diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
index a46176299..f924f6457 100644
--- a/.github/workflows/ci.yml
+++ b/.github/workflows/ci.yml
@@ -203,13 +203,16 @@ jobs:
     if: ${{ needs.eval.outputs.testCode == 'True' }}
     strategy:
       matrix:
-        os: [ubuntu-latest, windows-latest, macos-12]
+        os: [ubuntu-latest, windows-latest, macos-latest]
         python-version: ['3.8', '3.9', '3.10', '3.11', '3.12']
         kind: [serial, other, dml, main, treatment, ray]
         exclude:
           # Serial tests fail randomly on mac sometimes, so we don't run them there
-          - os: macos-12
+          - os: macos-latest
             kind: serial
+          # Mac ARM doesn't support tensorflow versions compatible with python 3.8
+          - os: macos-latest
+            python-version: '3.8'        
           # Ray tests run out of memory on Windows
           - os: windows-latest
             kind: ray
@@ -249,6 +252,10 @@ jobs:
       uses: actions/setup-python@v5
       with:
         python-version: ${{ matrix.python-version }}
+    - name: Install libomp
+      # lightgbm needs libomp on mac
+      run: brew install libomp
+      if: matrix.os == 'macos-latest'
     - name: Install uv
       # check if we're running on windows
       run: ${{ runner.os == 'Windows' && 'irm https://astral.sh/uv/install.ps1 | iex' || 'curl -LsSf https://astral.sh/uv/install.sh | sh' }}
@@ -293,6 +300,56 @@ jobs:
         name: tests-${{ env.id_string }}
         path: ${{ env.id_string }}-test-results.xml
 
+  store-reqs-per-env:
+    name: Store requirements for LKG updates
+    if: (success() || failure()) && (github.event_name == 'workflow_dispatch' && !inputs.use_lkg || github.event_name == 'schedule')
+    strategy:
+      matrix:
+        kind: [tests]
+        os: [ubuntu-latest, windows-latest, macos-latest]
+        python-version: ['3.8', '3.9', '3.10', '3.11', '3.12']
+        extras: ["[tf,plt,dowhy,ray]"]
+        include:
+          # explicitly add the two notebook extras
+          - kind: notebooks
+            os: ubuntu-latest
+            python-version: '3.8'
+            extras: "[tf,plt,ray]"
+          - kind: notebooks
+            os: ubuntu-latest
+            python-version: '3.9'
+            extras: "[plt,dowhy]"
+        exclude:
+          # Mac ARM doesn't support tensorflow versions compatible with python 3.8
+          - os: macos-latest
+            python-version: '3.8'
+    runs-on: ${{ matrix.os }}
+    steps:
+    - name: Checkout repository
+      uses: actions/checkout@v4
+      with:
+        ref: ${{ env.ref }}
+    - name: Setup Python
+      uses: actions/setup-python@v5
+      with:
+        python-version: ${{ matrix.python-version }}
+    - name: Install uv
+      # check if we're running on windows
+      run: ${{ runner.os == 'Windows' && 'irm https://astral.sh/uv/install.ps1 | iex' || 'curl -LsSf https://astral.sh/uv/install.sh | sh' }}
+    - name: Install econml
+      # add all extras used by any of the individual tests, which is a superset of what's actually used anywhere
+      run: uv pip install --system -e .${{ matrix.extras }}
+    - name: Install notebook requirements
+      run: uv pip install --system jupyter jupyter-client nbconvert nbformat seaborn xgboost tqdm
+      if: matrix.kind == 'notebooks'
+    - name: Save installed packages
+      run: pip freeze --exclude-editable > lkg-${{ matrix.kind }}-${{ matrix.os }}-${{ matrix.python-version }}.txt
+    - name: Upload installed packages
+      uses: actions/upload-artifact@v4
+      with:
+        name: lkg-${{ matrix.kind }}-${{ matrix.os }}-${{ matrix.python-version }}
+        path: lkg-${{ matrix.kind }}-${{ matrix.os }}-${{ matrix.python-version }}.txt
+
   coverage-report:
     name: "Coverage report"
     needs: [tests, notebooks]
@@ -329,11 +386,11 @@ jobs:
       
   merge-artifacts:
     name: "Merge artifacts"
-    needs: [coverage-report, tests, notebooks]
+    needs: [coverage-report, tests, notebooks, store-reqs-per-env]
     if: success() || failure()
     strategy:
       matrix:
-        artifact: [requirements, tests, coverage, notebooks]
+        artifact: [requirements, tests, coverage, notebooks, lkg-tests, lkg-notebooks]
     runs-on: ubuntu-latest
     steps:
     - name: "Merge artifacts"
@@ -346,7 +403,54 @@ jobs:
       # Unlike with plain upload-artifact, there's no way to ignore the situation where no files are found when using the v4 merge action
       # (see https://github.com/actions/upload-artifact/issues/520), so just continue on error isntead
       continue-on-error: true
+
+  generate-lkg:
+    name: "Generate updated last known good files"
+    needs: [merge-artifacts]
+    if: (success() || failure()) && (github.event_name == 'workflow_dispatch' && !inputs.use_lkg || github.event_name == 'schedule')
+    strategy:
+      matrix:
+        kind: [tests, notebooks]
+        include:
+        - kind: tests
+          output-name: lkg
+        - kind: notebooks
+          output-name: lkg-notebook
+    runs-on: ubuntu-latest
+    steps:
+    - name: Checkout repository
+      uses: actions/checkout@v4
+      with:
+        ref: ${{ env.ref }}
+    - name: Setup Python
+      uses: actions/setup-python@v5
+    - name: Download files
+      uses: actions/download-artifact@v4
+      with:
+        name: lkg-${{ matrix.kind }}
+        path: requirements
+    - name: "Generate last known good versions"
+      run: python .github/workflows/generate_lkg.py requirements "lkg-${{ matrix.kind }}-(?P<os>[a-z]+)-(latest|[0-9.]+)-(?P<pyversion>[0-9.]+).txt" ${{ matrix.output-name }}.txt
+    - name: "Upload last known good versions"
+      uses: actions/upload-artifact@v4
+      with:
+        name: ${{ matrix.output-name }}
+        path: ${{ matrix.output-name }}.txt
         
+  merge-lkg:
+    name: "Merge last known good versions"
+    needs: [generate-lkg]
+    if: success() || failure()
+    runs-on: ubuntu-latest
+    steps:
+    - name: "Merge last known good versions"
+      uses: actions/upload-artifact/merge@v4
+      with:
+        # can't use just "lkg" for the name since that artifact already exists and merge won't overwrite it even when delete-merged is true
+        name: lkg-files
+        pattern: "lkg*"
+        delete-merged: true
+
   build:
     name: Build package
     needs: [eval]
diff --git a/.github/workflows/generate_lkg.py b/.github/workflows/generate_lkg.py
new file mode 100644
index 000000000..bc6a98a8d
--- /dev/null
+++ b/.github/workflows/generate_lkg.py
@@ -0,0 +1,217 @@
+# Copyright (c) PyWhy contributors. All rights reserved.
+# Licensed under the MIT License.
+
+import argparse
+import re
+from collections import defaultdict, namedtuple
+from os import listdir, path
+
+import packaging.version
+from packaging.version import Version
+
+# We have a list of requirements files, one per python version and OS.
+# We want to generate a single requirements file that specifies the requirements
+# for each package contained in any of those files, along with the constraints on python version
+# and OS that apply to each package.
+
+Combo = namedtuple('Combo', ['os', 'py_version'])
+
+# For each version of a package (say numpy==0.24.1), we'll have a set of os/py_version combos
+# where it was installed; the correct constraint will be the union of all these pairs.
+# However, for readability we'd like to simplify that when possible to something more readable.
+# For example, if numpy==0.24.1 is installed on all versions of python and all OSes, we can just say
+# "numpy==0.24.1"; if it's installed on all versions of python on ubuntu, we can say
+# "numpy==0.24.1; platform_system=='Linux'".
+
+
+# We'll precompute a dictionary of simple constraints, mapping from the sets of combos to a string representation
+# of the constraint.
+# For simplicity, we won't consider all possible constraints, just some easy to generate ones.
+# In the most general case we'll OR together constraints grouped by os
+def simple_constraint_map(all_combos: frozenset[Combo]) -> tuple[dict[frozenset[Combo], str],
+                                                                 dict[tuple[str, frozenset[Version]], str]]:
+    """
+    Represent simple constraints via dictionaries.
+
+    Parameters
+    ----------
+    all_combos : frozenset[Combo]
+        All of the possible os/py_version pairs
+
+    Returns
+    -------
+    (d1, d2): tuple[dict[frozenset[Combo], str], dict[tuple[str, frozenset[Version]], str]]
+        A tuple of two dictionaries.
+        The first dictionary maps from a constrained set of os/py_version pairs to the string representation of the
+        constraint.
+        The second dictionary maps from a tuple of (os, set of py_versions) to the string representation of the
+        constraint.
+    """
+    all_os = frozenset({combo.os for combo in all_combos})
+    all_py_versions = frozenset({combo.py_version for combo in all_combos})
+
+    # constraint_map will map from sets of os/py_version pairs to a string representation of the constraint
+    # that would restrict all possible combos to just that set;
+    # we'll look up the sets of os/py_version pairs that a package is installed on and use this map to generate
+    # the correct constraint for that package.
+    constraint_map = {}
+
+    # first generate simple os constraints, like "platform_system=='Linux'" or "platform_system!='Linux'"
+    for os in all_os:
+        # Get the set of all os/py_version pairs where the os is the given os and the py_version is anything
+        filtered_combos = frozenset({combo for combo in all_combos if combo.os == os})
+        constraint_map[filtered_combos] = f"; platform_system=='{os}'"
+        constraint_map[all_combos - filtered_combos] = f"; platform_system!='{os}'"
+
+    # now generate simple python version constraints,
+    # like "python_version=='3.8'"", "python_version!='3.8'"; "python_version<'3.8'", "python_version>'3.8'"
+    for i, py_version in enumerate(sorted(all_py_versions)):
+        # Get the set of all os/py_version pairs where the py_version is the given py_version and the os is anything
+        filtered_combos = frozenset({combo for combo in all_combos if combo.py_version == py_version})
+        constraint_map[filtered_combos] = f"; python_version=='{py_version}'"
+        constraint_map[all_combos - filtered_combos] = f"; python_version!='{py_version}'"
+
+        if i > 0:
+            less_than = frozenset({combo for combo in all_combos if combo.py_version < py_version})
+            constraint_map[less_than] = f"; python_version<'{py_version}'"
+        if i < len(all_py_versions)-2:
+            greater_than = frozenset({combo for combo in all_combos if combo.py_version > py_version})
+            constraint_map[greater_than] = f"; python_version>'{py_version}'"
+
+    # if every combination is present, we don't need to add any constraint
+    constraint_map[all_combos] = ""
+
+    # generate simple per-os python version constraints
+    # we include the os in the key because we might not have every combination for every os
+    # (e.g. maybe macos doesn't support python 3.8, in which case there won't be a combo for that, but there might
+    # be a combo for ubuntu with python 3.8; then if we see all versions of python 3.9 and up on macos, we don't need
+    # any python version constraint, whereas if we see all versions of python 3.9 and up on ubuntu,
+    # we still do need a constraint since 3.8 is missing")
+    os_map = {}
+    for os in all_os:
+        for i, py_version in enumerate(all_py_versions):
+            filtered_combos = frozenset({combo for combo in all_combos
+                                         if combo.os == os and combo.py_version == py_version})
+            os_map[(os, frozenset({py_version}))] = f"python_version=='{py_version}'"
+            if i > 0 and i < len(all_py_versions)-1:
+                os_map[(os, all_py_versions - frozenset({py_version}))] = f"python_version!='{py_version}'"
+
+            if i > 0:
+                os_map[(os, frozenset({py for py in all_py_versions
+                                       if py < py_version}))] = f"python_version<'{py_version}'"
+            if i < len(all_py_versions)-1:
+                os_map[(os, frozenset({py for py in all_py_versions
+                                       if py > py_version}))] = f"python_version>'{py_version}'"
+
+        # if every combination is present, we don't need to add any constraint for that os
+        os_map[(os, all_py_versions)] = ""
+
+    return constraint_map, os_map
+
+
+# Convert between GitHub Actions' platform names and Python's platform.system() names
+platform_map = {'macos': 'Darwin', 'ubuntu': 'Linux', 'windows': 'Windows'}
+
+
+def make_req_file(requirements_directory, regex):
+    """
+    Make a unified requirements file from a directory of requirements files.
+
+    Parameters
+    ----------
+    requirements_directory : str
+        Directory containing requirements files
+
+    regex : str
+        Regex to match requirements file names, must have named groups "os" and "pyversion"
+    """
+    req_regex = r'^(?P<pkg>.*?)==(?P<version>.*)$'  # parses requirements from pip freeze results
+    files = listdir(requirements_directory)
+
+    all_combos = set()
+
+    # We'll store the requirements for each version of each package in a dictionary
+    # (e.g. "numpy" -> {0.24.1 -> {Combo1, Combo2, ...}, 0.24.2 -> {Combo3, Combo4, ...}, ...})
+    # each entry of the inner dictionary will become a line in the requirements file
+    # (e.g. "numpy==0.24.1; platform_system=='Linux' and python_version=='3.8' or ...")
+    req_dict = defaultdict(lambda: defaultdict(set))  # package -> package_version -> set of Combos
+
+    for file in files:
+        match = re.match(regex, file)
+        if not match:
+            print(f"Skipping {file} because it doesn't match the regex")
+            continue
+        os = platform_map[match.group('os')]
+        py_version = packaging.version.parse(match.group('pyversion'))
+        combo = Combo(os, py_version)
+        all_combos.add(combo)
+
+        # read each line of the file
+        with open(path.join(requirements_directory, file)) as lines:
+            for line in lines:
+                match = re.search(req_regex, line)
+                pkg_version = packaging.version.parse(match.group('version'))
+                req_dict[match.group('pkg')][pkg_version].add(combo)
+
+    constraint_map, os_map = simple_constraint_map(frozenset(all_combos))
+    # list of all requirements, sorted by package name and version
+    reqs = []
+    for pkg, versions in sorted(req_dict.items()):
+        for version, combos in sorted(versions.items()):
+            combos = frozenset(combos)
+            req = f"{pkg}=={version}"
+
+            if combos in constraint_map:
+                suffix = constraint_map[combos]
+
+            else:
+                # we don't have a simple constraint for this package, so we need to generate a more complex one
+                # which will generally be of the form:
+                # "(platform_system=='os1' and (python_version=='py1' or python_version=='py2') or ...) or
+                #  (platform_system=='os2' and (python_version=='py3' or ...) ..."
+                #
+                # that is, we will OR together constraints grouped by os
+                # for some oses, we might find a nice representation for their python version constraints in the os_map
+                # (e.g. "python_version=='3.8'", or "python_version<'3.8'"), in which case we'll use that;
+                # for others, we'll have to OR together all of the relevant individual versions
+                os_constraints = []
+
+                os_versions = defaultdict(set)  # dictionary from os to set of python versions
+                for combo in combos:
+                    os_versions[combo.os].add(combo.py_version)
+
+                # for each os, generate the corresponding constraint
+                for os in sorted(os_versions.keys()):
+                    versions = os_versions[os]
+                    os_key = (os, frozenset(os_versions[os]))
+                    if os_key in os_map:
+                        constraint = os_map[os_key]
+                        if constraint == "":
+                            os_constraints.append(f"platform_system=='{os}'")
+                        else:
+                            os_constraints.append(f"platform_system=='{os}' and {constraint}")
+                    else:
+                        version_constraint = " or ".join([f"python_version=='{py_version}'"
+                                                          for py_version in sorted(os_versions[os])])
+                        os_constraints.append(f"platform_system=='{os}' and ({version_constraint})")
+                if len(os_constraints) == 1:  # just one os with correspondig python versions, can use it directly
+                    suffix = f"; {os_constraints[0]}"
+                else:  # need to OR them together
+                    suffix = f"; ({') or ('.join(os_constraints)})"
+
+            reqs.append(f"{req}{suffix}")
+
+    return '\n'.join(reqs)
+
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser(description='Generate requirements files for CI')
+    parser.add_argument('requirements_directory', type=str, help='Directory containing requirements files')
+    parser.add_argument('regex', type=str,
+                        help='Regex to match requirements file names, must have named groups "os" and "pyversion"')
+    parser.add_argument('output_name', type=str, help='File to write requirements to')
+    args = parser.parse_args()
+
+    reqs = make_req_file(args.requirements_directory, args.regex)
+    with open(args.output_name, 'w') as f:
+        f.write(reqs)
diff --git a/econml/tests/test_grf_python.py b/econml/tests/test_grf_python.py
index 2f4a5a773..3140dad72 100644
--- a/econml/tests/test_grf_python.py
+++ b/econml/tests/test_grf_python.py
@@ -494,10 +494,10 @@ def test_projection(self,):
         forest = RegressionForest(**config).fit(X, y)
         projector = np.ones((X.shape[0], 2)) / 2.0
         mean_proj, var_proj = forest.predict_projection_and_var(X, projector)
-        np.testing.assert_array_equal(mean_proj, mean)
-        np.testing.assert_array_equal(var_proj, var)
-        np.testing.assert_array_equal(var_proj, forest.predict_projection_var(X, projector))
-        np.testing.assert_array_equal(mean_proj, forest.predict_projection(X, projector))
+        np.testing.assert_allclose(mean_proj, mean)
+        np.testing.assert_allclose(var_proj, var)
+        np.testing.assert_allclose(var_proj, forest.predict_projection_var(X, projector))
+        np.testing.assert_allclose(mean_proj, forest.predict_projection(X, projector))
         return
 
     def test_feature_importances(self,):
@@ -547,7 +547,7 @@ def test_feature_importances(self,):
                             feature_importances /= tw[0]
                             totest = tree.tree_.compute_feature_importances(normalize=False,
                                                                             max_depth=max_depth, depth_decay=2.0)
-                            np.testing.assert_array_equal(feature_importances, totest)
+                            np.testing.assert_allclose(feature_importances, totest)
 
                             het_importances = np.zeros(n_features)
                             for it, (feat, depth, left, right, w) in\