Add plain-language captions and mode-aware framing for policy readers

Eval card (/evals listing):
Under-title description in policy mode now prefers the
Auto-BenchmarkCards `goal` field over the technical `overview`.
Research mode keeps the overview-first ordering.

Model card (/models listing):
- Reproducibility-gap badge: research -> "N reproducibility gaps",
policy -> "N re-run gaps", with a tooltip that explains what the
gap means in plain language for each audience.
- "Score span" -> "Score range" in policy mode.
- "Architecture" row hidden in policy mode (technical detail not
useful to non-experts).
- "Artifact type" -> "Source type" in policy mode.
- "Re-runnability" row uses "Re-run readiness" + plain-language
description in policy mode.

Addresses UXR items:
- plain-language descriptions beneath each benchmark name (Slavina,
Ryan, Aviya)
- avoid surfacing long technical names in policy view (Aviya)
- clarify what reproducibility gaps mean for non-experts

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

components/benchmark-evaluation-card.tsx

@@ -305,9 +305,18 @@ export function BenchmarkEvaluationCard({
               <Badge variant="outline">{data.benchmarks_count} benchmark suites</Badge>
               <Badge variant="outline">{data.evaluations_count} reported results</Badge>
               {reproducibilityGapCount > 0 && (
-                <Badge className="border-amber-300 bg-amber-50 text-amber-900 hover:bg-amber-50 dark:border-amber-900/60 dark:bg-amber-950/40 dark:text-amber-100">
+                <Badge
+                  className="border-amber-300 bg-amber-50 text-amber-900 hover:bg-amber-50 dark:border-amber-900/60 dark:bg-amber-950/40 dark:text-amber-100"
+                  title={
+                    isResearchView
+                      ? `${reproducibilityGapCount} of ${reproducibilityTotal} reported scores have at least one missing setup field.`
+                      : `${reproducibilityGapCount} of ${reproducibilityTotal} reported scores cannot be independently re-run because the setup is not documented.`
+                  }
+                >
                   <AlertTriangle className="h-3 w-3" />
-                  {reproducibilityGapCount} setup gaps
+                  {isResearchView
+                    ? `${reproducibilityGapCount} reproducibility gaps`
+                    : `${reproducibilityGapCount} re-run gaps`}
                 </Badge>
               )}
             </div>
@@ -454,16 +463,27 @@ export function BenchmarkEvaluationCard({
               {topBenchmarks.length > 0 && (
                 <KeyValueRow label="Benchmarks" value={topBenchmarks.slice(0, 6).join(", ")} />
               )}
-              {scoreRange && <KeyValueRow label="Score span" value={scoreRange} />}
+              {scoreRange && (
+                <KeyValueRow label={isResearchView ? "Score span" : "Score range"} value={scoreRange} />
+              )}
               <KeyValueRow label="Updated" value={formatDate(data.latest_timestamp)} />
-              {data.architecture && <KeyValueRow label="Architecture" value={data.architecture} />}
+              {data.architecture && isResearchView && (
+                <KeyValueRow label="Architecture" value={data.architecture} />
+              )}
               {data.source_types.length > 0 && (
-                <KeyValueRow label="Artifact type" value={data.source_types.map((s) => s.replace(/_/g, " ")).join(", ")} />
+                <KeyValueRow
+                  label={isResearchView ? "Artifact type" : "Source type"}
+                  value={data.source_types.map((s) => s.replace(/_/g, " ")).join(", ")}
+                />
               )}
               {reproducibilityGapCount > 0 && (
                 <KeyValueRow
-                  label="Re-runnability"
-                  value={`${reproducibilityGapCount} of ${reproducibilityTotal} reported scores are not fully documented`}
+                  label={isResearchView ? "Re-runnability" : "Re-run readiness"}
+                  value={
+                    isResearchView
+                      ? `${reproducibilityGapCount} of ${reproducibilityTotal} reported scores are not fully documented`
+                      : `${reproducibilityGapCount} of ${reproducibilityTotal} scores cannot be re-run with the information available`
+                  }
                 />
               )}
             </div>

components/eval-card.tsx

@@ -123,7 +123,15 @@ export function EvalCard({ summary, delayMs = 0 }: EvalCardProps) {
             Suite: {summary.composite_benchmark_name}
           </div>
           <div className="mt-1 text-sm text-muted-foreground line-clamp-2">
-            {overviewText ?? summary.metric_config.evaluation_description}
+            {/*
+             * Policy readers benefit from a plain-language framing of what
+             * the benchmark is for. The Auto-BenchmarkCards `goal` field is
+             * usually written for non-experts; the `overview` field is more
+             * technical. Prefer goal in policy mode, overview in research.
+             */}
+            {isResearchView
+              ? overviewText ?? policyGoal ?? summary.metric_config.evaluation_description
+              : policyGoal ?? overviewText ?? summary.metric_config.evaluation_description}
           </div>
         </div>