Mark the serving-sensitivity re-runs on the Claude rows instead of a banner

app/src/app/model/[id]/page.tsx

@@ -1,6 +1,11 @@
 import type { Metadata } from "next";
 import Link from "next/link";
 import { notFound } from "next/navigation";
+import {
+  formatDelta,
+  servingSensitivityFor,
+} from "../../../lib/servingSensitivity";
+import { wouldRank } from "../../../lib/wouldRank";
 
 import rawData from "../../../data-summary.json";
 import versionRegistryJson from "../../../data.versions.json";
@@ -183,6 +188,7 @@ export default async function ModelPage({
           const rows = programRows(bench, id);
           const cases = hardestCases(bench, id);
           const coverage = coverageAccuracy(bench, id);
+          const sensitivity = servingSensitivityFor(id);
           const currencySymbol = summary.country === "uk" ? "£" : "$";
           return (
             <section
@@ -224,6 +230,37 @@ export default async function ModelPage({
                   }
                 />
               </div>
+              {summary.country === "us" && sensitivity ? (
+                <p className="mt-4 max-w-3xl text-sm leading-relaxed text-text-secondary">
+                  <span className="text-[10px] uppercase tracking-[0.14em] text-text-muted font-medium">
+                    Serving sensitivity
+                  </span>{" "}
+                  This row {sensitivity.boardTreatment}. The same model,{" "}
+                  {sensitivity.autoTreatment}, scores{" "}
+                  <span className="font-[family-name:var(--font-mono)] text-text">
+                    {sensitivity.autoExact.toFixed(1)}%
+                  </span>{" "}
+                  exact (
+                  {formatDelta(
+                    sensitivity.autoExact,
+                    summary.stat.exact ?? summary.stat.score,
+                  )}
+                  ) and would rank #
+                  {wouldRank(
+                    sensitivity.autoExact,
+                    bench.modelStats.filter((row) => row.condition === "no_tools"),
+                  )}{" "}
+                  on the live board; the board keeps the request shape its
+                  model card records.{" "}
+                  <Link
+                    href={sensitivity.noteHref}
+                    className="text-primary-strong underline-offset-2 hover:underline"
+                  >
+                    Read the note
+                  </Link>
+                  .
+                </p>
+              ) : null}
 
               <div className="mt-8 grid gap-8 lg:grid-cols-2">
                 <div>

app/src/components/ModelLeaderboard.tsx

@@ -19,17 +19,12 @@ import { canonicalScoreByModel } from "../lib/canonicalScore";
 import { wouldRank } from "../lib/wouldRank";
 import { isCurrentBoard, modelPageHref } from "../lib/boardScope";
 
-// Exact-match scores of the tool_choice: auto sensitivity runs
-// (sensitivity/claude-thinking-2026-08.md), scored on the same 1,973 outputs
-// the board scores (eleven outputs are excluded for every model). Their
-// "would rank" positions are derived from the live board rows at render time,
-// never typed by hand.
-const SENSITIVITY_EXACT = {
-  "claude-fable-5": 87.5,
-  "claude-opus-5": 86.2,
-  "claude-sonnet-5": 80.8,
-  "claude-fable-5.1": 88.2,
-} as const;
+import ServingSensitivityChip from "./ServingSensitivityChip";
+import {
+  NEXT_BOARD_HREF,
+  SENSITIVITY_DOC_HREF,
+  servingSensitivityFor,
+} from "../lib/servingSensitivity";
 import {
   rankWithFallbackScore,
   rankWithRecomputedScores,
@@ -238,6 +233,16 @@ export default function ModelLeaderboard({
     canRecomputeScores && baseNoTools.length > 0 && noTools.length === 0;
 
   const activeView = SENSITIVITY_VIEWS.find((v) => v.id === sensitivityView)!;
+  // The serving-sensitivity markers describe the live US board's rows. Their
+  // comparison is against the row's canonical exact-match score on the
+  // unfiltered board (the universe the re-runs were scored on), whatever
+  // metric, weighting or program filter the reader has selected.
+  const showSensitivity =
+    selectedView === "us" && isCurrentBoard(versionId, liveVersionId);
+  const unfilteredExact = (model: string): number => {
+    const row = baseNoTools.find((entry) => entry.model === model);
+    return row?.exact ?? row?.score ?? 0;
+  };
 
   // "Exact" means "within one currency unit," and that unit is country-
   // specific. Surface the right word in tooltips, captions, and the Options
@@ -291,7 +296,7 @@ export default function ModelLeaderboard({
       >
         Model rankings
       </h2>
-      {selectedView === "us" && isCurrentBoard(versionId, liveVersionId) && (
+      {showSensitivity && (
         <div
           className="card mt-5 px-5 py-4 animate-fade-up"
           style={{ animationDelay: "120ms" }}
@@ -300,35 +305,20 @@ export default function ModelLeaderboard({
             Serving sensitivity · August 2026
           </div>
           <p className="mt-2 text-sm leading-relaxed text-text-secondary">
-            Claude models skip extended thinking when the answer tool call is
-            forced, as it is in the request this board sends every model whose
-            selected answer contract is the forced tool (rows on the JSON
-            contract, whether their provider rejects a forced tool or their
-            model card selects JSON, answer as a JSON object); other
-            reasoning-by-default providers reason regardless. Re-run with <code>tool_choice: auto</code>,
-            Claude Fable 5 scores {SENSITIVITY_EXACT["claude-fable-5"]} (would
-            rank #{wouldRank(SENSITIVITY_EXACT["claude-fable-5"], baseNoTools)}
-            ), Claude Opus 5 {SENSITIVITY_EXACT["claude-opus-5"]} (#
-            {wouldRank(SENSITIVITY_EXACT["claude-opus-5"], baseNoTools)}), and
-            Claude Sonnet 5 {SENSITIVITY_EXACT["claude-sonnet-5"]} (#
-            {wouldRank(SENSITIVITY_EXACT["claude-sonnet-5"], baseNoTools)}).
-            Claude Fable 5.1 rejects forced tool calls outright, so its board
-            row answers as a JSON object and reasons at the provider default;
-            with the tool declared under <code>auto</code> it scores{" "}
-            {SENSITIVITY_EXACT["claude-fable-5.1"]} (#
-            {wouldRank(SENSITIVITY_EXACT["claude-fable-5.1"], baseNoTools)}).
-            The board below is unchanged — those runs sit beside it as a{" "}
-            <a
-              href="https://github.com/PolicyEngine/policybench/blob/main/sensitivity/claude-thinking-2026-08.md"
-              className="text-primary hover:underline"
-            >
-              labeled sensitivity
+            Three Claude rows ran without extended thinking: this board forces
+            the answer tool call for every row whose model card selects the
+            tool contract, which switches Claude&apos;s thinking off. Claude
+            Fable 5.1 rejects forced calls, so its row selects JSON, answers as
+            a JSON object and reasons either way; its marker compares
+            transports. The{" "}
+            <code>tool_choice: auto</code> re-runs are marked on the four rows
+            (open a marker for the score and why); the board itself is
+            unchanged. The{" "}
+            <a href={SENSITIVITY_DOC_HREF} className="text-primary hover:underline">
+              sensitivity note
             </a>{" "}
-            — and the{" "}
-            <a
-              href="https://github.com/PolicyEngine/policybench/issues/139"
-              className="text-primary hover:underline"
-            >
+            has all four runs, and the{" "}
+            <a href={NEXT_BOARD_HREF} className="text-primary hover:underline">
               next board version
             </a>{" "}
             moves every model to auto.
@@ -432,6 +422,17 @@ export default function ModelLeaderboard({
                       >
                         {MODEL_LABELS[m.model] || m.model}
                       </Link>
+                      {showSensitivity && servingSensitivityFor(m.model) ? (
+                        <ServingSensitivityChip
+                          modelLabel={MODEL_LABELS[m.model] || m.model}
+                          boardExact={unfilteredExact(m.model)}
+                          sensitivity={servingSensitivityFor(m.model)!}
+                          wouldRank={wouldRank(
+                            servingSensitivityFor(m.model)!.autoExact,
+                            baseNoTools,
+                          )}
+                        />
+                      ) : null}
                     </div>
                     <div className="mt-1.5 pl-[26px] font-[family-name:var(--font-mono)] text-[11px] text-text-muted">
                       {fmtCost(m.costPerHousehold, currencySymbol)} per
@@ -470,6 +471,17 @@ export default function ModelLeaderboard({
                   >
                     {MODEL_LABELS[m.model] || m.model}
                   </Link>
+                  {showSensitivity && servingSensitivityFor(m.model) ? (
+                    <ServingSensitivityChip
+                      modelLabel={MODEL_LABELS[m.model] || m.model}
+                      boardExact={unfilteredExact(m.model)}
+                      sensitivity={servingSensitivityFor(m.model)!}
+                      wouldRank={wouldRank(
+                        servingSensitivityFor(m.model)!.autoExact,
+                        baseNoTools,
+                      )}
+                    />
+                  ) : null}
                 </div>
 
                 <div className="col-span-2 text-right font-[family-name:var(--font-mono)] text-sm text-text-muted">