diff --git a/packages/app/src/components/evaluation/RecipeComparison.tsx b/packages/app/src/components/evaluation/RecipeComparison.tsx
index dd961a9d3..6d0f13a1d 100644
--- a/packages/app/src/components/evaluation/RecipeComparison.tsx
+++ b/packages/app/src/components/evaluation/RecipeComparison.tsx
@@ -143,14 +143,15 @@ const COLUMNS: DataTableColumn<RecipeRow>[] = [
 ];
 
 export default function RecipeComparison() {
-  const { selectedModel, selectedRunDate } = useGlobalFilters();
+  const { selectedModel } = useGlobalFilters();
   const [category, setCategory] = useState<TechniqueCategory | 'all'>('all');
 
-  const {
-    data: benchmarks,
-    isLoading: bmkLoading,
-    error: bmkError,
-  } = useBenchmarks(selectedModel, selectedRunDate || undefined);
+  // Always use the no-date "latest state across all dates" path. The recipe
+  // table is meant to surface current recipe variants, not historical data —
+  // a stale `selectedRunDate` inherited from /inference would hide rows
+  // ingested on more recent dates (e.g. the user reported missing the
+  // batch-size filter chip because mnbt rows landed after the sticky date).
+  const { data: benchmarks, isLoading: bmkLoading, error: bmkError } = useBenchmarks(selectedModel);
   const { data: evals, isLoading: evalLoading, error: evalError } = useEvaluations();
 
   const allRows = useMemo(() => {
@@ -231,14 +232,7 @@ export default function RecipeComparison() {
 
       {!loading && rows.length === 0 && (
         <div className="rounded-md border border-border bg-card p-6 text-sm text-muted-foreground">
-          No data for <strong>{selectedModel || '(no model)'}</strong>
-          {selectedRunDate ? (
-            <>
-              {' '}
-              on <strong>{selectedRunDate}</strong>
-            </>
-          ) : null}
-          . Pick a different model or date.
+          No data for <strong>{selectedModel || '(no model)'}</strong>. Pick a different model.
         </div>
       )}
 
diff --git a/packages/app/src/lib/chart-utils.ts b/packages/app/src/lib/chart-utils.ts
index 554bb7a5b..8f761da56 100644
--- a/packages/app/src/lib/chart-utils.ts
+++ b/packages/app/src/lib/chart-utils.ts
@@ -174,6 +174,16 @@ export const getHardwareKey = (entry: AggDataEntry): string => {
   } else if (entry['spec_decoding'] && entry['spec_decoding'] !== 'none') {
     normalizedHwName = `${normalizedHwName}_${entry['spec_decoding']}`;
   }
+  // Topology suffix when there's more than one GPU in the deployment, so that
+  // 1× and 2× (or larger) deployments don't collapse into a single chart
+  // series. Suppressed for single-GPU runs to keep parity with upstream
+  // legends. Disagg = prefill + decode pools; non-disagg = single pool.
+  const totalGpus = entry.disagg
+    ? (entry.num_prefill_gpu ?? 0) + (entry.num_decode_gpu ?? 0)
+    : (entry.num_prefill_gpu ?? entry.num_decode_gpu ?? 0);
+  if (totalGpus > 1) {
+    normalizedHwName = `${normalizedHwName}_${totalGpus}x`;
+  }
   return normalizedHwName;
 };
 
@@ -224,6 +234,7 @@ export function buildAvailabilityHwKey(
   framework?: string,
   specMethod?: string,
   disagg?: boolean,
+  totalGpus?: number,
 ): string {
   let hwKey = hardware.split('-')[0];
   const fw = framework ? resolveFrameworkAlias(framework) : undefined;
@@ -241,6 +252,11 @@ export function buildAvailabilityHwKey(
   }
   if (specMethod === 'mtp') hwKey = `${hwKey}_mtp`;
   else if (specMethod && specMethod !== 'none') hwKey = `${hwKey}_${specMethod}`;
+  // Optional topology suffix — keeps 1× and 2× deployments as separate
+  // legend entries. Mirror of the suffix in getHardwareKey above. Callers
+  // that don't know the topology (e.g. legacy availability rows lacking GPU
+  // counts) pass undefined and the suffix is skipped.
+  if (totalGpus !== undefined && totalGpus > 1) hwKey = `${hwKey}_${totalGpus}x`;
   return hwKey;
 }