diff --git a/src/lib/cycle/__tests__/prediction.test.ts b/src/lib/cycle/__tests__/prediction.test.ts
index 019daff4d..acd39024d 100644
--- a/src/lib/cycle/__tests__/prediction.test.ts
+++ b/src/lib/cycle/__tests__/prediction.test.ts
@@ -88,6 +88,17 @@ describe("cycle/prediction — cycle-length estimator (§1, §5)", () => {
     expect(est.lengthRounded).toBe(28);
   });
 
+  it("keeps a legitimate long (oligomenorrhea) cycle inside the hard bound", () => {
+    // A run of long-but-real cycles (ACOG: > 35 d is oligomenorrhea, still a
+    // real cycle; perimenopause routinely runs longer). A 52-day length must
+    // not be force-excluded by the hard ceiling — only the MAD fence / missed-
+    // log heuristic should ever drop it. Here the whole run is long, so the
+    // robust median sits near 50 and 52 is well inside the fence.
+    const est = estimateCycleLength([48, 50, 52, 49, 51], "PERIMENOPAUSE");
+    expect(est.cyclesObserved).toBe(5); // none clipped by HARD_CYCLE_MAX
+    expect(est.lengthRounded).toBe(50);
+  });
+
   it("perimenopause widens the fence (OUTLIER_K 4 keeps more)", () => {
     const lengths = [28, 30, 26, 40, 24, 32];
     const normal = estimateCycleLength(lengths, "TRYING_TO_CONCEIVE");
diff --git a/src/lib/cycle/types.ts b/src/lib/cycle/types.ts
index cab9114ee..425cd6e53 100644
--- a/src/lib/cycle/types.ts
+++ b/src/lib/cycle/types.ts
@@ -105,9 +105,23 @@ export const FERTILE_POST = 1;
 export const OUTLIER_K = 3.0;
 export const OUTLIER_K_PERIMENOPAUSE = 4.0;
 
-/** §5 — hard physiological cycle-length bounds (always outlier candidates). */
+/**
+ * §5 — hard physiological cycle-length bounds (always outlier candidates).
+ *
+ * The ceiling is a deliberately generous outlier *backstop*, not a clinical
+ * "normal" cap — the robust MAD fence + missed-log heuristic do the real
+ * outlier work. ACOG flags cycles > 35 d as oligomenorrhea, but such cycles are
+ * still genuine, and adolescents / perimenopausal users routinely run to ~50+ d.
+ * A hard cap of 45 force-excluded those legitimate long cycles, biasing the
+ * estimate toward the population middle for exactly the irregular users who most
+ * need an honest estimate. 60 d keeps the rare gross-error guard (a length this
+ * long is almost always a missed log, which the missed-log rule also catches)
+ * while letting real long cycles reach the MAD fence. The MAD fence is
+ * computed pre-exclusion, so raising the ceiling does not move the fence — it
+ * only stops the hard rule from pre-empting it on long but consistent cycles.
+ */
 export const HARD_CYCLE_MIN = 21;
-export const HARD_CYCLE_MAX = 45;
+export const HARD_CYCLE_MAX = 60;
 
 /** §5 — a single length ≥ this × median is treated as a probable missed-log. */
 export const MISSED_LOG_FACTOR = 1.75;