diff --git a/README.md b/README.md
index 2ad5dbd..e35577d 100644
--- a/README.md
+++ b/README.md
@@ -94,11 +94,13 @@ uv run --python 3.13.7 --no-project --no-sync --offline python -m compileall -q
 uv run --python 3.13.7 --no-project --no-sync --offline python -m unittest discover -s tests -v
 ```
 
-The 12 tests execute the real SQLite evaluator and comparator against temporary
-fixture databases. They cover denied mutations and unchanged database contents,
+The regression tests execute the real SQLite evaluator and comparator against
+temporary fixture databases. They cover denied mutations and unchanged database contents,
 safe literals/comments/identifiers, instruction limits and connection cleanup,
 ordered versus unordered results, duplicate counts, types, NULLs, and the
-versioned evaluation/report contract. A failed test exits nonzero and fails CI.
+versioned evaluation/report contract. Fairness regressions cover short gold SQL
+leaks such as `SELECT 1`, clean prompts, and score eligibility in saved audits
+and reports. A failed test exits nonzero and fails CI.
 These are local fixture regressions, not measured model accuracy or acceptance
 of downloaded benchmarks. CI also checks Python syntax and commit whitespace;
 the repository currently has no configured formatter, linter, or type checker.
diff --git a/scripts/eval_sql_benchmark.py b/scripts/eval_sql_benchmark.py
index 234a6b9..344d1f4 100644
--- a/scripts/eval_sql_benchmark.py
+++ b/scripts/eval_sql_benchmark.py
@@ -146,10 +146,11 @@ def audit_run(run_dir: Path, cases: list[dict[str, Any]], run_metadata: dict[str
         if not case:
             continue
         gold_sql = normalize_sql(case.get("gold_sql", ""))
-        if len(gold_sql) < 40:
+        if not gold_sql:
             continue
         prompt_norm = normalize_sql(prompt_path.read_text(encoding="utf-8", errors="ignore"))
-        if gold_sql and gold_sql in prompt_norm:
+        # Check short queries too, without matching SELECT 1 inside SELECT 10 or SELECT 1.0.
+        if re.search(rf"(?<![\w$]){re.escape(gold_sql)}(?![\w$]|\.\d)", prompt_norm):
             findings.append(
                 {
                     "severity": "critical",
diff --git a/tests/test_fairness_audit.py b/tests/test_fairness_audit.py
new file mode 100644
index 0000000..e6abf0a
--- /dev/null
+++ b/tests/test_fairness_audit.py
@@ -0,0 +1,114 @@
+"""Regression coverage for exact gold SQL leaking into benchmark prompts."""
+
+import contextlib
+import io
+import json
+from pathlib import Path
+import sqlite3
+import sys
+import tempfile
+import unittest
+from unittest.mock import patch
+
+sys.path.insert(0, str(Path(__file__).resolve().parents[1] / "scripts"))
+import eval_sql_benchmark as evaluator
+import report_sql_benchmarks as reporter
+
+
+class FairnessAuditTests(unittest.TestCase):
+    def setUp(self):
+        self.temp = tempfile.TemporaryDirectory()
+        self.addCleanup(self.temp.cleanup)
+        self.root = Path(self.temp.name)
+        self.run = self.root / "run"
+        self.case_dir = self.run / "case-1"
+        self.case_dir.mkdir(parents=True)
+        self.metadata = {
+            "benchmark": "kaggledbqa",
+            "track": "raw",
+            "provider": "fixture",
+            "prompt_set_sha256": "fixture",
+        }
+        root_patch = patch.object(evaluator, "ROOT", self.root)
+        root_patch.start()
+        self.addCleanup(root_patch.stop)
+
+    def audit(self, gold_sql, prompt, **metadata):
+        (self.case_dir / "prompt.txt").write_text(prompt, encoding="utf-8")
+        return evaluator.audit_run(
+            self.run,
+            [{"id": "case-1", "gold_sql": gold_sql}],
+            self.metadata | metadata,
+        )
+
+    def test_gold_sql_in_prompt_blocks_scoring_at_any_length(self):
+        examples = [
+            ("SELECT 1", "The correct SQL is:\nSELECT 1;"),
+            ("SELECT 1", "The correct SQL is SELECT 1."),
+            ("SELECT NULL", "The correct SQL is: SELECT NULL"),
+            (" SELECT 1; ", "The correct SQL is:\nselect\t1;"),
+            ("SELECT first_column, second_column FROM example_table",
+             "The correct SQL is: SELECT first_column, second_column FROM example_table;"),
+        ]
+        for gold_sql, prompt in examples:
+            with self.subTest(gold_sql=gold_sql):
+                audit = self.audit(gold_sql, prompt)
+                self.assertEqual(audit["classification"], "diagnostic_contaminated")
+                self.assertFalse(audit["score_eligible"])
+                self.assertEqual(len(audit["findings"]), 1)
+                finding = audit["findings"][0]
+                self.assertEqual(finding["code"], "gold_sql_in_prompt")
+                self.assertEqual(finding["severity"], "critical")
+                self.assertEqual(finding["case_id"], "case-1")
+                self.assertEqual(Path(finding["file"]), Path("run/case-1/prompt.txt"))
+
+    def test_similar_numeric_literals_are_not_exact_gold_sql(self):
+        for sql in ("SELECT 10", "SELECT 1.0", "SELECT 1e3"):
+            with self.subTest(sql=sql):
+                audit = self.audit("SELECT 1", f"Example SQL: {sql};")
+                self.assertEqual(audit["classification"], "fair")
+                self.assertTrue(audit["score_eligible"])
+                self.assertEqual(audit["findings"], [])
+
+    def test_clean_prompts_and_empty_gold_sql_remain_fair(self):
+        for gold_sql in ("SELECT 1", "", " ;;; "):
+            with self.subTest(gold_sql=gold_sql):
+                audit = self.audit(gold_sql, "Return the constant one.")
+                self.assertEqual(audit["classification"], "fair")
+                self.assertTrue(audit["score_eligible"])
+                self.assertEqual(audit["findings"], [])
+
+    def test_contamination_cannot_be_promoted_by_track_or_mode(self):
+        for metadata in (
+            {"track": "schema-plan"},
+            {"track": "raw+retrieval-context"},
+            {"track": "bounded-agent+schema-plan"},
+            {"mode": "agentic_dbt"},
+        ):
+            with self.subTest(metadata=metadata):
+                audit = self.audit("SELECT 1", "The correct SQL is SELECT 1;", **metadata)
+                self.assertEqual(audit["classification"], "diagnostic_contaminated")
+                self.assertFalse(audit["score_eligible"])
+
+    def test_evaluator_and_report_preserve_short_query_contamination(self):
+        database = self.root / "fixture.sqlite"
+        sqlite3.connect(database).close()
+        (self.run / "run_metadata.json").write_text(json.dumps(self.metadata), encoding="utf-8")
+        (self.case_dir / "prompt.txt").write_text("The correct SQL is SELECT 1;", encoding="utf-8")
+        (self.case_dir / "predicted.sql").write_text("SELECT 1", encoding="utf-8")
+        cases = [{"id": "case-1", "mode": "sql", "db_path": str(database), "gold_sql": "SELECT 1"}]
+        with patch.object(evaluator, "load_run_cases", return_value=cases), \
+                patch.object(sys, "argv", ["eval", "--run-dir", str(self.run)]), \
+                contextlib.redirect_stdout(io.StringIO()):
+            self.assertEqual(evaluator.main(), 0)
+        audit = json.loads((self.run / "fairness_audit.json").read_text(encoding="utf-8"))
+        self.assertEqual(audit["classification"], "diagnostic_contaminated")
+        self.assertFalse(audit["score_eligible"])
+        row = reporter.row_for_run(self.run)
+        self.assertEqual(row["exact_matches"], 1)
+        self.assertEqual(row["fairness_classification"], "diagnostic_contaminated")
+        self.assertFalse(row["score_eligible"])
+
+
+if __name__ == "__main__":
+    unittest.main()