diff --git a/.pre-commit-config.yaml b/.pre-commit-config.yaml
index c11e500cc..5cd15ce3a 100644
--- a/.pre-commit-config.yaml
+++ b/.pre-commit-config.yaml
@@ -25,3 +25,12 @@ repos:
     - id: ruff-format
       name: format with ruff
       types_or: [ python, pyi, jupyter ]
+
+- repo: local
+  hooks:
+    - id: clean-numba-cache
+      name: clear stale numba caches
+      entry: python ci/clean_numba_cache.py
+      language: python
+      files: ^uxarray/(utils/(computing|numba_math)|grid/(angles|arcs|area|coordinates|geometry|intersections|point_in_face|utils))\.py$
+      pass_filenames: false
diff --git a/ci/clean_numba_cache.py b/ci/clean_numba_cache.py
new file mode 100644
index 000000000..c245f78b8
--- /dev/null
+++ b/ci/clean_numba_cache.py
@@ -0,0 +1,56 @@
+"""Delete numba's on-disk cache for ``uxarray``.
+
+Numba stamps each ``cache=True`` index (``*.nbi``) with the ``(mtime, size)`` of the
+source file that *defines* the function, and nothing else. A kernel that calls a jitted
+function from another module has that callee's code linked into its own cached object,
+so editing the callee leaves the caller's cache entry "valid" with the old body baked in.
+
+Run this after editing any shared jitted module, or let the pre-commit hook do it::
+
+    python ci/clean_numba_cache.py [--dry-run]
+"""
+
+import argparse
+import os
+from pathlib import Path
+
+PACKAGE_DIR = Path(__file__).resolve().parent.parent / "uxarray"
+CACHE_SUFFIXES = {".nbi", ".nbc"}
+
+
+def _cache_files():
+    """Yield numba cache files stored next to the package's sources."""
+    for path in PACKAGE_DIR.rglob("__pycache__/*"):
+        if path.suffix in CACHE_SUFFIXES:
+            yield path
+
+
+def main(argv=None):
+    parser = argparse.ArgumentParser(description=__doc__.split("\n\n")[0])
+    parser.add_argument(
+        "-n",
+        "--dry-run",
+        action="store_true",
+        help="list what would be removed without deleting anything",
+    )
+    args = parser.parse_args(argv)
+
+    files = sorted(set(_cache_files()))
+    for path in files:
+        if args.dry_run:
+            print(path)
+        else:
+            path.unlink(missing_ok=True)
+
+    if os.environ.get("NUMBA_CACHE_DIR"):
+        # Numba files caches there under a hash of each source directory, so they cannot
+        # be told apart from other projects' entries; leave them alone rather than guess.
+        print("NUMBA_CACHE_DIR is set: clear it by hand, it is not touched here")
+
+    verb = "would remove" if args.dry_run else "removed"
+    print(f"{verb} {len(files)} numba cache file(s)")
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())
diff --git a/docs/contributing.rst b/docs/contributing.rst
index cffde8e8f..a07d30b70 100644
--- a/docs/contributing.rst
+++ b/docs/contributing.rst
@@ -360,6 +360,16 @@ following command::
 
     $ pre-commit run --all-files
 
+One of the hooks, ``clean-numba-cache``, deletes numba's on-disk kernel cache when you stage a
+change to one of the shared jitted modules (``uxarray/utils/computing.py``, ``numba_math.py``
+and the ``uxarray/grid`` modules that call them). Numba only checks a cached kernel against
+the file that defines it, so a kernel that calls a jitted function from another file would
+otherwise keep running the old code after that function is edited. The hook only fires at
+commit time, so after editing one of these files, run the same cleanup by hand before you
+test the change::
+
+    $ python ci/clean_numba_cache.py
+
 3.5. Use Feature Branches
 -------------------------
 
diff --git a/test/grid/geometry/test_numba_flags.py b/test/grid/geometry/test_numba_flags.py
new file mode 100644
index 000000000..d283e80f2
--- /dev/null
+++ b/test/grid/geometry/test_numba_flags.py
@@ -0,0 +1,112 @@
+"""Guard the numba flags that the compensated (EFT) kernels depend on.
+
+``fastmath`` lets LLVM reassociate ``e = (a - (s - bp)) + (b - bp)`` to 0, silently undoing
+every compensated result from ``uxarray.utils.computing``. ``acc_sqrt_re``, ``_accux_gca``
+and ``_accux_constlat`` need ``error_model="numpy"`` so that 1/0 and sqrt(<0) give inf/nan
+for the status layer to mask, instead of raising.
+
+An ``inline="always"`` callee is compiled with its caller's flags, so both checks apply to
+every kernel this code is inlined into, not just where it is defined.
+"""
+
+import ast
+import importlib
+import inspect
+import pkgutil
+import sys
+import textwrap
+from pathlib import Path
+
+import pytest
+from numba.core.registry import CPUDispatcher
+
+import uxarray
+
+EFT_MODULE = "uxarray.utils.computing"
+NUMPY_ERROR_MODEL_KERNELS = {
+    "uxarray.utils.computing.acc_sqrt_re",
+    "uxarray.grid.intersections._accux_gca",
+    "uxarray.grid.intersections._accux_constlat",
+}
+
+
+def _qualname(kernel):
+    return f"{kernel.py_func.__module__}.{kernel.py_func.__name__}"
+
+
+def _callees(kernel):
+    """Yield the jitted functions called by name in ``kernel``'s body."""
+    namespace = kernel.py_func.__globals__
+    for node in ast.walk(ast.parse(textwrap.dedent(inspect.getsource(kernel.py_func)))):
+        if not isinstance(node, ast.Call):
+            continue
+        func = node.func
+        if isinstance(func, ast.Name):
+            target = namespace.get(func.id)
+        elif isinstance(func, ast.Attribute) and isinstance(func.value, ast.Name):
+            target = getattr(namespace.get(func.value.id), func.attr, None)
+        else:
+            continue
+        if isinstance(target, CPUDispatcher):
+            yield target
+
+
+def _compiled_into(kernel):
+    """Names of ``kernel`` and every kernel inlined into it, all built with its flags."""
+    names, stack = {_qualname(kernel)}, [kernel]
+    while stack:
+        for callee in _callees(stack.pop()):
+            name = _qualname(callee)
+            if callee.targetoptions.get("inline") == "always" and name not in names:
+                names.add(name)
+                stack.append(callee)
+    return names
+
+
+@pytest.fixture(scope="module")
+def kernels():
+    """Map each module-level kernel in uxarray to ``(kernel, names compiled into it)``."""
+    found = {}
+    for info in pkgutil.walk_packages(uxarray.__path__, "uxarray."):
+        try:
+            module = importlib.import_module(info.name)
+        except ImportError:  # missing optional dependency: its kernels can't run either
+            continue
+        for obj in vars(module).values():
+            if isinstance(obj, CPUDispatcher) and obj.py_func.__module__ == info.name:
+                found[_qualname(obj)] = (obj, _compiled_into(obj))
+    assert NUMPY_ERROR_MODEL_KERNELS <= found.keys()  # else the checks pass vacuously
+    return found
+
+
+def test_numpy_error_model(kernels):
+    offenders = sorted(
+        name
+        for name, (kernel, compiled) in kernels.items()
+        if compiled & NUMPY_ERROR_MODEL_KERNELS
+        and kernel.targetoptions.get("error_model") != "numpy"
+    )
+    assert not offenders, f"need error_model='numpy' to mask inf/nan: {offenders}"
+
+
+def test_no_fastmath(kernels):
+    eft = {
+        name: kernel
+        for name, (kernel, compiled) in kernels.items()
+        if any(n.startswith(EFT_MODULE + ".") for n in compiled)
+    }
+    offenders = [
+        name for name, kernel in eft.items() if kernel.targetoptions.get("fastmath")
+    ]
+    # Also scan the source, for kernels the import walk can't see: nested ones and the
+    # ``two_prod`` variant not selected on this machine.
+    for module in {kernel.py_func.__module__ for kernel in eft.values()}:
+        tree = ast.parse(Path(sys.modules[module].__file__).read_text())
+        offenders += [
+            f"{module}:{node.lineno}"
+            for node in ast.walk(tree)
+            if isinstance(node, ast.keyword) and node.arg == "fastmath"
+        ]
+    assert not offenders, (
+        f"fastmath would cancel the EFT error terms: {sorted(offenders)}"
+    )
diff --git a/uxarray/grid/utils.py b/uxarray/grid/utils.py
index bfa389371..458283303 100644
--- a/uxarray/grid/utils.py
+++ b/uxarray/grid/utils.py
@@ -529,7 +529,7 @@ def setter(self, value):
 #
 # NOTE: these are inlined into ``cache=True`` kernels in another module, and numba stamps its
 # cache against the defining file alone, so editing them does not invalidate a caller's cached
-# object. Clear ``uxarray/grid/__pycache__/*.nbi *.nbc`` after changing anything here.
+# object. Run ``python ci/clean_numba_cache.py`` after changing anything here.
 
 
 @njit(cache=True)
diff --git a/uxarray/utils/computing.py b/uxarray/utils/computing.py
index aff800372..098c39ee5 100644
--- a/uxarray/utils/computing.py
+++ b/uxarray/utils/computing.py
@@ -8,6 +8,12 @@
 hardware FMA when one is available (validated bit-exact at import time) and
 falls back to the Veltkamp split otherwise, so there is no hard FMA dependency.
 
+.. warning::
+    Every primitive here is inlined into ``cache=True`` kernels in other modules. Numba
+    stamps a cache entry against the file that *defines* the kernel only,
+    so after editing this file those callers keep running the old body from their cache,
+    with no warning. Run ``python ci/clean_numba_cache.py`` before testing changes here.
+
 Python/Numba port of the AccuSphGeom C++ library (EFT tier only; the adaptive
 Shewchuk predicate and exact-arithmetic fallback tiers are not ported):