diff --git a/assets/site.css b/assets/site.css
index 37a54ea..42ebb66 100644
--- a/assets/site.css
+++ b/assets/site.css
@@ -429,6 +429,45 @@ h4 { font-size: .8125rem; }
 .copy-btn svg { width: 14px; height: 14px; }
 .copy-btn[data-copied] { color: var(--accent-2); border-color: var(--accent); }
 
+/* A measured table inside a project block. The caption carries the conditions,
+   so the figures above it can be read without the repository open. */
+.measure {
+  margin: 0 0 24px;
+  overflow-x: auto;
+}
+.measure table {
+  width: 100%;
+  border-collapse: collapse;
+  font-variant-numeric: tabular-nums;
+}
+.measure caption {
+  caption-side: bottom;
+  padding-top: 12px;
+  text-align: left;
+  font-size: .8125rem;
+  line-height: 1.5;
+  color: var(--fg-3);
+}
+.measure th {
+  text-align: left;
+  font-size: .6875rem;
+  font-weight: 600;
+  letter-spacing: .12em;
+  text-transform: uppercase;
+  color: var(--fg-3);
+  padding: 0 20px 10px 0;
+  border-bottom: 1px solid var(--line-2);
+}
+.measure td {
+  padding: 11px 20px 11px 0;
+  border-bottom: 1px solid var(--line);
+  font-size: .9375rem;
+}
+.measure td:first-child { color: var(--fg-2); }
+.measure td:nth-child(2) { font-weight: 600; }
+.measure td:last-child { color: var(--fg-3); font-size: .8125rem; padding-right: 0; }
+.measure code { color: var(--fg-2); }
+
 /* --- Course --------------------------------------------------------------- */
 
 .course {
diff --git a/index.html b/index.html
index 85d4871..bb56ff7 100644
--- a/index.html
+++ b/index.html
@@ -298,6 +298,29 @@ <h3><a href="https://github.com/DrobyshevDev/stadion" rel="noopener">stadion</a>
             with a confidence interval and placed on that scale, so «better» means a distance from
             the best achievable result rather than a position on a leaderboard.
           </p>
+
+          <div class="measure">
+            <table>
+              <caption>Headroom between the tuned classical method and the computed optimum, with a 95% bootstrap interval on the paired difference.</caption>
+              <thead>
+                <tr><th>Task</th><th>Headroom</th><th>95% interval</th></tr>
+              </thead>
+              <tbody>
+              <tr><td><code>inventory</code></td><td>+0.4%</td><td>[+0.561, +1.031]</td></tr>
+              <tr><td><code>joint-pricing</code></td><td>+4.3%</td><td>[+3.794, +5.656]</td></tr>
+              <tr><td><code>pricing</code></td><td>+6.2%</td><td>[+1.265, +1.761]</td></tr>
+              <tr><td><code>queueing</code></td><td>+16.9%</td><td>[+3.418, +3.988]</td></tr>
+              <tr><td><code>supply-chain</code></td><td>+17.3%</td><td>[+5.505, +7.493]</td></tr>
+              <tr><td><code>energy</code></td><td>+26.6%</td><td>[+4.221, +4.690]</td></tr>
+              </tbody>
+            </table>
+          </div>
+          <p>
+            The spread is the point. Where the textbook answer is already near-optimal there
+            is almost nothing to win; where it cannot plan ahead, a quarter of the value is
+            still on the table. A benchmark whose tasks all leave generous headroom has
+            selected for problems the standard method happens to be bad at.
+          </p>
           <div class="project-links">
             <a href="https://github.com/DrobyshevDev/stadion" rel="noopener">Repository</a>
             <a href="https://pypi.org/project/stadion-rl/" rel="noopener">PyPI</a>
diff --git a/ru/index.html b/ru/index.html
index 9282466..5348927 100644
--- a/ru/index.html
+++ b/ru/index.html
@@ -300,6 +300,29 @@ <h3><a href="https://github.com/DrobyshevDev/stadion" rel="noopener">stadion</a>
             выдаётся с доверительным интервалом и ставится на эту шкалу, поэтому «лучше» означает
             расстояние до достижимого предела, а не позицию в таблице лидеров.
           </p>
+
+          <div class="measure">
+            <table>
+              <caption>Запас между настроенным классическим методом и посчитанным оптимумом, с 95% бутстрап-интервалом на парной разности.</caption>
+              <thead>
+                <tr><th>Задача</th><th>Запас</th><th>95% интервал</th></tr>
+              </thead>
+              <tbody>
+              <tr><td><code>inventory</code></td><td>+0.4%</td><td>[+0.561, +1.031]</td></tr>
+              <tr><td><code>joint-pricing</code></td><td>+4.3%</td><td>[+3.794, +5.656]</td></tr>
+              <tr><td><code>pricing</code></td><td>+6.2%</td><td>[+1.265, +1.761]</td></tr>
+              <tr><td><code>queueing</code></td><td>+16.9%</td><td>[+3.418, +3.988]</td></tr>
+              <tr><td><code>supply-chain</code></td><td>+17.3%</td><td>[+5.505, +7.493]</td></tr>
+              <tr><td><code>energy</code></td><td>+26.6%</td><td>[+4.221, +4.690]</td></tr>
+              </tbody>
+            </table>
+          </div>
+          <p>
+            Разброс здесь и есть содержание. Там, где учебный ответ уже почти оптимален,
+            выигрывать нечего; там, где он не умеет планировать вперёд, на столе остаётся
+            четверть ценности. Бенчмарк, у которого все задачи оставляют щедрый запас,
+            отобран под проблемы, на которых стандартный метод плох.
+          </p>
           <div class="project-links">
             <a href="https://github.com/DrobyshevDev/stadion" rel="noopener">Репозиторий</a>
             <a href="https://pypi.org/project/stadion-rl/" rel="noopener">PyPI</a>