes the QP weights + recalibrator (the blind test has no redshift, so recal is fit on a
+held-out slice of the training file). Pre-made estimates + pretrained models are hosted at
+SUBMISSION_URL and unpacked into submissions/conclave/.
+"""
+import os
+from pathlib import Path
+
+import numpy as np
+import pytest
+from rail.core.data import TableHandle
+from rail.utils import catalog_utils
+
+from pz_data_challenge import submit_utils
+from pz_data_challenge.taskset_1 import run_taskset_1
+
+from conclave.submission import (
+    run_taskset_1_training_and_estimation as _conclave_train_and_estimate,
+    run_taskset_1_estimation_only as _conclave_estimate_only,
+)
+
+# CI concession: the full method trains 3 estimators (PZFlow/GPz/FlexZBoost) on 100k x 4
+# sim/scenario — hours, over the GitHub-runner budget. For the train+estimate CI path we train
+# on a subsample so CI proves the pipeline runs and emits valid p(z); the REAL full-scale-trained
+# models + estimates ship in the release tarball (SUBMISSION_URL) and feed the estimation-only
+# path. PZDC_CI_MAX_TRAIN=0 (default) trains on the full data (the real submission behaviour).
+CI_MAX_TRAIN: int = int(os.environ.get("PZDC_CI_MAX_TRAIN", "0"))
+
+SUBMISSION_NAME: str = "conclave"
+# GitHub release .tgz of the pre-made estimates + pretrained models (set before opening the PR).
+SUBMISSION_URL: str = "https://github.com/rhw/conclave/releases/download/ts1-v1/conclave_submission.tgz"
+
+SUBMIT_DIR: str = f"submissions/{SUBMISSION_NAME}"
+PUBLIC_AREA: str = os.environ.get("PZDC_PUBLIC_AREA", "tests/public")
+
+# Official catalog tag for reading the challenge catalogs. conclave's estimators additionally
+# apply their own (equivalent LSST+Roman) tag internally via conclave.bands.apply_band_set.
+CATALOG_TAG = "cardinal_roman_rubin"
+
+
+@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, SUBMIT_DIR)
+
+    def teardown_submit_area() -> None:
+        if not os.environ.get("NO_TEARDOWN"):
+            os.system(f"\\rm -rf {SUBMIT_DIR}/outputs_2 {SUBMIT_DIR}/outputs_3")
+
+    for sub in ("outputs_2", "outputs_3"):
+        try:
+            os.makedirs(os.path.join(SUBMIT_DIR, sub))
+        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 _maybe_subsample_train(train_file: str) -> str:
+    """Return train_file unchanged unless PZDC_CI_MAX_TRAIN>0, in which case write a
+    stratified-by-nothing random subsample to a temp hdf5 (keeps CI train fast)."""
+    if CI_MAX_TRAIN <= 0:
+        return train_file
+    import tempfile
+    import tables_io
+    d = tables_io.read(train_file)
+    keys = list(d.keys())
+    n = len(d[keys[0]])
+    if n <= CI_MAX_TRAIN:
+        return train_file
+    idx = np.sort(np.random.default_rng(0).choice(n, CI_MAX_TRAIN, replace=False))
+    sub = {k: np.asarray(d[k])[idx] for k in keys}
+    stem = os.path.join(tempfile.mkdtemp(), "ci_train")
+    tables_io.write(sub, stem, "hdf5")
+    return stem + ".hdf5"
+
+
+def run_taskset_1_estimation_only(
+    model_file: str | Path, test_file: str | Path, output_file: str | Path,
+) -> None:
+    _conclave_estimate_only(str(model_file), str(test_file), str(output_file))
+
+
+def run_taskset_1_training_and_estimation(
+    train_file: str | Path, test_file: str | Path, output_file: str | Path,
+) -> None:
+    train_file = _maybe_subsample_train(str(train_file))
+    _conclave_train_and_estimate(train_file, str(test_file), str(output_file))
+
+
+def test_example_taskset_1(setup_public_area: int, setup_submit_area: int) -> None:
+    # TS1 submission: only Task Set 1 deliverables are shipped (Task Set 2 closes later).
+    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,
+    )