diff --git a/docs/agentic-dev-kit/fallback-review-panel-evidence.md b/docs/agentic-dev-kit/fallback-review-panel-evidence.md
index f0b3808..8b0ca3f 100644
--- a/docs/agentic-dev-kit/fallback-review-panel-evidence.md
+++ b/docs/agentic-dev-kit/fallback-review-panel-evidence.md
@@ -122,6 +122,47 @@ clause is `safety-critical-changes.md` rule 3's own wording, quoted from the
 sibling doctrine file rather than from a tracker issue — a review lens went
 looking for it on the tracker and could not source it.)
 
+### The session that measured withdrawal
+
+PR `#225` (`#41`, the required/optional manifest axis). Five rounds:
+
+| Round | What it reviewed | HIGH |
+|---|---|---|
+| 1 | the change | 2 |
+| 2 | the round-1 fix | 2 |
+| 3 | the round-2 fix | 2 |
+| 4 | the round-3 fix | 1 |
+| 5 | the withdrawal | 0 |
+
+Only round 1 reviewed the change. Its second HIGH was a real fail-open — the
+derived graph missed `lib/repo_root.sh`, which two shell engines `source`, so
+`kit_doctor` called a broken tree clean at exit 0 — and the fix was a bash
+`source` scanner: a new mechanism, prompted by a finding, which the filing rule
+says to file rather than build. It was built.
+
+Every HIGH in rounds 2–4 was inside that scanner, and each was a defect in the
+previous round's fix for it:
+
+- **round 2** — the scan missed guarded `source` (`[ -f "$L" ] && source "$L"`),
+  and its exit-code test was confounded by its own fixture, so deleting the
+  mechanism under test left the suite green
+- **round 3** — the widened anchor matched `source` inside a quoted string, so
+  `echo "run this; source lib/dep.sh"` became a real edge; and the heredoc
+  tracker read `<<<` and arithmetic `<<` as openers, blinding whole files
+- **round 4** — the heredoc tracker recognised fewer opener forms than bash
+  accepts (`cmd <<A <<B`, `<<123`, `<<'MULTI WORD'`), so a real heredoc body was
+  scanned as code and manufactured an edge
+
+Both failure directions are tokenizer problems — command position, and heredoc
+boundaries — and no regex approximates either safely. The threshold was declared
+before round 4's findings ("a HIGH means the scanner comes out"), a HIGH arrived,
+and round 5 reviewed a delta of 90 insertions against 560 deletions and found
+none. The scanner is `#228`, filed with the four constructs that defeated it and
+an executable acceptance bar.
+
+The counterfactual is the whole point: filed at round 1, as the rule already
+required, that PR is two rounds rather than five.
+
 ### The cost measurement that does not exist
 
 `#163` Sink 2 asked for a cost measurement before "aim the re-run at the delta"
diff --git a/docs/agentic-dev-kit/fallback-review-panel.md b/docs/agentic-dev-kit/fallback-review-panel.md
index 65dac84..bfc006f 100644
--- a/docs/agentic-dev-kit/fallback-review-panel.md
+++ b/docs/agentic-dev-kit/fallback-review-panel.md
@@ -443,6 +443,22 @@ the trap inside them. So make each round *smaller*, not
 fewer: `safety-critical-changes.md` rule 3 ("a fix round addresses only what the
 review found") — a new mechanism gets filed, however squarely a finding prompted it.
 
+**When that rule was not followed, withdraw rather than patch.** The filing rule
+is prevention; this is its recovery, and it exists because the prevention was
+measured failing. A mechanism a round-1 finding prompted was built rather than
+filed, and every later round then found a HIGH *inside that mechanism* while the
+rest of the change drew nothing — rounds 2–5 reviewed the fixes, not the work.
+The signal is not a round count but **a fix that itself draws a HIGH, twice, in
+the same mechanism**: patching further is evidence-free, each patch a fresh guess
+at a problem the last guess misread. Take the mechanism out, file it with what
+defeated it, and let the round that removes it be the last — removal is the
+cheapest round there is, deleting surface rather than adding more for the next
+round to find.
+
+**Declare the withdrawal threshold before seeing a round's findings**, beside the
+stopping criterion. Named in advance it is a decision; reached afterwards it is
+indistinguishable from fatigue, and no reviewer can tell which it was.
+
 **Batch the fix round into one commit and one push, and aim the re-run at the
 delta.** Each push invalidates the current-head receipt, so each new head
 costs another required review — a fix round landed as four pushes buys four
diff --git a/kit-manifest.json b/kit-manifest.json
index 97b30c4..9265b6d 100644
--- a/kit-manifest.json
+++ b/kit-manifest.json
@@ -20,11 +20,11 @@
     },
     "docs/agentic-dev-kit/fallback-review-panel-evidence.md": {
       "role": "doctrine",
-      "sha256": "74548e09fc06caae09556e11b7501a0f29171188bc033ce20ddcdce2dccd2eb6"
+      "sha256": "599fc0c1995cd0510cf6d7fa1b46b19a7c84acd77bf56cf8f8db91441b9e1d26"
     },
     "docs/agentic-dev-kit/fallback-review-panel.md": {
       "role": "doctrine",
-      "sha256": "d397bacce89892cc425232668c91cee33b6b41a87ff6cbd7e72d27055ab402bd"
+      "sha256": "0ce06ccb47a2130fb30e1796aaae9ea747757493e77f2a11c568ab75d948f408"
     },
     "docs/agentic-dev-kit/safety-critical-changes.md": {
       "role": "doctrine",