diff --git a/scripts/audit_annotations.yml b/scripts/audit_annotations.yml
index 19ada0c..2841d17 100644
--- a/scripts/audit_annotations.yml
+++ b/scripts/audit_annotations.yml
@@ -22,9 +22,6 @@ datasets:
     description: ACS occupation summary
     provenance: constructed-lost
     note: construction (ACS filters, grouping, sorting) described in lecture prose only
-  assignat.xlsx:
-    description: Assignat issuance / price data, French Revolution
-    provenance: author-assembled
   bbh_macro_quarterly.csv:
     description: Macro quarterly series (Bhandari et al.)
     provenance: constructed-lost
@@ -33,24 +30,10 @@ datasets:
     description: Michigan survey monthly series (Bhandari et al.)
     provenance: constructed-lost
     note: same Zenodo replication package; extraction not scripted
-  caron.npy:
-    description: French Revolution money balances (Caron)
-    provenance: author-assembled
-    note: hand-prepared NumPy array, no construction record
-  chapter_3.xlsx:
-    description: Hyperinflation tables, Sargent "Ends of Four Big Inflations"
-    provenance: author-assembled
-    note: hand-transcribed by authors
   dataBHS.mat:
     description: US consumption/income series, MATLAB replication bundle
     provenance: verbatim
     flags: [lectures-root]
-  dette.xlsx:
-    description: French government debt series
-    provenance: author-assembled
-  fig_3.xlsx:
-    description: French Revolution fiscal data
-    provenance: author-assembled
   fred_data.csv:
     description: FRED snapshot — GS1, GS5, GS10, DFII5, DFII10, USREC
     provenance: constructed-lost
@@ -71,17 +54,6 @@ datasets:
     description: Hansen–Jagannathan (1991) asset-returns bundle
     provenance: constructed-lost
     note: lecture documents 3 sources (FRED yields deflated by CPIAUCSL, …); no build script
-  longprices.xls:
-    description: Long-run price levels (Sargent–Velde)
-    provenance: author-assembled
-    note: hand-transcribed by authors
-  mpd2020.xlsx:
-    description: Maddison Project Database 2020 (GDP per capita, long run)
-    provenance: verbatim
-  nom_balances.npy:
-    description: Nominal balances, French Revolution
-    provenance: author-assembled
-    note: hand-prepared NumPy array, no construction record
   test_pwt.csv:
     description: Penn World Table 7.0 extract
     provenance: author-assembled
diff --git a/scripts/build_audit.py b/scripts/build_audit.py
index 673a700..26bca89 100644
--- a/scripts/build_audit.py
+++ b/scripts/build_audit.py
@@ -354,6 +354,15 @@ def scan(repos_dir: Path):
     manifests = load_manifests()
     migration = load_yaml(MIGRATION)
 
+    # A migrated dataset's manifest is its source of truth, so it must NOT also
+    # carry an `audit_annotations.yml` entry (that file's own header states the
+    # rule). Nothing enforced it before: annotations are consulted only as a
+    # fallback when a manifest is absent, so a stale entry is inert at runtime
+    # and a green audit said nothing about the invariant. Computed from the two
+    # static sources rather than from the scan, so it still fires for a dataset
+    # no lecture currently references.
+    dual_recorded = sorted(set(manifests) & set(datasets_ann))
+
     repos, all_refs, all_api, unscanned_nb = {}, [], [], []
     for name in SCAN_REPOS:
         res = scan_repo(name, repos_dir)
@@ -540,6 +549,12 @@ def scan(repos_dir: Path):
             "missing_api_annotations": sorted(api_missing),
             "migration_inconsistencies": mig_problems,
             "unscanned_notebooks": sorted(unscanned_nb),
+            "dual_recorded": [
+                f"{f}: has a lectures/{f}.yml manifest AND an "
+                f"audit_annotations.yml datasets: entry — delete the annotation, "
+                f"the manifest is the source of truth"
+                for f in dual_recorded
+            ],
         },
         "stats": {
             "static_files": len(datasets),