diff --git a/scripts/r49-corpus/audit_pairs_sourcenames.py b/scripts/r49-corpus/audit_pairs_sourcenames.py
index b1bc598..3f16b84 100644
--- a/scripts/r49-corpus/audit_pairs_sourcenames.py
+++ b/scripts/r49-corpus/audit_pairs_sourcenames.py
@@ -65,6 +65,16 @@ def main() -> int:
                     help="rename.py's renameable threshold; mirrored from leak_gate.py")
     ap.add_argument("--report", default=None, help="write the full JSON breakdown here")
     ap.add_argument("--show", type=int, default=25, help="example beats to print")
+    ap.add_argument("--filter-out", default=None, metavar="PATH",
+                    help="write a copy of the pairs with every contaminated ROW removed. "
+                         "Turns this detector into the fix for an already-built pair set: "
+                         "the contamination is in the beat, so dropping the row removes it "
+                         "outright, and on Hemingway that costs 0.96%% of the training data. "
+                         "Cheaper and cleaner than regenerating 70 beats against a second "
+                         "generator session, which would leave the set mixed-provenance. "
+                         "Refuses to write when a RESPONSE is contaminated -- that is a "
+                         "different fault (pairs built against an unrenamed corpus) and "
+                         "dropping rows would hide it rather than fix it.")
     a = ap.parse_args()
 
     ents_all = json.loads(Path(a.entities).read_text())
@@ -157,6 +167,40 @@ def main() -> int:
         }, ensure_ascii=False, indent=2), encoding="utf-8")
         print(f"\n  wrote {a.report}")
 
+    if a.filter_out:
+        if nr:
+            print("\n== REFUSING to write a filtered copy: the RESPONSE column is contaminated.")
+            print("   That is not the beat-generator fault and dropping rows would hide it.")
+            print("   Check the provenance `corpus` path -- the pairs were probably built")
+            print("   against an unrenamed corpus, and the fix is upstream of this file.")
+            return 2
+        if not (pos_ok and neg_ok):
+            print("\n== REFUSING to write a filtered copy: the controls did not pass, so the")
+            print("   set of rows to drop is not trustworthy.")
+            return 2
+        if len({r["_src"] for r in rows}) != 1:
+            print("\n== REFUSING to write a filtered copy from more than one --pairs file:")
+            print("   the output is a single file and would silently merge train and val.")
+            print("   Filter each input separately.")
+            return 2
+        drop = set()
+        for i, r in enumerate(rows):
+            if removed_pat.search(r.get("beat") or ""):
+                drop.add(i)
+        out = Path(a.filter_out)
+        kept = 0
+        with out.open("w", encoding="utf-8") as fh:
+            for i, r in enumerate(rows):
+                if i in drop:
+                    continue
+                r = {k: v for k, v in r.items() if k != "_src"}
+                fh.write(json.dumps(r, ensure_ascii=False) + "\n")
+                kept += 1
+        print(f"\n  wrote {out}: {kept} pairs kept, {len(drop)} dropped "
+              f"({len(drop) / len(rows):.2%})")
+        print("  ⚠ Re-run this audit against the filtered file before training on it. A fix")
+        print("    that is not read back is a claim, not a result.")
+
     ok = pos_ok and neg_ok and nb == 0 and nr == 0
     print(f"\n  GATE: {'PASS' if ok else 'FAIL'}")
     return 0 if ok else 2