--- a
+++ b/tests/acceptance/test_external_studies.py
@@ -0,0 +1,119 @@
+import contextlib
+from pathlib import Path
+
+import pytest
+
+from ehrql.main import load_dataset_definition, load_measure_definitions
+
+
+# These tests specify dataset definitions in other repositories which we want to ensure
+# we don't accidentally break. In order to keep tests hermetic and deterministic, we
+# copy the study code into the repo and commit it (taking care to copy just the files
+# needed to evalute the dataset definition). The `update_external_studies.py` script
+# handles this. It can be invoked as:
+#
+#     python -m tests.acceptance.update_external_studies
+#
+# Or via just as:
+#
+#     just update-external-studies
+#
+# This is run automatically by a scheduled action which will create a PR if there are
+# any changes to be made.
+EXTERNAL_STUDIES = {
+    "test-age-distribution": dict(
+        repo="opensafely/test-age-distribution",
+        branch="main",
+        file_globs=[
+            "analysis/dataset_definition.py",
+        ],
+        dataset_definitions=["analysis/dataset_definition.py"],
+    ),
+    "qof-diabetes": dict(
+        repo="opensafely/qof-diabetes",
+        branch="main",
+        file_globs=[
+            "analysis/dataset_definition_*.py",
+            "analysis/dm_dataset.py",
+            "analysis/codelists.py",
+            "analysis/variable_lib_helper.py",
+            "codelists/*.csv",
+        ],
+        dataset_definitions=[],
+        measure_definitions=[
+            "analysis/dataset_definition_dm017.py",
+            ("analysis/dataset_definition_dm020.py", ["--ifcchba-cutoff-val", "58"]),
+            ("analysis/dataset_definition_dm021.py", ["--ifcchba-cutoff-val", "75"]),
+        ],
+    ),
+    "waiting-list": dict(
+        repo="opensafely/waiting-list",
+        branch="main",
+        file_globs=[
+            "analysis/codelists.py",
+            "analysis/dataset_definition_clockstops.py",
+            "analysis/measures_checks.py",
+            "analysis/measures_opioid.py",
+            "codelists/*.csv",
+        ],
+        dataset_definitions=[
+            "analysis/dataset_definition_clockstops.py",
+        ],
+        measure_definitions=[
+            "analysis/measures_checks.py",
+            ("analysis/measures_opioid.py", ["--codelist", "opioid_codes"]),
+        ],
+    ),
+    "mainroute_cancer": dict(
+        repo="opensafely/mainroute_cancer",
+        branch="main",
+        file_globs=[
+            "analysis/codelists.py",
+            "analysis/dataset_definition.py",
+            "analysis/define_static_dataset.py",
+            "analysis/measures_demo.py",
+            "codelists/*.csv",
+        ],
+        dataset_definitions=[
+            "analysis/define_static_dataset.py",
+        ],
+        measure_definitions=[
+            (
+                "analysis/measures_demo.py",
+                ["--start-date", "2018-03-23", "--intervals", "67"],
+            ),
+        ],
+    ),
+}
+
+STUDY_DIR = Path(__file__).parent / "external_studies"
+
+
+@pytest.mark.parametrize(
+    "study_name,definition_file,load_function",
+    [
+        (study_name, dataset_def, load_dataset_definition)
+        for (study_name, config) in EXTERNAL_STUDIES.items()
+        for dataset_def in config["dataset_definitions"]
+    ]
+    + [
+        (study_name, measure_def, load_measure_definitions)
+        for (study_name, config) in EXTERNAL_STUDIES.items()
+        for measure_def in config.get("measure_definitions", ())
+    ],
+)
+def test_external_study(study_name, definition_file, load_function):
+    if isinstance(definition_file, tuple):
+        definition_file, user_args = definition_file
+    else:
+        user_args = []
+    study_path = STUDY_DIR / study_name
+    definition_path = study_path / definition_file
+    # Studies often use project-relative paths so ensure these resolve correctly
+    with contextlib.chdir(study_path):
+        # Load the dataset or measure definition. This tests that we can construct a
+        # valid query model graph from the definition. I think this is sufficient for
+        # these tests which are intended to ensure we don't accidentally break the API.
+        # If we're unable to execute a valid query, that's a separate class of problem
+        # for which we need separate tests.
+        assert load_function(definition_path, user_args=user_args, environ={})