diff --git a/README.md b/README.md
index 375d060..255f91e 100644
--- a/README.md
+++ b/README.md
@@ -174,7 +174,8 @@ From the 2026 evaluation (details, seeds and downloadable data on the site):
   but returns a shortest path on 74.3% of solvable boards (71.4% to 77.0%) against 85.6% (83.2% to
   87.7%) for Euclidean. On the same 931 boards that is 11.3 points fewer (paired bootstrap 95% CI
   9.1 to 13.3); where only one heuristic was optimal it was Manhattan on 1 board and Euclidean on
-  106 (exact McNemar p < 0.001). Neither heuristic is admissible on this hex grid.
+  106 (exact McNemar p < 0.001). Neither heuristic is admissible on this hex grid, although
+  our 2022 report argued that both were.
 - **No detectable disagreement with the original Python** on the pairings feasible in Python
   (480 games; every Newcombe 95% interval for the six win-rate differences contains 0). With 80
   games per pairing this check can only detect differences larger than about ±15 points.
diff --git a/web/src/app/astar/page.tsx b/web/src/app/astar/page.tsx
index 36b5ac1..38510b9 100644
--- a/web/src/app/astar/page.tsx
+++ b/web/src/app/astar/page.tsx
@@ -68,10 +68,11 @@ export default function AstarPage() {
               {formatPStatement(o.mcnemar.pValue)}).
             </Finding>
             <Finding>
-              <strong>Why: neither heuristic is admissible here.</strong> On this hex grid a step
-              like (1, −1) costs 1, but Manhattan scores it 2 and Euclidean √2. Overestimating makes
-              the search greedier: fewer expansions, longer paths. The two sample inputs happen to
-              be solved optimally (8 and 13 cells, confirmed by BFS).
+              <strong>Why: neither heuristic is admissible here.</strong> Our 2022 report argued
+              that both were; that was wrong. On this hex grid a step like (1, −1) costs 1, but
+              Manhattan scores it 2 and Euclidean √2. Overestimating makes the search greedier:
+              fewer expansions, longer paths. The two sample inputs happen to be solved optimally (8
+              and 13 cells, confirmed by BFS).
             </Finding>
           </ul>
           <PairedStudySummaryView summary={summary} />
diff --git a/web/src/app/page.tsx b/web/src/app/page.tsx
index 3467f6b..8c71ed8 100644
--- a/web/src/app/page.tsx
+++ b/web/src/app/page.tsx
@@ -239,6 +239,11 @@ export default function HomePage() {
               colour that blocks the search. A notebook experiment compared the heuristics on
               hundreds of random boards.
             </p>
+            <p className="text-muted-foreground mt-3">
+              <strong className="text-foreground">Corrected in 2026:</strong> our report argued that
+              both heuristics were admissible. On this hex grid neither is, so A* can return a
+              longer path than the shortest one. The A* Lab measures how often.
+            </p>
             <Button asChild variant="outline" className="mt-5">
               <Link href="/astar">
                 Open the A* Lab <ArrowRight />