diff --git a/domain/prediction/src/main/kotlin/dev/privacyllc/period/domain/prediction/PersonalPredictionEngine.kt b/domain/prediction/src/main/kotlin/dev/privacyllc/period/domain/prediction/PersonalPredictionEngine.kt
index e4c7efd..4181d0d 100644
--- a/domain/prediction/src/main/kotlin/dev/privacyllc/period/domain/prediction/PersonalPredictionEngine.kt
+++ b/domain/prediction/src/main/kotlin/dev/privacyllc/period/domain/prediction/PersonalPredictionEngine.kt
@@ -421,6 +421,32 @@ class PersonalPredictionEngine : PredictionEngine {
         const val NOT_YET_MAX_PENALTY = 0.30
         const val MINIMUM_FOR_MEDIUM = 2
         const val HIGH_THRESHOLD = 0.55
-        const val MEDIUM_THRESHOLD = 0.44
+        /**
+         * Raised from 0.30 when [EXPECTED_ERROR_PER_SCALE] landed, and not as a
+         * separate opinion — the two are one change.
+         *
+         * The old accuracy term deflated every score, including scores that had
+         * earned better, so the thresholds sat where they had to sit to let a
+         * stable user through at all. Once accuracy stopped punishing errors the
+         * window already admits, every calibrated forecast scored near its own
+         * agreement and 0.30 no longer separated anything: the §51 variable user
+         * — cycles from 25 to 37 — landed at 0.32 and began reading Medium
+         * beside an eleven-day window.
+         *
+         * Measured at 400 seeds, not guessed. The ordering agreement produces is
+         * already right (stable ~0.66, drifting ~0.52, variable ~0.32); only the
+         * line was in the wrong place. 0.40 puts it in the gap: stable reads High
+         * 88% of the time late in her history, variable reads Low 93%, and
+         * drifting — tracked to within a day and a quarter — keeps Medium. 0.44
+         * was measured too and rejected: it pushes drifting back toward Low at
+         * 21%, telling a woman the app has lost the thread of a cycle it is
+         * following well.
+         *
+         * Tightening AGREEMENT_SENSITIVITY instead was tried across 1.0, 1.2 and
+         * 1.4 and rejected: agreement multiplies into every score, so every value
+         * that quietened the variable user also took High from the stable one,
+         * and 1.0 broke §51 outright.
+         */
+        const val MEDIUM_THRESHOLD = 0.40
     }
 }