diff --git a/CHANGELOG.md b/CHANGELOG.md
index 7221346c..1dab4e2a 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -57,6 +57,32 @@ in `CMakeLists.txt`) is derived from it.
   Suite-side only — no bngsim behavior changes. `parity_checks/rr_parity/
   overrides.py` gains the entry and `ode_jobs.json` re-bakes the one job record.
 
+- **Three `rr_parity` override keys that could never match are retired, and a
+  test now says so (#482 follow-up).** `MODEL0913003363:ode` (`TOL_OVERRIDES`),
+  `MODEL1006230083:ode` and `MODEL6963432821:ode` (both `NO_ORACLE_ADJUDICATED`)
+  name models the ODE manifest does not build: all three are `drop,no_model` in
+  the stage-1 corpus filter, their fetched files zero bytes (sha256 = the empty
+  string's), one of 305 corpus entries in that state. They have been stale since
+  the initial public release.
+
+  A stale key is silent rot. It reads as a live disposition while covering
+  nothing, and it is the shape a typo'd or renamed `model_id` takes — an override
+  meant to reclassify a row that never did. The only detection was a
+  `print("WARN: ...")` in `build_ode_jobs.py` / `build_ssa_jobs.py`, which
+  nothing fails on and which cannot run at all in a checkout without
+  `$BIOMODELS_SEDML_DIR`. That is why three accumulated unnoticed.
+
+  Each is retired as a comment rather than deleted, carrying its disposition
+  forward so it is not re-derived if the corpus ever fetches the model
+  successfully — the same treatment `MODEL2205030001:ode` got when GH #119
+  retired it. `test_committed_overrides_have_no_stale_keys` then asserts the
+  committed overrides against the committed manifests for both regimes, which
+  needs neither the mirror nor the build. Verified to fail on an injected stale
+  key.
+
+  Suite-side only; no bngsim behaviour changes, and no job's disposition moves —
+  none of these keys was reaching a job to begin with.
+
 ## [0.15.0] - 2026-08-23
 
 ### Added
diff --git a/parity_checks/rr_parity/overrides.py b/parity_checks/rr_parity/overrides.py
index 776042b3..24d8d7dd 100644
--- a/parity_checks/rr_parity/overrides.py
+++ b/parity_checks/rr_parity/overrides.py
@@ -348,16 +348,18 @@
             "(both engines) zero real columns remain -> PASS. Stiff IVP."
         ),
     },
-    "MODEL0913003363:ode": {
-        "rtol": 1e-10,
-        "atol": 1e-16,
-        "reason": (
-            "Tolerance artifact (both engines converge when tightened). 'm' + 3 "
-            "other columns diverge sharply at the sweep tol (worst reldiv 0.951, "
-            "an early transient); at 1e-10/1e-16 (both engines) zero real columns "
-            "remain -> PASS. Ill-conditioned transient at the loose default."
-        ),
-    },
+    # MODEL0913003363:ode — RETIRED 2026-08-25 as a stale key (GH #482 follow-up).
+    # The model is `drop,no_model` in the stage-1 corpus filter
+    # (benchmarks/suites/biomodels/manifest.csv): its fetched file is zero bytes
+    # (sha256 = the empty string's, e3b0c442...b855), one of 305 corpus entries in
+    # that state. build_ode_jobs.py therefore builds no job for it and the key
+    # could never match — `stale_keys` has flagged it since the initial public
+    # release. Recorded here rather than deleted outright, so the disposition is
+    # not re-derived if the corpus ever fetches the model successfully:
+    # tolerance artifact, both engines converge when tightened. 'm' + 3 other
+    # columns diverged sharply at the sweep tol (worst reldiv 0.951, an early
+    # transient); at rtol/atol 1e-10/1e-16 on BOTH engines zero real columns
+    # remained -> PASS. Ill-conditioned transient at the loose default.
     "BIOMD0000000827:ode": {
         "rtol": 1e-10,
         "atol": 1e-16,
@@ -457,15 +459,13 @@
             "below the significance floor). bngsim confirmed."
         ),
     },
-    "MODEL1006230083:ode": {
-        "verdict": "confirm",
-        "issue": "GH #117",
-        "reason": (
-            "RR CVODE CV_TOO_MUCH_WORK + COPASI failed. Pure rate-rule ODE system "
-            "on 5 parameters (no reactions). Independent scipy LSODA/BDF reproduce "
-            "to max_rel=0 (peak-rel ~1e-8). bngsim confirmed."
-        ),
-    },
+    # MODEL1006230083:ode — RETIRED 2026-08-25 as a stale key (GH #482 follow-up),
+    # same cause as MODEL6963432821 below and MODEL0913003363 in TOL_OVERRIDES:
+    # `drop,no_model` in the stage-1 corpus filter, so no job is built and the key
+    # can never match. The GH #117 adjudication it carried, kept for the record:
+    # RR CVODE CV_TOO_MUCH_WORK + COPASI failed; pure rate-rule ODE system on 5
+    # parameters (no reactions); independent scipy LSODA/BDF reproduce to
+    # max_rel=0 (peak-rel ~1e-8); verdict confirm — bngsim confirmed.
     "MODEL1112050001:ode": {
         "verdict": "confirm",
         "issue": "GH #117",
@@ -495,15 +495,13 @@
             "confirmed."
         ),
     },
-    "MODEL6963432821:ode": {
-        "verdict": "confirm",
-        "issue": "GH #117",
-        "reason": (
-            "RR CVODE CV_TOO_MUCH_WORK + COPASI failed. Pure rate-rule ODE system "
-            "on 7 parameters. Independent scipy LSODA/BDF reproduce to max_rel=0 "
-            "(peak-rel ~4e-9). bngsim confirmed."
-        ),
-    },
+    # MODEL6963432821:ode — RETIRED 2026-08-25 as a stale key (GH #482 follow-up),
+    # same cause as MODEL1006230083 above: `drop,no_model` in the stage-1 corpus
+    # filter, so no job is built and the key can never match. The GH #117
+    # adjudication it carried, kept for the record: RR CVODE CV_TOO_MUCH_WORK +
+    # COPASI failed; pure rate-rule ODE system on 7 parameters; independent scipy
+    # LSODA/BDF reproduce to max_rel=0 (peak-rel ~4e-9); verdict confirm — bngsim
+    # confirmed.
     "MODEL9811206584:ode": {
         "verdict": "confirm",
         "issue": "GH #117",
diff --git a/parity_checks/tests/test_overrides.py b/parity_checks/tests/test_overrides.py
index 5cfeea6a..c284223a 100644
--- a/parity_checks/tests/test_overrides.py
+++ b/parity_checks/tests/test_overrides.py
@@ -15,6 +15,10 @@
 stale_keys
   * flags an authored key whose model is absent from the built set, and stays
     silent when every key is matched
+  * the COMMITTED overrides carry no stale key against the COMMITTED manifests
+    (the regression guard: build_*_jobs.py only `print`s a WARN, and cannot run
+    at all without $BIOMODELS_SEDML_DIR, so nothing else notices a key going
+    stale — three had accumulated undetected since the initial public release)
 
 _job_overrides (runner side)
   * reads a tol Override into a {rtol, atol} dict, and a known_artifact /
@@ -25,10 +29,15 @@
 
 from __future__ import annotations
 
+from pathlib import Path
+
 import overrides as ov
-from _core import Job, Oracle, Override
+import pytest
+from _core import Job, Oracle, Override, read_manifest
 from rr_run import _job_overrides
 
+RR_PARITY = Path(__file__).resolve().parents[1] / "rr_parity"
+
 
 # --------------------------------------------------------------------------- #
 # overrides_for
@@ -177,3 +186,29 @@ def test_job_overrides_reads_no_oracle_adjudicated():
     tol, artifact, invalid_ref, adjudication = _job_overrides(job)
     assert tol is None and artifact is None and invalid_ref is None
     assert adjudication == ("confirm", "scipy BDF reproduces to max_rel=0", "GH #117")
+
+
+# --------------------------------------------------------------------------- #
+# The committed state
+# --------------------------------------------------------------------------- #
+@pytest.mark.parametrize("method,manifest", [("ode", "ode_jobs.json"), ("ssa", "ssa_jobs.json")])
+def test_committed_overrides_have_no_stale_keys(method, manifest):
+    """No authored override key may name a model the committed manifest doesn't
+    build.
+
+    A stale key is silent rot: it reads as a live disposition while covering
+    nothing, and it is the shape a typo'd or renamed model_id takes — an
+    override that was meant to reclassify a row and never did. The only existing
+    detection is a ``print("WARN: ...")`` in build_ode_jobs.py / build_ssa_jobs.py,
+    which nothing fails on and which cannot run at all in a checkout without
+    ``$BIOMODELS_SEDML_DIR``; three ODE keys had accumulated undetected since the
+    initial public release (retired in the GH #482 follow-up, with their
+    dispositions kept as comments in overrides.py). This asserts the committed
+    overrides against the committed manifests, which needs neither.
+
+    Retiring a key is the fix when the model genuinely left the corpus; correcting
+    it is the fix when the model_id is wrong. Both are better than a warning
+    nobody reads.
+    """
+    _, jobs = read_manifest(RR_PARITY / manifest)
+    assert ov.stale_keys({j.model_id for j in jobs}, method) == []