diff --git a/README.md b/README.md
index 2a914d1..07018d3 100644
--- a/README.md
+++ b/README.md
@@ -1047,7 +1047,7 @@ provenance) instead of being tested; raise `--max-categorical-card` to test it a
 For CI, `--fail-under N` returns exit code `1` and explains the failing score
 on stderr when the overall representation score is below `N`; report output,
 including `--json`, remains on stdout. A score at or above the threshold exits
-successfully.
+successfully. `N` must be between 0 and 100; any other value exits `2` with an error.
 
 `--json` output carries a `provenance` block recording the SHA-256 of the input
 file, the `faircode` version, and the thresholds actually resolved for the run,
diff --git a/faircode/SPEC.md b/faircode/SPEC.md
index 115bcb3..2b71007 100644
--- a/faircode/SPEC.md
+++ b/faircode/SPEC.md
@@ -200,6 +200,11 @@ Grade bands:
 | D     | 40–54   | Strong imbalance / sparse subgroups                  |
 | F     | 0–39    | Severe imbalance or single-group dimensions          |
 
+`faircode profile --fail-under N` compares the score with `N` for a CI gate: exit 1 when the score is
+below `N`, exit 0 otherwise. `N` must be between 0 and 100, the range of the score. Any other value,
+including `nan`, exits 2 with "--fail-under must be between 0 and 100, got N" before the dataset is
+read, because a value above 100 would fail every run and a negative one would never fail (#1037).
+
 The score intentionally reflects **balance only**. Missing-data %, imbalance ratios, and
 intersectional gaps are surfaced as **flags** alongside the score, not folded into it - this keeps
 the two engines trivially in sync and the score easy to explain.
diff --git a/faircode/cli.py b/faircode/cli.py
index 5ba262d..80547a0 100644
--- a/faircode/cli.py
+++ b/faircode/cli.py
@@ -556,6 +556,11 @@ def main(argv: list[str] | None = None) -> int:
         if not args.sample and not args.csv:
             print("error: profile needs a csv argument (or --sample)", file=sys.stderr)
             return 2
+        if args.fail_under is not None and not 0 <= args.fail_under <= 100:
+            # The score is 0-100: above that the gate always fails, below 0 it never does (#1037).
+            print(f"error: --fail-under must be between 0 and 100, got {args.fail_under:g}",
+                  file=sys.stderr)
+            return 2
         if args.proxy_alpha is not None and not 0 < args.proxy_alpha <= 1:
             print("error: --proxy-alpha must be in (0, 1]", file=sys.stderr)
             return 2
diff --git a/tests/test_cli.py b/tests/test_cli.py
index ea220cd..aa83a5a 100644
--- a/tests/test_cli.py
+++ b/tests/test_cli.py
@@ -1730,3 +1730,22 @@ def test_reference_csv_with_a_blank_share_ignores_that_row(tmp_path, capsys):
     assert "NaN" not in out and "Infinity" not in out
     groups = {g["label"] for g in json.loads(out)["dimensions"][0]["reference"]["groups"]}
     assert groups == {"M", "F"}
+
+
+def test_fail_under_outside_0_to_100_is_a_usage_error(tmp_path, capsys):
+    """#1037: 150 would fail every run and -5 would never fail, so both are rejected."""
+    path = tmp_path / "balanced.csv"
+    path.write_text("sex\nM\nF\nM\nF\n", encoding="utf-8")
+    for value in ("150", "100.5", "-5", "-0.1", "nan", "inf"):
+        assert main(["profile", str(path), "--fail-under", value]) == 2
+        captured = capsys.readouterr()
+        assert "--fail-under must be between 0 and 100, got " in captured.err
+        assert captured.out == ""
+
+
+def test_fail_under_accepts_both_ends_of_the_range(tmp_path, capsys):
+    path = tmp_path / "balanced.csv"
+    path.write_text("sex\nM\nF\nM\nF\n", encoding="utf-8")
+    assert main(["profile", str(path), "--fail-under", "0"]) == 0
+    assert main(["profile", str(path), "--fail-under", "100"]) == 0
+    assert "must be between" not in capsys.readouterr().err