diff --git a/app/api/routes/extraction.py b/app/api/routes/extraction.py
index bf6dc29..28f26ea 100644
--- a/app/api/routes/extraction.py
+++ b/app/api/routes/extraction.py
@@ -17,6 +17,7 @@
 from app.api.schemas.incident_contract import IncidentContract
 from app.core.config import (
     ESTIMATED_EXTRACTION_SECONDS,
+    EXTRACTION_ALLOW_RERUN,
     EXTRACTION_POLL_INTERVAL_SECONDS,
 )
 from app.core.errors.base import AppError
@@ -91,7 +92,7 @@ def create_extraction(
         )
 
     existing = get_extraction_by_input(db, input_id)
-    if existing is not None:
+    if existing is not None and not EXTRACTION_ALLOW_RERUN:
         raise AppError(
             "An extraction already exists for this input",
             status_code=409,
diff --git a/app/core/config.py b/app/core/config.py
index e956cc6..40ba28b 100644
--- a/app/core/config.py
+++ b/app/core/config.py
@@ -123,6 +123,12 @@
 # manual entry rather than guessed at.
 EXTRACTION_CHUNK_RETRIES = int(os.getenv("EXTRACTION_CHUNK_RETRIES", "1"))
 
+# One input gets one extraction, which makes re-testing the same narrative mean
+# re-uploading it. Switching this on lets a repeat run through; every run still
+# gets its own extraction and its own draft incident, so the old ones stay put.
+# Development only, leave it off anywhere real.
+EXTRACTION_ALLOW_RERUN = os.getenv("EXTRACTION_ALLOW_RERUN", "false").strip().lower() == "true"
+
 # The contract file the chunk registry reads its tiers and triggers from.
 INCIDENT_CONTRACT_PATH = Path(
     os.getenv("INCIDENT_CONTRACT_PATH", BASE_DIR / "contracts" / "schemas" / "incident-contract.yaml")
diff --git a/docker/.env.example b/docker/.env.example
index 446336b..3bb20f6 100644
--- a/docker/.env.example
+++ b/docker/.env.example
@@ -83,6 +83,10 @@ LLM_EXTRA_HEADERS=
 # --- Extraction -----------------------------------------------------------
 # Extra tries after a chunk's first answer fails validation.
 EXTRACTION_CHUNK_RETRIES=1
+# Let the same input be extracted more than once. Development only: it makes
+# testing a narrative repeatedly possible without re-uploading it. Each run
+# still gets its own extraction and its own draft incident.
+EXTRACTION_ALLOW_RERUN=false
 # Deployment context the extractor falls back on when the narrative is silent.
 # A request can override any of these through the extraction body's "defaults".
 FIREFORM_DEFAULT_COUNTRY=US
diff --git a/tests/test_v1_extraction.py b/tests/test_v1_extraction.py
index ec2f226..94d54fe 100644
--- a/tests/test_v1_extraction.py
+++ b/tests/test_v1_extraction.py
@@ -163,12 +163,26 @@ def test_409_input_not_ready(self, client, db):
     def test_409_extraction_already_exists(self, client, db):
         inp = _ready_input(db)
         existing = create_extraction(db, Extraction(input_id=inp.input_id))
-        resp, _ = self._post(client, inp.input_id)
+        # Pinned rather than left to the environment, so a developer running
+        # with the rerun flag on still tests the shipped behaviour.
+        with patch("app.api.routes.extraction.EXTRACTION_ALLOW_RERUN", False):
+            resp, _ = self._post(client, inp.input_id)
         assert resp.status_code == 409
         body = resp.json()
         assert body["error_code"] == "EXTRACTION_EXISTS"
         assert body["detail"]["existing_extract_id"] == str(existing.extract_id)
 
+    def test_202_rerun_allowed_when_flag_is_on(self, client, db):
+        # Development escape hatch: the same narrative can be extracted again
+        # instead of having to be re-uploaded. The earlier extraction stays.
+        inp = _ready_input(db)
+        existing = create_extraction(db, Extraction(input_id=inp.input_id))
+        with patch("app.api.routes.extraction.EXTRACTION_ALLOW_RERUN", True):
+            resp, _ = self._post(client, inp.input_id)
+        assert resp.status_code == 202
+        assert resp.json()["extract_id"] != str(existing.extract_id)
+        assert get_extraction(db, existing.extract_id) is not None
+
     def test_503_ollama_unavailable(self, client, db):
         inp = _ready_input(db)
         resp, _ = self._post(client, inp.input_id, ollama_up=False)