Add six September 2026 models (GPT-6 Astra #2 of 39), exclude eleven ambiguous outputs from scoring, publish dashboard-data-20260905c

app/src/app/paper/page.tsx

@@ -4,10 +4,10 @@ import Link from "next/link";
 
 import SiteHeader from "../../components/SiteHeader";
 
-const SNAPSHOT_DATE_LABEL = "Snapshot 2026-09-01";
+const SNAPSHOT_DATE_LABEL = "Snapshot 2026-09-05";
 
 const PAPER_DESCRIPTION =
-  "PolicyBench paper: the 2026-09-01 manuscript snapshot reporting the household-impact-weighted exact-match rate, with model responses collected from June 12 through September 1 and PolicyEngine reference outputs on the US populace microdata.";
+  "PolicyBench paper: the 2026-09-05 manuscript snapshot reporting the household-impact-weighted exact-match rate, with model responses collected from June 12 through September 1 and PolicyEngine reference outputs on the US populace microdata.";
 
 export const metadata: Metadata = {
   title: "Paper",
@@ -49,7 +49,7 @@ export default function PaperPage() {
     <>
       <p className="max-w-2xl text-sm leading-relaxed text-text-secondary sm:text-base">
         Benchmarking no-tool tax-and-benefit estimation in frontier language
-        models. This page embeds the 2026-09-01 scored manuscript snapshot: a
+        models. This page embeds the 2026-09-05 scored manuscript snapshot: a
         100-household public preview reporting the household-impact-weighted
         exact-match rate against PolicyEngine reference outputs.
       </p>

app/src/components/ExclusionNote.tsx

@@ -0,0 +1,84 @@
+import type { ReferenceExclusion, ScenarioPrediction } from "../types";
+import { formatCurrency } from "../format";
+import { binaryFlag } from "../lib/scoring";
+import { describeExclusionReason } from "../lib/predictionStatus";
+
+function formatValue(
+  value: number,
+  isBinary: boolean,
+  currencySymbol: "$" | "£",
+): string {
+  if (isBinary) {
+    const flag = binaryFlag(value);
+    return flag === null ? "Invalid" : flag === 1 ? "Yes" : "No";
+  }
+  return formatCurrency(value, currencySymbol);
+}
+
+/**
+ * Why an output contributes to no model's score. Rendered in the prediction
+ * detail dialog for rows the payload marks `scored: false`; the release's
+ * exclusion record, when present, supplies the alternative reading and the
+ * reference under it.
+ */
+export default function ExclusionNote({
+  pred,
+  exclusion,
+  isBinary,
+  currencySymbol,
+}: {
+  pred: ScenarioPrediction;
+  exclusion?: ReferenceExclusion;
+  isBinary: boolean;
+  currencySymbol: "$" | "£";
+}) {
+  const unlistedInput = exclusion?.unlistedInput ?? pred.excludedInput;
+  const reason = describeExclusionReason(
+    exclusion?.reasonCode ?? pred.excludedReason,
+  );
+  return (
+    <section
+      className="mt-4 rounded-lg border border-border-subtle bg-surface px-4 py-3"
+      data-testid="exclusion-note"
+    >
+      <div className="text-[10px] uppercase tracking-[0.14em] text-text-muted font-medium">
+        Excluded from scoring
+      </div>
+      <p className="mt-2 text-sm text-text-secondary leading-relaxed">
+        This output is scored for no model: {reason}
+        {unlistedInput ? (
+          <>
+            {" "}
+            (
+            <code className="font-[family-name:var(--font-mono)] text-[12px]">
+              {unlistedInput}
+            </code>
+            )
+          </>
+        ) : null}
+        . The prediction above is shown for inspection and counts neither for
+        nor against this model.
+      </p>
+      {exclusion ? (
+        <>
+          <p className="mt-2 text-sm text-text-secondary leading-relaxed">
+            {exclusion.alternativeReading} Under that reading the reference is{" "}
+            <span className="font-[family-name:var(--font-mono)]">
+              {formatValue(exclusion.alternativeValue, isBinary, currencySymbol)}
+            </span>{" "}
+            rather than the frozen{" "}
+            <span className="font-[family-name:var(--font-mono)]">
+              {formatValue(exclusion.frozenValue, isBinary, currencySymbol)}
+            </span>
+            .
+          </p>
+          <p className="mt-2 text-xs text-text-muted leading-relaxed">
+            Both values recomputed with {exclusion.engineVersion}; excluded by
+            the developers on {exclusion.decidedOn}.
+            {exclusion.note ? ` ${exclusion.note}` : ""}
+          </p>
+        </>
+      ) : null}
+    </section>
+  );
+}

app/src/components/Methodology.tsx

@@ -162,7 +162,7 @@ export default function Methodology({
           selects JSON for the family (the older Gemini and DeepSeek rows).{" "}
           {currentBoard ? (
             <>
-              Ten of the 33 models answer the same facts in subsets of one or
+              Ten of the 39 models answer the same facts in subsets of one or
               three outputs per request, an accommodation that predates the
               whole-scenario rule. The per-model transport and request shape are
               recorded in the paper&apos;s serving-configuration table and the

app/src/components/ModelLeaderboard.tsx

@@ -20,13 +20,15 @@ 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). Their "would rank" positions are
-// derived from the live board rows at render time, never typed by hand.
+// (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": 86.9,
-  "claude-opus-5": 85.6,
-  "claude-sonnet-5": 80.2,
-  "claude-fable-5.1": 87.5,
+  "claude-fable-5": 87.5,
+  "claude-opus-5": 86.2,
+  "claude-sonnet-5": 80.8,
+  "claude-fable-5.1": 88.2,
 } as const;
 import {
   rankWithFallbackScore,