diff --git a/Include/nmNum/nmData/nmDataSensitive.h b/Include/nmNum/nmData/nmDataSensitive.h
index 056f58e8..4acf003d 100644
--- a/Include/nmNum/nmData/nmDataSensitive.h
+++ b/Include/nmNum/nmData/nmDataSensitive.h
@@ -56,7 +56,7 @@ public:
         Distribution m_distribution;
         bool m_distributionInitialized;
         double m_mean, m_stdDev, m_peak, m_distributionMin, m_distributionMax;
-        bool distributionQuantile(double probability, double& value) const;
+        bool distributionQuantile(double probability, double& value, bool limitToPhysicalRange = true) const;
 		VariableSampling();
 
 		// JSON 序列化 / 反序列化
diff --git a/Src/nmNum/nmData/nmDataSensitive.cpp b/Src/nmNum/nmData/nmDataSensitive.cpp
index ff645df9..f7af77a8 100644
--- a/Src/nmNum/nmData/nmDataSensitive.cpp
+++ b/Src/nmNum/nmData/nmDataSensitive.cpp
@@ -16,10 +16,9 @@ double normalCdf(double z)
     return z < 0.0 ? tail : 1.0 - tail;
 }
 
-double normalQuantile(double probability)
+double normalQuantile(double probability, double low = -10.0, double high = 10.0)
 {
     // 在累积分布上二分求逆，避免引入新的数学库。
-    double low = -10.0, high = 10.0;
     for(int i = 0; i < 60; ++i) {
         const double mid = (low + high) * 0.5;
         if(normalCdf(mid) < probability) low = mid;
@@ -409,12 +408,14 @@ QVector<double> nmDataSensitive::VariableSampling::samplingValues() const
 }
 
 
-bool nmDataSensitive::VariableSampling::distributionQuantile(double probability, double& value) const
+bool nmDataSensitive::VariableSampling::distributionQuantile(double probability, double& value, bool limitToPhysicalRange) const
 {
     // 六种分布都由分位数取值；对数正态的均值、标准差仍使用参数原单位。
     if(!qIsFinite(probability) || probability <= 0.0 || probability >= 1.0) return false;
-    const bool bounded = m_distribution >= TRUNCATED_NORMAL;
-    const double low = m_distributionMin, high = m_distributionMax;
+    const bool specifiedBounds = m_distribution >= TRUNCATED_NORMAL;
+    const bool bounded = specifiedBounds || limitToPhysicalRange;
+    const double low = specifiedBounds ? m_distributionMin : m_modelValue.getMinValue().toDouble();
+    const double high = specifiedBounds ? m_distributionMax : m_modelValue.getMaxValue().toDouble();
     if(bounded && (!qIsFinite(low) || !qIsFinite(high) || low >= high)) return false;
     if(m_distribution == UNIFORM) value = low + probability * (high - low);
     else if(m_distribution == TRIANGULAR) {
@@ -436,14 +437,23 @@ bool nmDataSensitive::VariableSampling::distributionQuantile(double probability,
             sigma = std::sqrt(variance);
         }
         if(!qIsFinite(mean) || !qIsFinite(sigma) || sigma <= 0.0) return false;
+        double quantile = 0.0;
         if(bounded) {
-            const double a = logarithmic && low == 0.0 ? 0.0
-                : normalCdf(((logarithmic ? std::log(low) : low) - mean) / sigma);
-            const double b = normalCdf(((logarithmic ? std::log(high) : high) - mean) / sigma);
+            // 在允许范围内重新分配概率后取样，保持拉丁超立方分层，不裁剪或重抽越界样本。
+            double zLow = logarithmic && low == 0.0 ? -std::numeric_limits<double>::infinity()
+                : ((logarithmic ? std::log(low) : low) - mean) / sigma;
+            double zHigh = ((logarithmic ? std::log(high) : high) - mean) / sigma;
+            // 右尾利用正态分布对称性转到左尾，避免两个接近 1 的累积概率相减。
+            const bool reflected = zLow > 0.0;
+            if(reflected) { const double previousLow = zLow; zLow = -zHigh; zHigh = -previousLow; }
+            const double a = normalCdf(zLow), b = normalCdf(zHigh);
             if(!(b > a)) return false;
-            probability = a + probability * (b - a);
-        }
-        value = mean + sigma * normalQuantile(probability);
+            probability = a + (reflected ? 1.0 - probability : probability) * (b - a);
+            // 超过 40 个标准差时双精度尾概率已下溢；缩小二分区间，同时保留实际边界。
+            quantile = normalQuantile(probability, qMax(zLow, -40.0), qMin(zHigh, 40.0));
+            if(reflected) quantile = -quantile;
+        } else quantile = normalQuantile(probability);
+        value = mean + sigma * quantile;
         if(logarithmic) value = std::exp(value);
     }
     return qIsFinite(value);
diff --git a/Src/nmNum/nmSubWxs/nmWxSensitive.cpp b/Src/nmNum/nmSubWxs/nmWxSensitive.cpp
index 8f2b9714..8ffbf8bc 100644
--- a/Src/nmNum/nmSubWxs/nmWxSensitive.cpp
+++ b/Src/nmNum/nmSubWxs/nmWxSensitive.cpp
@@ -323,7 +323,9 @@ protected:
             // 两端加密以平滑显示尾部，有界分布的横轴保留精确上下限。
             const double probability = (1.0 - std::cos(3.141592653589793 * i / 200.0)) * 0.5;
             cumulativeValues[i] = bounded ? probability : 0.0001 + 0.9998 * probability;
-            valid = m_variable.distributionQuantile(qBound(1e-10, cumulativeValues[i], 1.0 - 1e-10), x[i]);
+            // Skin 保留原分布绘图，其余参数显示与实际采样一致的受限分布。
+            valid = m_variable.distributionQuantile(qBound(1e-10, cumulativeValues[i], 1.0 - 1e-10), x[i],
+                                                     m_variable.m_varName != "Skin");
             if(bounded && i == 0) x[i] = m_variable.m_distributionMin;
             if(bounded && i == 200) x[i] = m_variable.m_distributionMax;
             valid = valid && convertSamplingValue(m_variable.m_varName, x[i], samplingBaseUnit(m_variable.m_varName),
@@ -1086,9 +1088,7 @@ bool nmWxSensitive::saveDataFromUI()
                 detailStack->setCurrentIndex(idx + 1);
                 return false;
             }
-            // 均值、众数及有界分布的上下限沿用确定性方法的系统允许范围。
-            if(valid && variable.m_distribution < nmDataSensitive::VariableSampling::UNIFORM)
-                valid = validSamplingValue(name, variable.m_mean);
+            // 均值描述原分布，可以在截断区间之外；显式上下限及众数仍需符合系统范围。
             if(valid && variable.m_distribution >= nmDataSensitive::VariableSampling::TRUNCATED_NORMAL)
                 valid = validSamplingValue(name, variable.m_distributionMin) &&
                         validSamplingValue(name, variable.m_distributionMax);
@@ -2119,28 +2119,7 @@ void nmWxSensitive::onPreviewClicked()
             QMessageBox::warning(this, tr("Sensitivity"), error);
             return;
         }
-        // 普通正态、对数正态仍可能抽到越界值；整批检查通过后才显示预览，不裁剪或重抽。
-        QVector<double> values;
-        for(int run = 0; run < m_sensitiveData.getMonteCarloCount(); ++run) {
-            if(!m_sensitiveData.getRunValues(run, values)) return;
-            for(int i = 0; i < m_sensitiveData.getVariableCount(); ++i) {
-                const nmDataSensitive::VariableSampling& variable = m_sensitiveData.getVariable(i);
-                if(!variable.getEnabled() || validSamplingValue(variable.m_varName, values[i])) continue;
-                const QString name = variable.m_varName;
-                const QString unit = variable.m_modelValue.getUnit();
-                double lower = 0.0, upper = 0.0, value = values[i];
-                samplingPhysicalRange(name, lower, upper);
-                convertSamplingValue(name, lower, samplingBaseUnit(name), unit, lower);
-                convertSamplingValue(name, upper, samplingBaseUnit(name), unit, upper);
-                convertSamplingValue(name, value, samplingBaseUnit(name), unit, value);
-                QMessageBox::warning(this, tr("Sensitivity"),
-                    tr("%1: run %2 sampled %3 %4 outside the allowed range [%5, %6]. Adjust the distribution settings or use a bounded distribution.")
-                    .arg(name).arg(run + 1).arg(formatSamplingValue(value, 6)).arg(unit)
-                    .arg(formatSamplingValue(lower, 6)).arg(formatSamplingValue(upper, 6)));
-                detailStack->setCurrentIndex(findVariableIndex(name) + 1);
-                return;
-            }
-        }
+        // 样本已在允许范围内生成，直接显示预览。
     }
     // Saphir 风格的连续运行列表，组合数再大也不显示分页控件。
     QDialog dialog(this);