Separate policy and researcher views

- Foundations: glossary + Term + ModeSwitch helpers
- Policy: plain-language summary, scope chips, provenance row,
collapsible subtasks, technical details pushed into expander
- Researcher: per-row Reproducibility card with adaptive compact mode
for limited-disclosure rows; agent setup group for agentic evals;
lazy-fetched generation_config + score_details from model files
- Replace boilerplate IBM risk field with curated known-issues registry
- Apples-to-apples banner above leaderboards, anchored to
Comparability panel
- Methodology split: surface task setup, scoring, validation
- Compact row-signals indicator replacing repetitive coloured badges
- Cleaner divergence chips in Comparability panel; full-width when
only one divergence type fires
- Flag-this-score button that links to the dataset record on HF and
pre-fills a discussion title plus copyable context
- Defensive guard so an embedded ancestor card no longer leaks its
text onto a leaf-benchmark page

app/api/eval-row-config/route.ts

@@ -0,0 +1,65 @@
+import { NextResponse } from "next/server"
+
+import { fetchModelDetail } from "@/lib/hf-data"
+
+/**
+ * Lightweight lookup that returns the per-(model, benchmark) reproducibility
+ * payload — generation_config + sample_size/standard_error/confidence_interval
+ * — extracted from the model's full record. Used by the leaderboard's
+ * Reproducibility card on row expand so we avoid joining N model files into
+ * the eval-detail page on every load.
+ */
+export async function GET(request: Request) {
+  const { searchParams } = new URL(request.url)
+  const modelId = searchParams.get("model_id")
+  const benchmarkKey = (searchParams.get("benchmark_key") || "").toLowerCase()
+  const evalName = (searchParams.get("eval_name") || "").toLowerCase()
+
+  if (!modelId) {
+    return NextResponse.json({ error: "Missing model_id" }, { status: 400 })
+  }
+
+  const slug = modelId.replace(/[\/]/g, "__")
+  const detail = await fetchModelDetail(slug)
+  if (!detail) {
+    return NextResponse.json({ generation_config: null, score_details: null }, { status: 200 })
+  }
+
+  let bestGenerationConfig: unknown = null
+  let bestScoreDetails: unknown = null
+
+  const candidates = Object.values(detail.evaluations_by_category ?? {}).flat()
+  for (const ev of candidates) {
+    const evBench = ((ev.benchmark as string | undefined) ?? "").toLowerCase()
+    for (const r of ev.evaluation_results ?? []) {
+      const rName = (r.evaluation_name ?? "").toLowerCase()
+      const rDisplay = (r.display_name ?? "").toLowerCase()
+      const matches =
+        (benchmarkKey && (evBench === benchmarkKey || evBench.includes(benchmarkKey) || rName.includes(benchmarkKey))) ||
+        (evalName && (rName === evalName || rDisplay === evalName))
+      if (!matches) continue
+
+      const candidateGen = ev.generation_config ?? r.generation_config
+      const argsCount =
+        candidateGen && typeof candidateGen === "object" && "generation_args" in candidateGen
+          ? Object.keys((candidateGen as { generation_args?: Record<string, unknown> }).generation_args ?? {}).length
+          : 0
+      const currentCount =
+        bestGenerationConfig && typeof bestGenerationConfig === "object" && "generation_args" in bestGenerationConfig
+          ? Object.keys((bestGenerationConfig as { generation_args?: Record<string, unknown> }).generation_args ?? {}).length
+          : 0
+      if (candidateGen && argsCount > currentCount) {
+        bestGenerationConfig = candidateGen
+      }
+      if (!bestScoreDetails && r.score_details) {
+        bestScoreDetails = r.score_details
+      }
+    }
+    if (bestGenerationConfig && bestScoreDetails) break
+  }
+
+  return NextResponse.json({
+    generation_config: bestGenerationConfig,
+    score_details: bestScoreDetails,
+  })
+}

components/apples-to-apples-banner.tsx

@@ -0,0 +1,98 @@
+"use client"
+
+import { AlertTriangle, GitCompareArrows, UsersRound } from "lucide-react"
+import { useAudienceMode } from "@/components/audience-mode-provider"
+import { SignalTooltip } from "@/components/signals/signal-tooltip"
+import type { ComparabilitySummary } from "@/lib/backend-artifacts"
+
+interface ApplesToApplesBannerProps {
+  summary?: ComparabilitySummary | null
+  /**
+   * Optional anchor id so the "see details" link can scroll to the full
+   * Comparability panel further down the page.
+   */
+  detailsAnchorId?: string
+}
+
+/**
+ * High-visibility warning shown above the leaderboard when the row-level
+ * comparability summary indicates that some scores were collected under
+ * different setups (variant divergence) or only by the model developer
+ * (cross-party divergence). Designed to interrupt naive ranking comparisons
+ * before the reader scrolls to the per-row signals.
+ */
+export function ApplesToApplesBanner({ summary, detailsAnchorId }: ApplesToApplesBannerProps) {
+  const { mode } = useAudienceMode()
+
+  if (!summary) return null
+
+  const variantCount = summary.variant_divergent_count ?? 0
+  const crossPartyCount = summary.cross_party_divergent_count ?? 0
+  const totalConcerns = variantCount + crossPartyCount
+  if (totalConcerns === 0) return null
+
+  const variantsChecked = summary.groups_with_variant_check ?? 0
+  const crossPartyChecked = summary.groups_with_cross_party_check ?? 0
+
+  const concernPhrases: string[] = []
+  if (variantCount > 0) {
+    concernPhrases.push(
+      `${variantCount} group${variantCount === 1 ? "" : "s"} where models used different setups`,
+    )
+  }
+  if (crossPartyCount > 0) {
+    concernPhrases.push(
+      `${crossPartyCount} group${crossPartyCount === 1 ? "" : "s"} where reports come only from the model developer`,
+    )
+  }
+
+  const headline =
+    mode === "policy"
+      ? "Heads up: not every score here is directly comparable."
+      : "Apples-to-apples warning"
+
+  const body =
+    mode === "policy"
+      ? `${concernPhrases.join(" and ")}. Direct ranking comparisons may be misleading.`
+      : `${concernPhrases.join("; ")}. See the Comparability panel below for which models and which fields differ.`
+
+  return (
+    <div className="flex items-start gap-3 rounded-2xl border border-amber-300/70 bg-amber-50/70 px-4 py-3 text-sm text-amber-900 dark:border-amber-900/50 dark:bg-amber-950/25 dark:text-amber-100">
+      <AlertTriangle className="mt-0.5 h-4 w-4 shrink-0" />
+      <div className="min-w-0 flex-1 space-y-1">
+        <div className="font-semibold leading-tight">{headline}</div>
+        <p className="text-[13px] leading-5">{body}</p>
+        <div className="flex flex-wrap items-center gap-x-4 gap-y-1 text-[11px] opacity-90">
+          {variantCount > 0 && (
+            <SignalTooltip
+              content={`${variantCount} of ${variantsChecked} groups checked for setup divergence flagged a problem (e.g. different shots, prompts, or scoring).`}
+            >
+              <span className="inline-flex items-center gap-1 cursor-help">
+                <GitCompareArrows className="h-3 w-3" />
+                Setup divergence: {variantCount}/{variantsChecked || variantCount}
+              </span>
+            </SignalTooltip>
+          )}
+          {crossPartyCount > 0 && (
+            <SignalTooltip
+              content={`${crossPartyCount} of ${crossPartyChecked} groups checked for source divergence flagged a problem (only the model developer reported a number).`}
+            >
+              <span className="inline-flex items-center gap-1 cursor-help">
+                <UsersRound className="h-3 w-3" />
+                Source divergence: {crossPartyCount}/{crossPartyChecked || crossPartyCount}
+              </span>
+            </SignalTooltip>
+          )}
+          {detailsAnchorId && (
+            <a
+              href={`#${detailsAnchorId}`}
+              className="ml-auto font-medium underline-offset-4 hover:underline"
+            >
+              See details ↓
+            </a>
+          )}
+        </div>
+      </div>
+    </div>
+  )
+}

components/eval-detail.tsx

@@ -9,6 +9,7 @@ import { CompletenessPanel } from "@/components/signals/completeness-panel"
 import { ComparabilityPanel } from "@/components/signals/comparability-panel"
 import { ReproducibilityPanel } from "@/components/signals/reproducibility-panel"
 import { SignalsRowBadges } from "@/components/signals/signals-row-badges"
+import { RowSignalsCompact } from "@/components/signals/row-signals-compact"
 import { SignalTooltip } from "@/components/signals/signal-tooltip"
 import { getCompletenessPopulatedCount } from "@/components/signals/signal-utils"
 import { Card, CardContent, CardDescription, CardHeader, CardTitle } from "@/components/ui/card"
@@ -54,6 +55,12 @@ import {
 } from "lucide-react"
 import type { BenchmarkCard } from "@/lib/benchmark-schema"
 import type { BenchmarkEvalSummary, ModelResultForBenchmark } from "@/lib/eval-processing"
+import { PolicyOverview } from "@/components/policy-overview"
+import { ResearcherReproducibilityCard } from "@/components/researcher-reproducibility-card"
+import { KnownIssuesPanel } from "@/components/known-issues-panel"
+import { getKnownIssues, type KnownIssue } from "@/lib/known-issues"
+import { ApplesToApplesBanner } from "@/components/apples-to-apples-banner"
+import { FlagScoreButton } from "@/components/flag-score-button"
 
 interface EvalDetailProps {
   summary: BenchmarkEvalSummary
@@ -523,6 +530,11 @@ export function EvalDetail({ summary }: EvalDetailProps) {
     (summary.leaderboard_metrics?.length ?? 0) > 1 &&
     (summary.leaderboard_rows?.length ?? 0) > 0
   const [overviewOpen, setOverviewOpen] = useState(true)
+  // Collapse the dense technical overview by default in policy mode; expand
+  // for researchers. Reset whenever the user switches modes.
+  useEffect(() => {
+    setOverviewOpen(isResearchView)
+  }, [isResearchView])
   const [expandedRows, setExpandedRows] = useState<Record<string, boolean>>({})
   const [leaderboardPage, setLeaderboardPage] = useState(1)
   const [minParamStep, setMinParamStep] = useState(0)
@@ -636,6 +648,7 @@ export function EvalDetail({ summary }: EvalDetailProps) {
 
   return (
     <div className="space-y-6">
+      {!isResearchView && <PolicyOverview summary={summary} />}
       <Card className="overflow-hidden">
         <Collapsible open={overviewOpen} onOpenChange={setOverviewOpen}>
           <CollapsibleTrigger asChild>
@@ -645,7 +658,7 @@ export function EvalDetail({ summary }: EvalDetailProps) {
             >
               <div className="min-w-0 space-y-1">
                 <div className="text-[11px] font-semibold uppercase tracking-[0.2em] text-muted-foreground">
-                  {isResearchView ? "Benchmark overview" : "Reading overview"}
+                  {isResearchView ? "Benchmark overview" : "Technical details"}
                 </div>
                 <div className="flex flex-wrap items-center gap-2">
                   <span className="text-base font-semibold tracking-tight sm:text-lg">{summary.evaluation_name}</span>
@@ -710,11 +723,6 @@ export function EvalDetail({ summary }: EvalDetailProps) {
                     {summary.metric_config.evaluation_description}
                   </p>
 
-                  {!isResearchView && (
-                    <p className="max-w-3xl text-sm leading-6 text-muted-foreground">
-                      {`${summary.benchmark_card?.purpose_and_intended_users?.goal ?? "This benchmark reports a capability result."} Scores should be read alongside benchmark scope, metric definitions, and the source dataset context.`}
-                    </p>
-                  )}
                 </div>
 
                 <div className="grid w-full grid-cols-2 gap-2 xl:grid-cols-4">
@@ -890,6 +898,14 @@ export function EvalDetail({ summary }: EvalDetailProps) {
                   isResearchView={isResearchView}
                   defaultOpen
                   defaultRisksOpen={!isResearchView}
+                  knownIssues={getKnownIssues(
+                    summary.evaluation_name,
+                    summary.composite_benchmark_name,
+                    summary.composite_benchmark_key,
+                    summary.benchmark_family_key,
+                    summary.benchmark_leaf_key,
+                    summary.benchmark_card.benchmark_details?.name,
+                  )}
                 />
               )}
             </CardContent>
@@ -901,7 +917,11 @@ export function EvalDetail({ summary }: EvalDetailProps) {
         <MultiMetricLeaderboard summary={summary} isResearchView={isResearchView} />
       ) : (
         <Card className="overflow-hidden">
-          <CardHeader className="border-b bg-muted/10">
+          <CardHeader className="border-b bg-muted/10 space-y-3">
+            <ApplesToApplesBanner
+              summary={summary.comparability_summary}
+              detailsAnchorId="comparability-panel"
+            />
             <div className="flex flex-col gap-3 lg:flex-row lg:items-end lg:justify-between">
               <div className="space-y-2">
                 <div className="flex items-center gap-2">
@@ -1034,7 +1054,6 @@ export function EvalDetail({ summary }: EvalDetailProps) {
                     {isResearchView ? "Evaluator" : "Reporting Org"}
                   </TableHead>
                   <TableHead className="hidden min-w-[120px] lg:table-cell">Updated</TableHead>
-                  <TableHead className="w-16 px-4 text-right">Details</TableHead>
                 </TableRow>
               </TableHeader>
               <TableBody>
@@ -1044,6 +1063,7 @@ export function EvalDetail({ summary }: EvalDetailProps) {
                     ? Object.entries(modelResult.score_details.details).filter(([, value]) => typeof value === "number")
                     : []
                   const hasExpandableDetails =
+                    isResearchView ||
                     (modelResult.aggregate_components && modelResult.aggregate_components.length > 1) ||
                     subtasks.length > 1
 
@@ -1075,7 +1095,18 @@ export function EvalDetail({ summary }: EvalDetailProps) {
 
                         <TableCell className="whitespace-normal">
                           <div className="space-y-1">
-                            <div className="font-semibold leading-tight">
+                            <div className="flex items-center gap-1.5 font-semibold leading-tight">
+                              {hasExpandableDetails && (
+                                <button
+                                  type="button"
+                                  onClick={() => toggleRow(key)}
+                                  aria-label={isExpanded ? "Collapse details" : "Expand details"}
+                                  aria-expanded={isExpanded}
+                                  className="-ml-1 inline-flex h-5 w-5 items-center justify-center rounded text-muted-foreground transition-colors hover:bg-muted hover:text-foreground"
+                                >
+                                  {isExpanded ? <ChevronUp className="h-3.5 w-3.5" /> : <ChevronDown className="h-3.5 w-3.5" />}
+                                </button>
+                              )}
                               <Link
                                 href={`/models/${getModelFamilyRouteId(modelResult.model_info)}`}
                                 className="underline decoration-dotted underline-offset-4 hover:text-primary"
@@ -1113,8 +1144,10 @@ export function EvalDetail({ summary }: EvalDetailProps) {
                         </TableCell>
 
                         <TableCell className="text-right">
-                          <div className="text-xl font-semibold tabular-nums">{formatRawScore(modelResult.score, summary.metric_config.unit)}</div>
-                          <SignalsRowBadges annotations={rowAnnotations} />
+                          <div className="flex items-center justify-end gap-1.5">
+                            <div className="text-xl font-semibold tabular-nums">{formatRawScore(modelResult.score, summary.metric_config.unit)}</div>
+                            <RowSignalsCompact annotations={rowAnnotations} />
+                          </div>
                         </TableCell>
 
                         {isResearchView ? (
@@ -1160,24 +1193,11 @@ export function EvalDetail({ summary }: EvalDetailProps) {
                         <TableCell className="hidden lg:table-cell">
                           <div className="text-sm text-muted-foreground">{formatDate(modelResult.evaluation_timestamp)}</div>
                         </TableCell>
-
-                        <TableCell className="px-4 text-right">
-                          {hasExpandableDetails && (
-                            <Button
-                              variant="ghost"
-                              size="icon"
-                              aria-label={isExpanded ? "Collapse details" : "Expand details"}
-                              onClick={() => toggleRow(key)}
-                            >
-                              {isExpanded ? <ChevronUp className="h-4 w-4" /> : <ChevronDown className="h-4 w-4" />}
-                            </Button>
-                          )}
-                        </TableCell>
                       </TableRow>
 
                       {isExpanded && (
                         <TableRow className="hover:bg-transparent">
-                          <TableCell colSpan={8} className="bg-muted/10 px-0 py-0">
+                          <TableCell colSpan={7} className="bg-muted/10 px-0 py-0">
                             <div className="space-y-5 px-4 py-5 sm:px-6">
                               <div className="grid gap-4 xl:grid-cols-3">
                                 <DetailPanel
@@ -1244,13 +1264,15 @@ export function EvalDetail({ summary }: EvalDetailProps) {
                                   )}
                                 </DetailPanel>
 
-                                <ReproducibilityPanel gap={rowAnnotations?.reproducibility_gap} />
+                                {!isResearchView && (
+                                  <ReproducibilityPanel gap={rowAnnotations?.reproducibility_gap} />
+                                )}
 
                                 <DetailPanel
                                   title={isResearchView ? "Score Breakdown" : "Metric Summary"}
                                   subtitle={
                                     isResearchView
-                                      ? "Raw metric values and uncertainty details."
+                                      ? "Raw score and scale."
                                       : "Raw performance plus uncertainty and sample details."
                                   }
                                 >
@@ -1260,19 +1282,23 @@ export function EvalDetail({ summary }: EvalDetailProps) {
                                   />
                                   <MetaRow label="Score Type" value={modelResult.result.metric_config.score_type} />
                                   <MetaRow label="Range" value={`${minScore} - ${maxScore}`} />
-                                  <MetaRow
-                                    label="Sample Size"
-                                    value={modelResult.score_details.sample_size ?? "Unknown"}
-                                  />
-                                  <MetaRow
-                                    label="Standard Error"
-                                    value={modelResult.score_details.standard_error ?? "Unknown"}
-                                  />
-                                  {modelResult.score_details.confidence_interval && (
-                                    <MetaRow
-                                      label="Confidence Interval"
-                                      value={`${modelResult.score_details.confidence_interval.lower} - ${modelResult.score_details.confidence_interval.upper} (${modelResult.score_details.confidence_interval.confidence_level}%)`}
-                                    />
+                                  {!isResearchView && (
+                                    <>
+                                      <MetaRow
+                                        label="Sample Size"
+                                        value={modelResult.score_details.sample_size ?? "Unknown"}
+                                      />
+                                      <MetaRow
+                                        label="Standard Error"
+                                        value={modelResult.score_details.standard_error ?? "Unknown"}
+                                      />
+                                      {modelResult.score_details.confidence_interval && (
+                                        <MetaRow
+                                          label="Confidence Interval"
+                                          value={`${modelResult.score_details.confidence_interval.lower} - ${modelResult.score_details.confidence_interval.upper} (${modelResult.score_details.confidence_interval.confidence_level}%)`}
+                                        />
+                                      )}
+                                    </>
                                   )}
                                 </DetailPanel>
                               </div>
@@ -1334,53 +1360,74 @@ export function EvalDetail({ summary }: EvalDetailProps) {
                                 </div>
                               )}
 
-                              {modelResult.result.generation_config && (
+                              {isResearchView ? (
                                 <div className="space-y-3">
-                                  <div>
-                                    <div className="text-xs font-semibold uppercase tracking-[0.2em] text-muted-foreground">
-                                      Generation Config
-                                    </div>
-                                    <div className="text-sm text-muted-foreground">
-                                      Evaluation-time generation parameters.
-                                    </div>
+                                  <ResearcherReproducibilityCard
+                                    modelResult={modelResult}
+                                    benchmarkKey={summary.benchmark_leaf_key ?? summary.composite_benchmark_key}
+                                    evalName={summary.evaluation_name}
+                                  />
+                                  <div className="flex justify-end">
+                                    <FlagScoreButton
+                                      modelName={modelResult.model_info.name}
+                                      modelId={modelResult.model_info.id}
+                                      benchmarkName={summary.evaluation_name}
+                                      benchmarkId={summary.evaluation_id}
+                                      score={formatRawScore(modelResult.score, summary.metric_config.unit)}
+                                      sourceUrl={modelResult.source_metadata.source_url}
+                                      sourceRecordUrl={modelResult.source_record_url}
+                                    />
                                   </div>
+                                </div>
+                              ) : (
+                                modelResult.result.generation_config && (
+                                  <div className="space-y-3">
+                                    <div>
+                                      <div className="text-xs font-semibold uppercase tracking-[0.2em] text-muted-foreground">
+                                        Generation Config
+                                      </div>
+                                      <div className="text-sm text-muted-foreground">
+                                        Evaluation-time generation parameters.
+                                      </div>
+                                    </div>
+
+                                    <div className="grid gap-3 md:grid-cols-2 xl:grid-cols-3">
+                                      {modelResult.result.generation_config.generation_args &&
+                                        Object.entries(modelResult.result.generation_config.generation_args).map(([key, value]) => (
+                                          <div key={key} className="rounded-xl border bg-background/70 p-4">
+                                            <div className="text-xs uppercase tracking-[0.16em] text-muted-foreground">
+                                              {key.replace(/_/g, " ")}
+                                            </div>
+                                            <div className="mt-2 text-sm font-medium">
+                                              {formatMetadataValue(value)}
+                                            </div>
+                                          </div>
+                                        ))}
 
-                                  <div className="grid gap-3 md:grid-cols-2 xl:grid-cols-3">
-                                    {modelResult.result.generation_config.generation_args &&
-                                      Object.entries(modelResult.result.generation_config.generation_args).map(([key, value]) => (
-                                        <div key={key} className="rounded-xl border bg-background/70 p-4">
+                                      {modelResult.result.generation_config.additional_details && (
+                                        <div className="rounded-xl border bg-background/70 p-4 md:col-span-2 xl:col-span-3">
                                           <div className="text-xs uppercase tracking-[0.16em] text-muted-foreground">
-                                            {key.replace(/_/g, " ")}
+                                            Additional Details
                                           </div>
-                                          <div className="mt-2 text-sm font-medium">
-                                            {formatMetadataValue(value)}
+                                          <div className="mt-2 text-sm font-medium whitespace-pre-wrap">
+                                            {formatMetadataValue(modelResult.result.generation_config.additional_details)}
                                           </div>
                                         </div>
-                                      ))}
+                                      )}
 
-                                    {modelResult.result.generation_config.additional_details && (
-                                      <div className="rounded-xl border bg-background/70 p-4 md:col-span-2 xl:col-span-3">
-                                        <div className="text-xs uppercase tracking-[0.16em] text-muted-foreground">
-                                          Additional Details
-                                        </div>
-                                        <div className="mt-2 text-sm font-medium whitespace-pre-wrap">
-                                          {formatMetadataValue(modelResult.result.generation_config.additional_details)}
-                                        </div>
-                                      </div>
-                                    )}
-
-                                    {modelResult.result.generation_config.prompt_template && (
-                                      <div className="rounded-xl border bg-background/70 p-4 md:col-span-2 xl:col-span-3">
-                                        <div className="text-xs uppercase tracking-[0.16em] text-muted-foreground">
-                                          Prompt Template
-                                        </div>
-                                        <div className="mt-2 text-sm font-medium whitespace-pre-wrap">
-                                          {formatMetadataValue(modelResult.result.generation_config.prompt_template)}
+                                      {modelResult.result.generation_config.prompt_template && (
+                                        <div className="rounded-xl border bg-background/70 p-4 md:col-span-2 xl:col-span-3">
+                                          <div className="text-xs uppercase tracking-[0.16em] text-muted-foreground">
+                                            Prompt Template
+                                          </div>
+                                          <div className="mt-2 text-sm font-medium whitespace-pre-wrap">
+                                            {formatMetadataValue(modelResult.result.generation_config.prompt_template)}
+                                          </div>
                                         </div>
-                                      </div>
-                                    )}
+                                      )}
+                                    </div>
                                   </div>
-                                </div>
+                                )
                               )}
                             </div>
                           </TableCell>
@@ -1391,7 +1438,7 @@ export function EvalDetail({ summary }: EvalDetailProps) {
                 })}
                 {leaderboardRows.length === 0 && (
                   <TableRow>
-                    <TableCell colSpan={8} className="px-6 py-12 text-center text-sm text-muted-foreground">
+                    <TableCell colSpan={7} className="px-6 py-12 text-center text-sm text-muted-foreground">
                       No leaderboard entries match the selected parameter range.
                     </TableCell>
                   </TableRow>
@@ -1429,6 +1476,32 @@ function MultiMetricLeaderboard({
   const [activeSubtaskTab, setActiveSubtaskTab] = useState<string>("all")
   const [minParamStep, setMinParamStep] = useState(0)
   const [maxParamStep, setMaxParamStep] = useState(PARAM_RANGE_VALUES.length - 1)
+  const [expandedRows, setExpandedRows] = useState<Record<string, boolean>>({})
+
+  // Index ModelResultForBenchmark entries by model_info.id so we can power the
+  // research-mode reproducibility card from a multi-metric row. There may be
+  // several entries per model (one per metric); we prefer one with a recorded
+  // generation_config so the card has the most data to show.
+  const modelResultByModelId = useMemo(() => {
+    const map = new Map<string, ModelResultForBenchmark>()
+    for (const result of summary.model_results) {
+      const id = result.model_info.id
+      const existing = map.get(id)
+      if (!existing) {
+        map.set(id, result)
+        continue
+      }
+      const existingHasGen = existing.result.generation_config != null
+      const candidateHasGen = result.result.generation_config != null
+      if (!existingHasGen && candidateHasGen) {
+        map.set(id, result)
+      }
+    }
+    return map
+  }, [summary.model_results])
+
+  const toggleExpandedRow = (key: string) =>
+    setExpandedRows((current) => ({ ...current, [key]: !current[key] }))
   const leaderboardMetrics = summary.leaderboard_metrics ?? []
   const leaderboardRows = summary.leaderboard_rows ?? []
   const allMetricKeys = useMemo(() => leaderboardMetrics.map((metric) => metric.column_key), [leaderboardMetrics])
@@ -1703,7 +1776,11 @@ function MultiMetricLeaderboard({
 
   return (
     <Card className="overflow-hidden">
-      <CardHeader className="border-b bg-muted/10">
+      <CardHeader className="border-b bg-muted/10 space-y-3">
+        <ApplesToApplesBanner
+          summary={summary.comparability_summary}
+          detailsAnchorId="comparability-panel"
+        />
         <div className="flex flex-col gap-3 lg:flex-row lg:items-end lg:justify-between">
           <div className="space-y-2">
             <div className="flex items-center gap-2">
@@ -1940,9 +2017,13 @@ function MultiMetricLeaderboard({
             <TableBody>
               {pagedRows.map((row) => {
                 const rank = rankByModelId.get(row.model_info.id) ?? 0
+                const expandKey = row.model_info.id
+                const isExpanded = expandedRows[expandKey] ?? false
+                const matchingResult = modelResultByModelId.get(row.model_info.id)
 
                 return (
-                <TableRow key={row.model_info.id} className="hover:bg-muted/10">
+                <Fragment key={row.model_info.id}>
+                <TableRow className={cn("hover:bg-muted/10", isExpanded && "bg-muted/15")}>
                   <TableCell className="px-4">
                     <div
                       className={cn(
@@ -1955,13 +2036,28 @@ function MultiMetricLeaderboard({
                   </TableCell>
                   <TableCell className="px-4 whitespace-normal">
                     <div className="space-y-1">
-                      <div className="font-semibold leading-tight">
+                      <div className="flex items-center gap-1.5 font-semibold leading-tight">
+                        {isResearchView && matchingResult && (
+                          <button
+                            type="button"
+                            onClick={() => toggleExpandedRow(expandKey)}
+                            aria-label={isExpanded ? "Hide reproducibility" : "Show reproducibility"}
+                            aria-expanded={isExpanded}
+                            className="-ml-1 inline-flex h-5 w-5 items-center justify-center rounded text-muted-foreground transition-colors hover:bg-muted hover:text-foreground"
+                          >
+                            {isExpanded ? <ChevronUp className="h-3.5 w-3.5" /> : <ChevronDown className="h-3.5 w-3.5" />}
+                          </button>
+                        )}
                         <Link
                           href={`/models/${getModelFamilyRouteId(row.model_info)}`}
                           className="underline decoration-dotted underline-offset-4 hover:text-primary"
                         >
                           {row.model_info.name}
                         </Link>
+                        <RowSignalsCompact
+                          annotations={getRowLevelAnnotations(row, visibleMetrics)}
+                          className="ml-1"
+                        />
                       </div>
                       <div className="flex flex-wrap items-center gap-2 text-xs text-muted-foreground">
                         {row.model_info.parameter_count && (
@@ -1976,12 +2072,6 @@ function MultiMetricLeaderboard({
                         )}
                         <span className="lg:hidden">{row.model_info.developer ?? "Unknown developer"}</span>
                       </div>
-                      <SignalsRowBadges
-                        annotations={getRowLevelAnnotations(row, visibleMetrics)}
-                        variant="row"
-                        className="mt-1 justify-start"
-                        hideOnMobile={false}
-                      />
                     </div>
                   </TableCell>
 
@@ -2016,6 +2106,34 @@ function MultiMetricLeaderboard({
                     {formatDate(row.evaluation_timestamp)}
                   </TableCell>
                 </TableRow>
+                {isResearchView && isExpanded && matchingResult && (
+                  <TableRow className="hover:bg-transparent">
+                    <TableCell
+                      colSpan={visibleMetrics.length + 5}
+                      className="bg-muted/10 px-4 py-5 sm:px-6"
+                    >
+                      <div className="space-y-3">
+                        <ResearcherReproducibilityCard
+                          modelResult={matchingResult}
+                          benchmarkKey={summary.benchmark_leaf_key ?? summary.composite_benchmark_key}
+                          evalName={summary.evaluation_name}
+                        />
+                        <div className="flex justify-end">
+                          <FlagScoreButton
+                            modelName={matchingResult.model_info.name}
+                            modelId={matchingResult.model_info.id}
+                            benchmarkName={summary.evaluation_name}
+                            benchmarkId={summary.evaluation_id}
+                            score={formatRawScore(matchingResult.score, summary.metric_config.unit)}
+                            sourceUrl={matchingResult.source_metadata.source_url}
+                            sourceRecordUrl={matchingResult.source_record_url}
+                          />
+                        </div>
+                      </div>
+                    </TableCell>
+                  </TableRow>
+                )}
+                </Fragment>
               )})}
 
               {filteredRows.length === 0 && (
@@ -2046,11 +2164,13 @@ function BenchmarkCardCollapsible({
   isResearchView,
   defaultOpen = true,
   defaultRisksOpen = false,
+  knownIssues = [],
 }: {
   card: BenchmarkCard
   isResearchView: boolean
   defaultOpen?: boolean
   defaultRisksOpen?: boolean
+  knownIssues?: KnownIssue[]
 }) {
   const [open, setOpen] = useState(defaultOpen)
   return (
@@ -2079,6 +2199,7 @@ function BenchmarkCardCollapsible({
           card={card}
           isResearchView={isResearchView}
           defaultRisksOpen={defaultRisksOpen}
+          knownIssues={knownIssues}
         />
       </CollapsibleContent>
     </Collapsible>
@@ -2095,12 +2216,12 @@ function DetailPanel({
   children: React.ReactNode
 }) {
   return (
-    <div className="rounded-2xl border bg-background/70 p-4">
+    <div className="min-w-0 rounded-2xl border bg-background/70 p-4">
       <div className="mb-4">
         <div className="font-semibold">{title}</div>
         <div className="text-sm text-muted-foreground">{subtitle}</div>
       </div>
-      <div className="space-y-2.5">{children}</div>
+      <div className="min-w-0 space-y-2.5">{children}</div>
     </div>
   )
 }
@@ -2129,9 +2250,28 @@ function MetaRow({
     }
   }
   return (
-    <div className="flex gap-3 text-sm">
-      <span className="w-32 shrink-0 text-muted-foreground">{label}</span>
-      <span className="min-w-0 flex-1 break-words font-medium">{value}</span>
+    <div
+      className="text-sm"
+      style={{
+        display: "grid",
+        gridTemplateColumns: "8rem minmax(0, 1fr)",
+        columnGap: "0.75rem",
+        width: "100%",
+        minWidth: 0,
+      }}
+    >
+      <div className="text-muted-foreground">{label}</div>
+      <div
+        className="font-medium"
+        style={{
+          minWidth: 0,
+          maxWidth: "100%",
+          overflowWrap: "anywhere",
+          wordBreak: "break-word",
+        }}
+      >
+        {value}
+      </div>
     </div>
   )
 }
@@ -2183,10 +2323,12 @@ function BenchmarkCardPanel({
   card,
   isResearchView,
   defaultRisksOpen = false,
+  knownIssues = [],
 }: {
   card: BenchmarkCard
   isResearchView: boolean
   defaultRisksOpen?: boolean
+  knownIssues?: KnownIssue[]
 }) {
   const [risksOpen, setRisksOpen] = useState(defaultRisksOpen)
   const details = card.benchmark_details
@@ -2236,6 +2378,8 @@ function BenchmarkCardPanel({
       </CardHeader>
 
       <CardContent className="space-y-6 p-5 sm:p-6">
+        {knownIssues.length > 0 && <KnownIssuesPanel issues={knownIssues} variant="full" />}
+
         {/* Overview + domains */}
         <div className="space-y-3">
           <p className="text-sm leading-6 text-muted-foreground">{details.overview}</p>
@@ -2288,6 +2432,52 @@ function BenchmarkCardPanel({
           </div>
         </div>
 
+        {(methodology.methods?.length > 0 ||
+          (methodology.calculation && methodology.calculation !== "Not specified") ||
+          (methodology.validation && methodology.validation !== "Not specified")) && (
+          <div className="rounded-[1.25rem] border border-border/70 bg-muted/10 p-4">
+            <div className="mb-3 text-[11px] font-semibold uppercase tracking-[0.2em] text-muted-foreground">
+              How tasks were sourced and scored
+            </div>
+            <div className="grid gap-4 lg:grid-cols-2">
+              {methodology.methods?.length > 0 && (
+                <div>
+                  <div className="mb-1.5 text-xs font-semibold text-foreground/80">
+                    Task setup
+                  </div>
+                  <ol className="list-decimal space-y-1.5 pl-4 text-sm leading-5 text-muted-foreground">
+                    {methodology.methods.map((m, i) => (
+                      <li key={i}>{m}</li>
+                    ))}
+                  </ol>
+                </div>
+              )}
+              <div className="space-y-3">
+                {methodology.calculation && methodology.calculation !== "Not specified" && (
+                  <div>
+                    <div className="mb-1.5 text-xs font-semibold text-foreground/80">
+                      Score calculation
+                    </div>
+                    <p className="text-sm leading-5 text-muted-foreground">
+                      {methodology.calculation}
+                    </p>
+                  </div>
+                )}
+                {methodology.validation && methodology.validation !== "Not specified" && (
+                  <div>
+                    <div className="mb-1.5 text-xs font-semibold text-foreground/80">
+                      Validation
+                    </div>
+                    <p className="text-sm leading-5 text-muted-foreground">
+                      {methodology.validation}
+                    </p>
+                  </div>
+                )}
+              </div>
+            </div>
+          </div>
+        )}
+
         {/* Research-only: methodology + dataset details */}
         {isResearchView && (
           <div className="grid gap-4 sm:grid-cols-2">
@@ -2339,8 +2529,12 @@ function BenchmarkCardPanel({
           </div>
         )}
 
-        {/* Risks (collapsible) */}
-        {risks.length > 0 && (
+        {/* Generic IBM-style AI risks. These are boilerplate (per audit
+            feedback: "least useful feature for policy users"), so in policy
+            mode we hide them entirely — the curated known-issues panel above
+            carries the benchmark-specific concerns. Researchers still get the
+            full collapsible list. */}
+        {risks.length > 0 && isResearchView && (
           <Collapsible open={risksOpen} onOpenChange={setRisksOpen}>
             <CollapsibleTrigger asChild>
               <button

components/flag-score-button.tsx

@@ -0,0 +1,267 @@
+"use client"
+
+import { useMemo, useState } from "react"
+import {
+  Copy,
+  ExternalLink,
+  FileSearch,
+  Flag,
+  GitPullRequestArrow,
+  MessageSquare,
+} from "lucide-react"
+import {
+  Dialog,
+  DialogContent,
+  DialogDescription,
+  DialogHeader,
+  DialogTitle,
+} from "@/components/ui/dialog"
+import { Button } from "@/components/ui/button"
+
+interface FlagScoreButtonProps {
+  modelName: string
+  modelId: string
+  benchmarkName: string
+  benchmarkId?: string
+  score: number | string
+  /** URL to the published source the score was extracted from (paper, blog, leaderboard). */
+  sourceUrl?: string
+  /** URL to the processed record JSON in the card_backend HF dataset. */
+  sourceRecordUrl?: string
+  /** Optional explicit upstream record URL in evaleval/EEE_datastore (raw source of truth). */
+  eeeRecordUrl?: string
+}
+
+interface DatasetLinks {
+  repoSlug: string
+  recordViewUrl: string | null
+  recordRawUrl: string | null
+  discussionsUrl: string
+  newDiscussionUrl: string
+}
+
+/**
+ * Given a /resolve/main/... HF dataset URL, derive the helpful sibling URLs
+ * for the dataset (file viewer, discussions list, new-discussion form). Returns
+ * null when the input doesn't look like a HF dataset URL.
+ */
+function deriveDatasetLinks(recordUrl: string | undefined, prefilledTitle: string): DatasetLinks | null {
+  if (!recordUrl) return null
+  const m = recordUrl.match(
+    /^https:\/\/huggingface\.co\/datasets\/([^/]+\/[^/]+)\/(?:resolve|raw|blob)\/[^/]+\/(.*)$/,
+  )
+  if (!m) return null
+  const repoSlug = m[1]
+  const path = m[2]
+  const datasetBase = `https://huggingface.co/datasets/${repoSlug}`
+  const recordViewUrl = `${datasetBase}/blob/main/${path}`
+  const recordRawUrl = `${datasetBase}/resolve/main/${path}`
+  const discussionsUrl = `${datasetBase}/discussions`
+  const newDiscussionUrl = `${datasetBase}/discussions/new?title=${encodeURIComponent(prefilledTitle)}`
+  return { repoSlug, recordViewUrl, recordRawUrl, discussionsUrl, newDiscussionUrl }
+}
+
+/**
+ * "Flag this score" — researcher affordance that, instead of capturing the
+ * report ourselves, sends the user directly to the upstream HF dataset where
+ * the data lives. They can then file a discussion or submit a correction PR
+ * against the actual record. We also offer a copyable context snippet so the
+ * issue body has all the relevant identifiers without needing to retype them.
+ */
+export function FlagScoreButton({
+  modelName,
+  modelId,
+  benchmarkName,
+  benchmarkId,
+  score,
+  sourceUrl,
+  sourceRecordUrl,
+  eeeRecordUrl,
+}: FlagScoreButtonProps) {
+  const [open, setOpen] = useState(false)
+  const [copied, setCopied] = useState<"context" | null>(null)
+
+  const prefilledTitle = `Possible issue: ${modelName} on ${benchmarkName} (score ${score})`
+
+  const cardBackendLinks = useMemo(
+    () => deriveDatasetLinks(sourceRecordUrl, prefilledTitle),
+    [sourceRecordUrl, prefilledTitle],
+  )
+  const eeeLinks = useMemo(
+    () => deriveDatasetLinks(eeeRecordUrl, prefilledTitle),
+    [eeeRecordUrl, prefilledTitle],
+  )
+
+  // Prefer the EEE upstream as the "correction venue" when known — that's
+  // where raw evaluation records live. Fall back to card_backend (the
+  // pipeline output) when EEE isn't directly addressable for this row.
+  const correctionLinks = eeeLinks ?? cardBackendLinks
+
+  const contextSnippet = [
+    `Model: ${modelName} (${modelId})`,
+    `Benchmark: ${benchmarkName}${benchmarkId ? ` (${benchmarkId})` : ""}`,
+    `Reported score: ${score}`,
+    sourceUrl ? `Original source URL: ${sourceUrl}` : null,
+    sourceRecordUrl ? `Pipeline record: ${sourceRecordUrl}` : null,
+    eeeRecordUrl ? `EEE upstream record: ${eeeRecordUrl}` : null,
+  ]
+    .filter(Boolean)
+    .join("\n")
+
+  const handleCopyContext = async () => {
+    try {
+      await navigator.clipboard.writeText(contextSnippet)
+      setCopied("context")
+      setTimeout(() => setCopied(null), 1800)
+    } catch {
+      // Clipboard might be blocked; ignore — the user can still select text manually.
+    }
+  }
+
+  return (
+    <>
+      <Button
+        type="button"
+        size="sm"
+        variant="outline"
+        className="h-7 gap-1.5 text-xs"
+        onClick={() => setOpen(true)}
+      >
+        <Flag className="h-3 w-3" />
+        Flag this score
+      </Button>
+
+      <Dialog open={open} onOpenChange={setOpen}>
+        <DialogContent className="sm:max-w-lg">
+          <DialogHeader>
+            <DialogTitle>Flag this score</DialogTitle>
+            <DialogDescription>
+              Take this report directly to the dataset where the record lives. You can file a
+              discussion or open a correction PR against the actual file.
+            </DialogDescription>
+          </DialogHeader>
+
+          <div className="space-y-4">
+            <div className="rounded-xl border bg-muted/20 p-3 text-xs">
+              <div className="text-[10px] font-semibold uppercase tracking-[0.16em] text-muted-foreground">
+                Flagging
+              </div>
+              <div className="mt-1">
+                <span className="font-medium">{modelName}</span>{" "}
+                <span className="text-muted-foreground">on</span>{" "}
+                <span className="font-medium">{benchmarkName}</span>
+              </div>
+              <div className="text-muted-foreground tabular-nums">Score: {score}</div>
+            </div>
+
+            <div className="space-y-2">
+              {correctionLinks?.recordViewUrl && (
+                <a
+                  href={correctionLinks.recordViewUrl}
+                  target="_blank"
+                  rel="noreferrer"
+                  className="flex items-center justify-between rounded-xl border border-border/70 bg-background px-3 py-2.5 text-sm transition-colors hover:border-primary/40 hover:bg-muted/20"
+                >
+                  <div className="flex items-center gap-2">
+                    <FileSearch className="h-4 w-4 text-muted-foreground" />
+                    <div>
+                      <div className="font-semibold">View the underlying record</div>
+                      <div className="text-xs text-muted-foreground">
+                        Opens the JSON file on{" "}
+                        <span className="font-mono">{correctionLinks.repoSlug}</span>
+                      </div>
+                    </div>
+                  </div>
+                  <ExternalLink className="h-3.5 w-3.5 text-muted-foreground" />
+                </a>
+              )}
+
+              {correctionLinks && (
+                <a
+                  href={correctionLinks.newDiscussionUrl}
+                  target="_blank"
+                  rel="noreferrer"
+                  className="flex items-center justify-between rounded-xl border border-border/70 bg-background px-3 py-2.5 text-sm transition-colors hover:border-primary/40 hover:bg-muted/20"
+                >
+                  <div className="flex items-center gap-2">
+                    <MessageSquare className="h-4 w-4 text-muted-foreground" />
+                    <div>
+                      <div className="font-semibold">Open a discussion</div>
+                      <div className="text-xs text-muted-foreground">
+                        Pre-filled title; paste the context snippet into the body.
+                      </div>
+                    </div>
+                  </div>
+                  <ExternalLink className="h-3.5 w-3.5 text-muted-foreground" />
+                </a>
+              )}
+
+              {correctionLinks && (
+                <a
+                  href={correctionLinks.discussionsUrl}
+                  target="_blank"
+                  rel="noreferrer"
+                  className="flex items-center justify-between rounded-xl border border-border/70 bg-background px-3 py-2.5 text-sm transition-colors hover:border-primary/40 hover:bg-muted/20"
+                >
+                  <div className="flex items-center gap-2">
+                    <GitPullRequestArrow className="h-4 w-4 text-muted-foreground" />
+                    <div>
+                      <div className="font-semibold">Browse existing discussions</div>
+                      <div className="text-xs text-muted-foreground">
+                        Check whether someone already filed a similar correction.
+                      </div>
+                    </div>
+                  </div>
+                  <ExternalLink className="h-3.5 w-3.5 text-muted-foreground" />
+                </a>
+              )}
+
+              {!correctionLinks && (
+                <div className="rounded-xl border border-dashed border-amber-300/60 bg-amber-50/40 px-3 py-2.5 text-xs text-amber-800 dark:border-amber-900/60 dark:bg-amber-950/20 dark:text-amber-200">
+                  No upstream record URL is recorded for this row, so we can't link directly to
+                  the dataset. Copy the context below and file an issue at{" "}
+                  <a
+                    className="underline-offset-4 hover:underline"
+                    href="https://huggingface.co/datasets/evaleval/EEE_datastore/discussions"
+                    target="_blank"
+                    rel="noreferrer"
+                  >
+                    evaleval/EEE_datastore/discussions
+                  </a>
+                  .
+                </div>
+              )}
+            </div>
+
+            <div className="space-y-1.5">
+              <div className="flex items-center justify-between">
+                <label className="text-xs font-semibold text-muted-foreground">
+                  Context to paste into the issue
+                </label>
+                <Button
+                  type="button"
+                  size="sm"
+                  variant="ghost"
+                  onClick={handleCopyContext}
+                  className="h-7 gap-1.5 text-xs"
+                >
+                  <Copy className="h-3 w-3" />
+                  {copied === "context" ? "Copied" : "Copy"}
+                </Button>
+              </div>
+              <pre className="max-h-40 overflow-auto whitespace-pre-wrap break-all rounded-xl border bg-muted/10 p-3 text-[11px] leading-5 text-muted-foreground">
+                {contextSnippet}
+              </pre>
+            </div>
+
+            <div className="flex justify-end pt-1">
+              <Button type="button" variant="ghost" onClick={() => setOpen(false)}>
+                Close
+              </Button>
+            </div>
+          </div>
+        </DialogContent>
+      </Dialog>
+    </>
+  )
+}

components/known-issues-panel.tsx

@@ -0,0 +1,108 @@
+"use client"
+
+import { AlertOctagon, AlertTriangle, ExternalLink, Info } from "lucide-react"
+import type { KnownIssue } from "@/lib/known-issues"
+
+interface KnownIssuesPanelProps {
+  issues: KnownIssue[]
+  /**
+   * "compact" — single-line summary chip suitable for surfacing at the top of
+   * the policy overview. "full" — full bordered list with descriptions.
+   */
+  variant?: "compact" | "full"
+}
+
+const SEVERITY_STYLE: Record<KnownIssue["severity"], { wrap: string; icon: React.ComponentType<{ className?: string }>; label: string }> = {
+  info: {
+    wrap: "border-sky-300/60 bg-sky-50/60 text-sky-900 dark:border-sky-900/50 dark:bg-sky-950/20 dark:text-sky-100",
+    icon: Info,
+    label: "Note",
+  },
+  warning: {
+    wrap: "border-amber-300/60 bg-amber-50/60 text-amber-900 dark:border-amber-900/50 dark:bg-amber-950/20 dark:text-amber-100",
+    icon: AlertTriangle,
+    label: "Known issue",
+  },
+  critical: {
+    wrap: "border-rose-300/60 bg-rose-50/60 text-rose-900 dark:border-rose-900/50 dark:bg-rose-950/20 dark:text-rose-100",
+    icon: AlertOctagon,
+    label: "Critical issue",
+  },
+}
+
+export function KnownIssuesPanel({ issues, variant = "full" }: KnownIssuesPanelProps) {
+  if (issues.length === 0) return null
+
+  const sorted = [...issues].sort((a, b) => severityRank(b.severity) - severityRank(a.severity))
+  const headlineSeverity = sorted[0].severity
+  const Style = SEVERITY_STYLE[headlineSeverity]
+  const Icon = Style.icon
+
+  if (variant === "compact") {
+    return (
+      <div className={`flex items-start gap-2 rounded-2xl border px-3 py-2 text-sm ${Style.wrap}`}>
+        <Icon className="mt-0.5 h-4 w-4 shrink-0" />
+        <div className="min-w-0">
+          <span className="font-semibold">
+            {issues.length} known issue{issues.length === 1 ? "" : "s"} documented
+          </span>
+          <span className="ml-1 text-muted-foreground">— see below for detail.</span>
+        </div>
+      </div>
+    )
+  }
+
+  return (
+    <section className="space-y-2">
+      <div className="text-[11px] font-semibold uppercase tracking-[0.2em] text-muted-foreground">
+        Known issues with this benchmark
+      </div>
+      <ul className="space-y-2">
+        {sorted.map((issue, idx) => {
+          const S = SEVERITY_STYLE[issue.severity]
+          const I = S.icon
+          return (
+            <li
+              key={`${issue.title}-${idx}`}
+              className={`rounded-2xl border px-3.5 py-3 ${S.wrap}`}
+            >
+              <div className="flex items-start gap-2">
+                <I className="mt-0.5 h-4 w-4 shrink-0" />
+                <div className="min-w-0 flex-1">
+                  <div className="flex flex-wrap items-baseline gap-2">
+                    <span className="text-[10px] font-semibold uppercase tracking-[0.18em] opacity-80">
+                      {S.label}
+                    </span>
+                    <span className="font-semibold leading-tight">{issue.title}</span>
+                  </div>
+                  <p className="mt-1 text-sm leading-5 opacity-90">{issue.summary}</p>
+                  {(issue.source_url || issue.published) && (
+                    <div className="mt-2 flex flex-wrap items-center gap-3 text-xs opacity-80">
+                      {issue.source_url && (
+                        <a
+                          href={issue.source_url}
+                          target="_blank"
+                          rel="noreferrer"
+                          className="inline-flex items-center gap-1 underline-offset-4 hover:underline"
+                        >
+                          Source <ExternalLink className="h-3 w-3" />
+                        </a>
+                      )}
+                      {issue.published && <span>Published {issue.published}</span>}
+                    </div>
+                  )}
+                </div>
+              </div>
+            </li>
+          )
+        })}
+      </ul>
+    </section>
+  )
+}
+
+function severityRank(s: KnownIssue["severity"]): number {
+  if (s === "critical") return 3
+  if (s === "warning") return 2
+  return 1
+}

components/mode-switch.tsx

@@ -0,0 +1,28 @@
+"use client"
+
+import type { ReactNode } from "react"
+import { useAudienceMode } from "@/components/audience-mode-provider"
+
+interface ModeSwitchProps {
+  research?: ReactNode
+  policy?: ReactNode
+}
+
+/**
+ * Renders different content per audience mode. Use this instead of inline
+ * `mode === "research"` checks so that call sites stay readable and the
+ * two branches stay obviously parallel.
+ */
+export function ModeSwitch({ research, policy }: ModeSwitchProps) {
+  const { mode } = useAudienceMode()
+  return <>{mode === "research" ? research ?? null : policy ?? null}</>
+}
+
+/**
+ * Hook variant for cases where ModeSwitch can't be used (e.g. choosing
+ * between two non-JSX values like a className or a number).
+ */
+export function useModeValue<T>(values: { research: T; policy: T }): T {
+  const { mode } = useAudienceMode()
+  return mode === "research" ? values.research : values.policy
+}

components/policy-overview.tsx

@@ -0,0 +1,350 @@
+"use client"
+
+import { useMemo, useState } from "react"
+import { BookOpen, ChevronDown, ChevronUp, ExternalLink, FileText, Globe, Layers, ScrollText, Tag, Users } from "lucide-react"
+import { SignalTooltip } from "@/components/signals/signal-tooltip"
+import type { BenchmarkEvalSummary } from "@/lib/eval-processing"
+import { getKnownIssues } from "@/lib/known-issues"
+import { KnownIssuesPanel } from "@/components/known-issues-panel"
+
+interface PolicyOverviewProps {
+  summary: BenchmarkEvalSummary
+}
+
+const SUMMARY_PREVIEW_CHARS = 280
+
+function classifyResource(url: string): { kind: "paper" | "dataset" | "leaderboard" | "site"; label: string } {
+  const lower = url.toLowerCase()
+  if (lower.includes("arxiv.org") || lower.endsWith(".pdf") || lower.includes("/papers/")) {
+    return { kind: "paper", label: "Paper" }
+  }
+  if (lower.includes("huggingface.co/datasets") || lower.includes("/dataset")) {
+    return { kind: "dataset", label: "Dataset" }
+  }
+  if (lower.includes("leaderboard")) {
+    return { kind: "leaderboard", label: "Leaderboard" }
+  }
+  return { kind: "site", label: "Source" }
+}
+
+function shortHost(url: string) {
+  return url.replace(/^https?:\/\//, "").replace(/\/.*$/, "")
+}
+
+/**
+ * Plain-language summary surface shown at the top of the benchmark page in
+ * policy mode. Designed to answer three questions a non-technical reader has:
+ * what does this measure, who built it, and where does it come from?
+ *
+ * Technical detail (variants, metric specifications, score scales) lives in
+ * the existing overview card below this and is collapsed by default.
+ */
+function normalizeId(value: string | undefined | null): string {
+  if (!value) return ""
+  return value
+    .toLowerCase()
+    .trim()
+    .replace(/[\s_\-/]+/g, "")
+}
+
+export function PolicyOverview({ summary }: PolicyOverviewProps) {
+  const card = summary.benchmark_card
+
+  // Defensive check: pipelines older than the "ancestor card leak" fix
+  // sometimes attach the parent suite's card to a leaf benchmark (e.g.
+  // helm_classic's card embedded under XSUM). Detect when the card's own
+  // name is clearly not this benchmark and ignore its narrative text — the
+  // synthesized fallback below produces something accurate instead.
+  const cardName = card?.benchmark_details?.name
+  const cardNameNorm = normalizeId(cardName)
+  const evalIdentifiers = [
+    summary.evaluation_name,
+    summary.benchmark_leaf_key,
+    summary.composite_benchmark_key,
+    summary.composite_benchmark_name,
+    summary.canonical_display_name,
+    summary.evaluation_id,
+  ].map(normalizeId)
+  // Treat the card as belonging to this eval when its name fuzzily appears in
+  // any of the eval's identifiers (or vice versa). Otherwise the card is from
+  // a different (typically ancestor) benchmark.
+  const cardMatchesEval =
+    !cardName ||
+    evalIdentifiers.some(
+      (id) =>
+        id.length > 0 &&
+        cardNameNorm.length > 0 &&
+        (id.includes(cardNameNorm) || cardNameNorm.includes(id)),
+    )
+
+  const overview = cardMatchesEval ? card?.benchmark_details?.overview?.trim() || "" : ""
+  const goal = cardMatchesEval ? card?.purpose_and_intended_users?.goal?.trim() || "" : ""
+
+  // Detect "parent" benchmark pages — either an aggregated composite or a
+  // multi-metric matrix where each column is a subtask. In both cases the
+  // per-evaluation `metric_config.evaluation_description` belongs to whichever
+  // component was processed first (e.g. just the "airline" subset of Tau
+  // Bench 2) and would mislead a policy reader. Synthesize parent framing
+  // instead and surface the subtasks separately.
+  const isAggregated = summary.is_aggregated === true
+  const aggregateNames = (summary.aggregate_sources ?? [])
+    .map((s) => s.composite_benchmark_name)
+    .filter((s): s is string => typeof s === "string" && s.length > 0)
+
+  const subtaskLabels = useMemo(() => {
+    const seen = new Set<string>()
+    const labels: string[] = []
+    const add = (raw: string | undefined | null) => {
+      if (!raw) return
+      const trimmed = raw.trim()
+      if (!trimmed) return
+      const key = trimmed.toLowerCase()
+      if (seen.has(key)) return
+      seen.add(key)
+      labels.push(trimmed)
+    }
+    for (const subtask of summary.subtasks ?? []) {
+      add(subtask.display_name || subtask.subtask_name)
+    }
+    for (const metric of summary.leaderboard_metrics ?? []) {
+      if (metric.scope === "subtask") {
+        add(metric.subtask_name || metric.display_name)
+      }
+    }
+    for (const name of aggregateNames) add(name)
+    return labels
+  }, [summary.subtasks, summary.leaderboard_metrics, aggregateNames])
+
+  const isMatrix = (summary.leaderboard_metrics?.length ?? 0) > 1
+  const isParentPage = isAggregated || (isMatrix && subtaskLabels.length > 1)
+  const useComponentDescription = !isParentPage
+
+  const parentFallback = isParentPage && subtaskLabels.length > 1
+    ? `${summary.evaluation_name} reports results across ${subtaskLabels.length} ${
+        isAggregated ? "component benchmarks" : "subtasks"
+      }. Each is evaluated separately; the score shown is the ${
+        isAggregated ? "average" : "per-subtask result"
+      }.`
+    : null
+
+  const summaryText =
+    overview ||
+    goal ||
+    parentFallback ||
+    (useComponentDescription ? summary.metric_config.evaluation_description : summary.evaluation_name)
+
+  const [expanded, setExpanded] = useState(false)
+  const [subtasksOpen, setSubtasksOpen] = useState(false)
+  const isLong = summaryText.length > SUMMARY_PREVIEW_CHARS
+  const visibleText = expanded || !isLong
+    ? summaryText
+    : summaryText.slice(0, SUMMARY_PREVIEW_CHARS).replace(/\s+\S*$/, "") + "…"
+
+  const domains = useMemo(() => {
+    const fromTags = summary.tags?.domains ?? []
+    const fromCard = cardMatchesEval ? card?.benchmark_details?.domains ?? [] : []
+    return Array.from(new Set([...fromTags, ...fromCard].map((d) => d.trim()).filter(Boolean))).slice(0, 6)
+  }, [summary.tags?.domains, card?.benchmark_details?.domains, cardMatchesEval])
+
+  const languages = useMemo(() => {
+    const fromTags = summary.tags?.languages ?? []
+    const fromCard = cardMatchesEval ? card?.benchmark_details?.languages ?? [] : []
+    return Array.from(new Set([...fromTags, ...fromCard].map((d) => d.trim()).filter(Boolean))).slice(0, 4)
+  }, [summary.tags?.languages, card?.benchmark_details?.languages, cardMatchesEval])
+
+  const license = card?.ethical_and_legal_considerations?.data_licensing
+  const showLicense = license && license !== "Not specified"
+
+  const resources = useMemo(() => {
+    if (!cardMatchesEval) return []
+    const urls = (card?.benchmark_details?.resources ?? []).filter((r) => typeof r === "string" && r.startsWith("http"))
+    const seen = new Map<string, { kind: ReturnType<typeof classifyResource>["kind"]; label: string; url: string }>()
+    for (const url of urls) {
+      const c = classifyResource(url)
+      if (!seen.has(c.kind)) {
+        seen.set(c.kind, { ...c, url })
+      }
+    }
+    return Array.from(seen.values())
+  }, [card?.benchmark_details?.resources, cardMatchesEval])
+
+  const evaluators = (summary.evaluator_names ?? []).slice(0, 3)
+  const hasMoreEvaluators = (summary.evaluator_names?.length ?? 0) > evaluators.length
+
+  const knownIssues = useMemo(
+    () =>
+      getKnownIssues(
+        summary.evaluation_name,
+        summary.composite_benchmark_name,
+        summary.composite_benchmark_key,
+        summary.benchmark_family_key,
+        summary.benchmark_leaf_key,
+        card?.benchmark_details?.name,
+      ),
+    [
+      summary.evaluation_name,
+      summary.composite_benchmark_name,
+      summary.composite_benchmark_key,
+      summary.benchmark_family_key,
+      summary.benchmark_leaf_key,
+      card?.benchmark_details?.name,
+    ],
+  )
+
+  const directionLabel = summary.metric_config.lower_is_better
+    ? "Lower scores are better"
+    : "Higher scores are better"
+
+  return (
+    <section className="rounded-3xl border bg-card p-5 sm:p-6">
+      <header className="mb-4 flex flex-wrap items-center gap-2">
+        <BookOpen className="h-5 w-5 text-primary" />
+        <h2 className="text-lg font-semibold tracking-tight">{summary.evaluation_name}</h2>
+        <span className="text-xs text-muted-foreground">In plain language</span>
+      </header>
+
+      {knownIssues.length > 0 && (
+        <div className="mb-3">
+          <KnownIssuesPanel issues={knownIssues} variant="compact" />
+        </div>
+      )}
+
+      <p className="text-base leading-7 text-foreground/90">
+        {visibleText}
+        {isLong && (
+          <button
+            type="button"
+            onClick={() => setExpanded((v) => !v)}
+            className="ml-1 text-sm font-medium text-primary underline-offset-4 hover:underline"
+          >
+            {expanded ? "Show less" : "Read more"}
+          </button>
+        )}
+      </p>
+
+      <p className="mt-3 text-sm text-muted-foreground">
+        <SignalTooltip
+          content={
+            summary.metric_config.lower_is_better
+              ? "On this benchmark, a lower number means the model did better."
+              : "On this benchmark, a higher number means the model did better."
+          }
+        >
+          <span className="underline decoration-dotted decoration-muted-foreground/60 underline-offset-4 cursor-help">
+            {directionLabel}.
+          </span>
+        </SignalTooltip>{" "}
+        Compared across {summary.models_count} model{summary.models_count === 1 ? "" : "s"}.
+      </p>
+
+      {isParentPage && subtaskLabels.length > 1 && (
+        <div className="mt-4 rounded-2xl border border-border/60 bg-muted/10">
+          <button
+            type="button"
+            onClick={() => setSubtasksOpen((v) => !v)}
+            aria-expanded={subtasksOpen}
+            className="flex w-full items-center justify-between gap-3 px-3.5 py-2.5 text-left transition-colors hover:bg-muted/20 rounded-2xl"
+          >
+            <span className="flex items-center gap-1.5 text-[11px] font-semibold uppercase tracking-[0.18em] text-muted-foreground">
+              <Layers className="h-3.5 w-3.5" />
+              {isAggregated ? `Component benchmarks (${subtaskLabels.length})` : `Subtasks (${subtaskLabels.length})`}
+            </span>
+            {subtasksOpen ? (
+              <ChevronUp className="h-4 w-4 text-muted-foreground" />
+            ) : (
+              <ChevronDown className="h-4 w-4 text-muted-foreground" />
+            )}
+          </button>
+          {subtasksOpen && (
+            <ul className="grid list-disc gap-x-6 gap-y-1 px-3.5 pb-3.5 pl-9 text-sm sm:grid-cols-2 lg:grid-cols-3">
+              {subtaskLabels.map((name) => (
+                <li key={name} className="capitalize">
+                  {name}
+                </li>
+              ))}
+            </ul>
+          )}
+        </div>
+      )}
+
+      {(domains.length > 0 || languages.length > 0 || showLicense) && (
+        <div className="mt-4 flex flex-wrap gap-2">
+          {domains.map((d) => (
+            <span
+              key={`d-${d}`}
+              className="inline-flex items-center gap-1 rounded-full border border-border/60 bg-muted/30 px-2.5 py-1 text-xs font-medium capitalize"
+            >
+              <Tag className="h-3 w-3 shrink-0 text-muted-foreground" />
+              {d}
+            </span>
+          ))}
+          {languages.map((l) => (
+            <span
+              key={`l-${l}`}
+              className="inline-flex items-center gap-1 rounded-full border border-sky-200/70 bg-sky-50/60 px-2.5 py-1 text-xs font-medium dark:border-sky-900/40 dark:bg-sky-950/20"
+            >
+              <Globe className="h-3 w-3 shrink-0 text-sky-600" />
+              {l}
+            </span>
+          ))}
+          {showLicense && (
+            <SignalTooltip content="The license under which the benchmark dataset is released.">
+              <span className="inline-flex items-center gap-1 rounded-full border border-border/60 bg-background px-2.5 py-1 text-xs font-medium cursor-help">
+                <ScrollText className="h-3 w-3 shrink-0 text-muted-foreground" />
+                {license}
+              </span>
+            </SignalTooltip>
+          )}
+        </div>
+      )}
+
+      {(resources.length > 0 || evaluators.length > 0) && (
+        <div className="mt-5 border-t pt-4">
+          <div className="mb-2 text-[11px] font-semibold uppercase tracking-[0.18em] text-muted-foreground">
+            Where this comes from
+          </div>
+          <div className="flex flex-wrap items-center gap-3">
+            {resources.map((r) => (
+              <a
+                key={r.url}
+                href={r.url}
+                target="_blank"
+                rel="noreferrer"
+                className="inline-flex items-center gap-1.5 rounded-full border border-border/60 bg-background px-3 py-1.5 text-xs font-medium text-muted-foreground transition-colors hover:border-primary/40 hover:text-primary"
+              >
+                <FileText className="h-3 w-3 shrink-0" />
+                <span className="font-semibold text-foreground">{r.label}</span>
+                <span className="text-muted-foreground">{shortHost(r.url)}</span>
+                <ExternalLink className="h-3 w-3 shrink-0" />
+              </a>
+            ))}
+            {evaluators.length > 0 && (
+              <SignalTooltip
+                content={
+                  <span className="block space-y-1">
+                    <span className="block font-semibold">Who reported these scores</span>
+                    <span className="block">{summary.evaluator_names?.join(", ")}</span>
+                    {summary.third_party_ratio > 0 && (
+                      <span className="block text-muted-foreground">
+                        {Math.round(summary.third_party_ratio * 100)}% of results come from independent evaluators (not the model's own developer).
+                      </span>
+                    )}
+                  </span>
+                }
+              >
+                <span className="inline-flex items-center gap-1.5 rounded-full border border-border/60 bg-background px-3 py-1.5 text-xs font-medium cursor-help">
+                  <Users className="h-3 w-3 shrink-0 text-muted-foreground" />
+                  <span className="font-semibold text-foreground">Reported by</span>
+                  <span className="text-muted-foreground">
+                    {evaluators.join(", ")}
+                    {hasMoreEvaluators ? ` +${(summary.evaluator_names?.length ?? 0) - evaluators.length} more` : ""}
+                  </span>
+                </span>
+              </SignalTooltip>
+            )}
+          </div>
+        </div>
+      )}
+    </section>
+  )
+}

components/researcher-reproducibility-card.tsx

@@ -0,0 +1,483 @@
+"use client"
+
+import { useEffect, useState, type ReactNode } from "react"
+import { AlertTriangle, ExternalLink, FlaskConical } from "lucide-react"
+import { Term } from "@/components/term"
+import { SignalTooltip } from "@/components/signals/signal-tooltip"
+import type { ModelResultForBenchmark } from "@/lib/eval-processing"
+import type { GenerationConfig, ScoreDetails } from "@/lib/benchmark-schema"
+
+interface ResearcherReproducibilityCardProps {
+  modelResult: ModelResultForBenchmark
+  /**
+   * Benchmark identifier used to pick the right per-eval row when fetching
+   * enrichment from the model's full record. The eval-detail endpoint
+   * synthesizes leaderboard rows without `generation_config`, so we top up
+   * lazily on row expand from /api/eval-row-config.
+   */
+  benchmarkKey?: string
+  evalName?: string
+}
+
+const KNOWN_DECODING_KEYS = ["temperature", "top_p", "top_k", "max_tokens", "seed", "reasoning"] as const
+
+// Keys that belong to agentic eval setups, surfaced in their own group rather
+// than dumped under "decoding" extras as raw JSON.
+const KNOWN_AGENT_KEYS = [
+  "agentic_eval_config",
+  "max_attempts",
+  "eval_limits",
+  "eval_plan",
+  "sandbox",
+  "max_turns",
+  "message_limit",
+] as const
+
+const KEY_LABEL: Record<string, string> = {
+  temperature: "temperature",
+  top_p: "top-p",
+  top_k: "top-k",
+  max_tokens: "max tokens",
+  seed: "seed",
+  reasoning: "reasoning mode",
+  n: "samples per prompt",
+  best_of: "best-of-N",
+  num_samples: "samples per prompt",
+  num_runs: "runs",
+  n_shot: "n-shot",
+  num_fewshot: "few-shot examples",
+  fewshot: "few-shot examples",
+  agentic_eval_config: "tools available",
+  max_attempts: "max attempts",
+  eval_limits: "eval limits",
+  eval_plan: "eval plan",
+  sandbox: "sandbox",
+  max_turns: "max turns",
+  message_limit: "message limit",
+}
+
+/**
+ * Try to render a structured agentic config object as a short, readable
+ * string. Falls back to null so the caller can use the generic formatter.
+ */
+function formatAgentValue(key: string, value: unknown): string | null {
+  if (value == null) return null
+  if (typeof value !== "object") return null
+
+  if (key === "agentic_eval_config") {
+    const tools = (value as { available_tools?: unknown }).available_tools
+    if (Array.isArray(tools)) {
+      const names = tools
+        .map((t) => (t && typeof t === "object" ? (t as { name?: unknown }).name : null))
+        .filter((n): n is string => typeof n === "string" && n.length > 0)
+      if (names.length === 0) return "no tools"
+      if (names.length <= 4) return `${names.length} tools: ${names.join(", ")}`
+      return `${names.length} tools: ${names.slice(0, 4).join(", ")} +${names.length - 4}`
+    }
+  }
+
+  if (key === "eval_limits") {
+    const obj = value as Record<string, unknown>
+    const parts: string[] = []
+    for (const k of ["message_limit", "max_messages", "token_limit", "max_tokens"]) {
+      if (typeof obj[k] === "number") parts.push(`${k.replace(/_/g, " ")}: ${obj[k]}`)
+    }
+    if (parts.length > 0) return parts.join(", ")
+  }
+
+  if (key === "eval_plan") {
+    const name = (value as { name?: unknown }).name
+    const steps = (value as { steps?: unknown }).steps
+    if (typeof name === "string" && Array.isArray(steps)) return `${name} (${steps.length} step${steps.length === 1 ? "" : "s"})`
+    if (typeof name === "string") return name
+  }
+
+  if (key === "sandbox") {
+    const keys = Object.keys(value as Record<string, unknown>)
+    if (keys.length === 0) return "default"
+    return keys.join(", ")
+  }
+
+  return null
+}
+
+function pickFromAdditional(value: unknown, keys: string[]): unknown | undefined {
+  if (!value || typeof value !== "object") return undefined
+  const obj = value as Record<string, unknown>
+  for (const k of keys) {
+    if (obj[k] !== undefined && obj[k] !== null) return obj[k]
+  }
+  return undefined
+}
+
+/**
+ * Renders any value as a short string for the parameter cards. Returns null
+ * when the value is empty so the caller can render "Not disclosed" instead.
+ */
+function formatValue(value: unknown): string | null {
+  if (value === undefined || value === null) return null
+  if (typeof value === "boolean") return value ? "yes" : "no"
+  if (typeof value === "number") return Number.isInteger(value) ? value.toString() : value.toFixed(3)
+  if (typeof value === "string") {
+    const trimmed = value.trim()
+    return trimmed.length === 0 ? null : trimmed
+  }
+  try {
+    return JSON.stringify(value)
+  } catch {
+    return null
+  }
+}
+
+function ParamRow({
+  label,
+  value,
+  termKey,
+  hint,
+}: {
+  label: string
+  value: ReactNode | null
+  termKey?: string
+  hint?: string
+}) {
+  const isMissing = value === null || value === undefined
+  return (
+    <div className="flex items-baseline justify-between gap-3 border-b border-dashed border-border/50 py-1.5 text-sm last:border-0">
+      <span className="text-muted-foreground">
+        {termKey ? <Term term={termKey}>{label}</Term> : label}
+      </span>
+      {isMissing ? (
+        <SignalTooltip
+          content={
+            hint ??
+            "This parameter wasn't reported by the source. Without it, the result may not be exactly reproducible."
+          }
+        >
+          <span className="inline-flex items-center gap-1 text-xs font-medium text-amber-700 dark:text-amber-300 cursor-help">
+            <AlertTriangle className="h-3 w-3" /> Not disclosed
+          </span>
+        </SignalTooltip>
+      ) : (
+        <span className="font-medium tabular-nums">{value}</span>
+      )}
+    </div>
+  )
+}
+
+interface FieldSpec {
+  label: string
+  value: string | null
+  termKey?: string
+  hint?: string
+}
+
+interface FieldGroup {
+  title: string
+  fields: FieldSpec[]
+}
+
+/**
+ * Detailed reproducibility surface for researcher mode. Shown inside a
+ * leaderboard row's expanded panel. Adaptive: when most fields aren't disclosed
+ * (the common case today), it collapses to a compact "Limited disclosure"
+ * summary showing only fields that *are* present, with a button to reveal the
+ * full audit grid.
+ */
+export function ResearcherReproducibilityCard({
+  modelResult,
+  benchmarkKey,
+  evalName,
+}: ResearcherReproducibilityCardProps) {
+  const [enrichedGen, setEnrichedGen] = useState<GenerationConfig | null>(null)
+  const [enrichedScore, setEnrichedScore] = useState<ScoreDetails | null>(null)
+  const [loading, setLoading] = useState(false)
+
+  // Lazily top up generation_config / score_details from the model's full
+  // record. The eval-detail leaderboard endpoint omits these fields; we only
+  // pay the fetch cost when a researcher actually expands a row.
+  useEffect(() => {
+    const inlineGen = modelResult.result.generation_config
+    const inlineScoreOk =
+      modelResult.score_details.standard_error != null ||
+      modelResult.score_details.confidence_interval != null ||
+      modelResult.score_details.sample_size != null
+    const hasInlineArgs =
+      !!inlineGen &&
+      typeof inlineGen === "object" &&
+      "generation_args" in inlineGen &&
+      Object.keys((inlineGen as { generation_args?: Record<string, unknown> }).generation_args ?? {}).length > 0
+    if (hasInlineArgs && inlineScoreOk) return // nothing to fetch
+
+    const modelId = modelResult.model_info.id
+    if (!modelId) return
+
+    const params = new URLSearchParams({ model_id: modelId })
+    if (benchmarkKey) params.set("benchmark_key", benchmarkKey)
+    if (evalName) params.set("eval_name", evalName)
+
+    let cancelled = false
+    setLoading(true)
+    fetch(`/api/eval-row-config?${params.toString()}`)
+      .then((r) => (r.ok ? r.json() : null))
+      .then((data) => {
+        if (cancelled || !data) return
+        if (data.generation_config) setEnrichedGen(data.generation_config as GenerationConfig)
+        if (data.score_details) setEnrichedScore(data.score_details as ScoreDetails)
+      })
+      .catch(() => {})
+      .finally(() => {
+        if (!cancelled) setLoading(false)
+      })
+    return () => {
+      cancelled = true
+    }
+  }, [modelResult, benchmarkKey, evalName])
+
+  const gen = modelResult.result.generation_config ?? enrichedGen ?? undefined
+  const args = gen?.generation_args ?? {}
+  const scoreDetails = {
+    ...modelResult.score_details,
+    sample_size: modelResult.score_details.sample_size ?? enrichedScore?.sample_size,
+    standard_error: modelResult.score_details.standard_error ?? enrichedScore?.standard_error,
+    confidence_interval:
+      modelResult.score_details.confidence_interval ?? enrichedScore?.confidence_interval,
+  }
+  const additional =
+    typeof gen?.additional_details === "object" && gen?.additional_details !== null
+      ? (gen.additional_details as Record<string, unknown>)
+      : null
+
+  const argsKeys = Object.keys(args)
+  const extraDecodingKeys = argsKeys.filter(
+    (k) =>
+      !KNOWN_DECODING_KEYS.includes(k as (typeof KNOWN_DECODING_KEYS)[number]) &&
+      !KNOWN_AGENT_KEYS.includes(k as (typeof KNOWN_AGENT_KEYS)[number])
+  )
+  const agentKeysPresent = KNOWN_AGENT_KEYS.filter((k) => args[k] != null)
+  const hasAgentSetup = agentKeysPresent.length > 0
+
+  const shots = pickFromAdditional(additional, ["num_fewshot", "n_shot", "shots", "fewshot"])
+  const samplesPerPrompt = pickFromAdditional(additional, ["n", "num_samples", "samples_per_prompt"]) ?? args["n"]
+  const bestOf = pickFromAdditional(additional, ["best_of", "best_of_n"]) ?? args["best_of"]
+  const numRuns = pickFromAdditional(additional, ["num_runs", "n_runs", "runs"])
+  const scoringMethod = pickFromAdditional(additional, ["scoring", "scoring_method", "judge", "evaluator"])
+  const evalLibrary = pickFromAdditional(additional, ["eval_library", "harness", "framework"])
+  const evalLibraryVersion = pickFromAdditional(additional, ["eval_library_version", "harness_version"])
+  const promptTemplate = gen?.prompt_template?.trim() || null
+
+  const groups: FieldGroup[] = [
+    {
+      title: "Decoding",
+      fields: [
+        {
+          label: "temperature",
+          termKey: "temperature",
+          value: formatValue(args.temperature),
+          hint: "Temperature controls randomness — without it, others can't recreate the same outputs.",
+        },
+        { label: "top-p", termKey: "top-p", value: formatValue(args.top_p) },
+        { label: "top-k", termKey: "top-k", value: formatValue(args.top_k) },
+        { label: "max tokens", value: formatValue(args.max_tokens) },
+        { label: "seed", value: formatValue(args.seed) },
+        ...extraDecodingKeys.map((k) => ({
+          label: KEY_LABEL[k] ?? k.replace(/_/g, " "),
+          value: formatValue(args[k]),
+        })),
+      ],
+    },
+    {
+      title: "Sampling",
+      fields: [
+        {
+          label: "few-shot examples",
+          termKey: "few-shot",
+          value: formatValue(shots),
+          hint: "How many worked examples were included in the prompt before the question.",
+        },
+        {
+          label: "samples per prompt",
+          value: formatValue(samplesPerPrompt),
+          hint: "Number of completions generated per question.",
+        },
+        {
+          label: "best-of-N",
+          termKey: "best-of-n",
+          value: formatValue(bestOf),
+          hint: "Whether the score reflects the best of multiple attempts (inflates results vs. single-attempt).",
+        },
+        {
+          label: "runs averaged",
+          value: formatValue(numRuns),
+          hint: "How many evaluation runs were averaged to produce the reported number.",
+        },
+        {
+          label: "test instances",
+          value: formatValue(scoreDetails.sample_size),
+          hint: "Number of items in the test set the model was scored on.",
+        },
+      ],
+    },
+    ...(hasAgentSetup
+      ? [
+          {
+            title: "Agent setup",
+            fields: agentKeysPresent.map((k) => ({
+              label: KEY_LABEL[k] ?? k.replace(/_/g, " "),
+              value: formatAgentValue(k, args[k]) ?? formatValue(args[k]),
+              hint:
+                k === "agentic_eval_config"
+                  ? "Tools the agent could call during the run."
+                  : k === "max_attempts"
+                    ? "Maximum independent attempts the agent gets per task."
+                    : k === "eval_limits"
+                      ? "Hard caps the harness enforced on the run (messages, tokens, etc.)."
+                      : k === "eval_plan"
+                        ? "Solver/plan the harness used to drive the agent."
+                        : k === "sandbox"
+                          ? "Environment in which the agent ran (e.g. docker, local)."
+                          : undefined,
+            })),
+          } as FieldGroup,
+        ]
+      : []),
+    {
+      title: "Scoring & uncertainty",
+      fields: [
+        {
+          label: "scoring method",
+          value: formatValue(scoringMethod),
+          hint: "Exact match, LLM-as-judge, human grading, etc. Determines what 'correct' means.",
+        },
+        { label: "standard error", value: formatValue(scoreDetails.standard_error) },
+        {
+          label: "confidence interval",
+          value: scoreDetails.confidence_interval
+            ? `${scoreDetails.confidence_interval.lower}–${scoreDetails.confidence_interval.upper} (${scoreDetails.confidence_interval.confidence_level}%)`
+            : null,
+        },
+        { label: "eval library", value: formatValue(evalLibrary) },
+        { label: "library version", value: formatValue(evalLibraryVersion) },
+      ],
+    },
+  ]
+
+  const totalFields = groups.reduce((n, g) => n + g.fields.length, 0)
+  const disclosedFields = groups.reduce(
+    (n, g) => n + g.fields.filter((f) => f.value !== null).length,
+    0
+  )
+  const disclosureRatio = totalFields > 0 ? disclosedFields / totalFields : 0
+  const disclosedGroups = groups
+    .map((g) => ({ ...g, fields: g.fields.filter((f) => f.value !== null) }))
+    .filter((g) => g.fields.length > 0)
+
+  // If fewer than ~30% of fields are disclosed (and at least one is missing),
+  // start in compact mode so the card isn't a wall of "Not disclosed".
+  const shouldStartCompact = disclosureRatio < 0.3 && disclosedFields < totalFields
+  const [showAll, setShowAll] = useState(!shouldStartCompact)
+  const isCompact = shouldStartCompact && !showAll
+
+  return (
+    <section className="rounded-2xl border bg-background/70 p-4">
+      <header className="mb-3 flex items-start justify-between gap-3">
+        <div className="flex items-start gap-2 min-w-0">
+          <FlaskConical className="mt-0.5 h-4 w-4 shrink-0 text-primary" />
+          <div className="min-w-0">
+            <div className="text-sm font-semibold">Reproducibility</div>
+            <div className="text-xs text-muted-foreground">
+              {isCompact
+                ? `Limited disclosure — only ${disclosedFields} of ${totalFields} reproducibility fields are reported.`
+                : "Everything someone would need to re-run this evaluation. Missing fields are flagged."}
+            </div>
+          </div>
+        </div>
+        <span className="shrink-0 rounded-full border border-border/60 bg-muted/30 px-2 py-0.5 text-[11px] font-medium tabular-nums text-muted-foreground">
+          {loading ? "loading…" : `${disclosedFields}/${totalFields} disclosed`}
+        </span>
+      </header>
+
+      {isCompact ? (
+        <>
+          {disclosedGroups.length > 0 ? (
+            <div className="grid gap-4 sm:grid-cols-2 lg:grid-cols-3">
+              {disclosedGroups.map((g) => (
+                <div key={g.title}>
+                  <div className="mb-2 text-[10px] font-semibold uppercase tracking-[0.18em] text-muted-foreground">
+                    {g.title}
+                  </div>
+                  {g.fields.map((f) => (
+                    <ParamRow key={f.label} label={f.label} termKey={f.termKey} value={f.value} hint={f.hint} />
+                  ))}
+                </div>
+              ))}
+            </div>
+          ) : (
+            <div className="rounded-xl border border-dashed border-amber-300/60 bg-amber-50/40 px-3 py-2 text-xs text-amber-800 dark:border-amber-900/60 dark:bg-amber-950/20 dark:text-amber-200">
+              No reproducibility metadata was disclosed by the source.
+            </div>
+          )}
+        </>
+      ) : (
+        <div className="grid gap-4 lg:grid-cols-3">
+          {groups.map((g) => (
+            <div key={g.title}>
+              <div className="mb-2 text-[10px] font-semibold uppercase tracking-[0.18em] text-muted-foreground">
+                {g.title}
+              </div>
+              {g.fields.map((f) => (
+                <ParamRow key={f.label} label={f.label} termKey={f.termKey} value={f.value} hint={f.hint} />
+              ))}
+            </div>
+          ))}
+        </div>
+      )}
+
+      {shouldStartCompact && (
+        <button
+          type="button"
+          onClick={() => setShowAll((v) => !v)}
+          className="mt-3 inline-flex items-center text-xs font-medium text-primary underline-offset-4 hover:underline"
+        >
+          {showAll
+            ? "Hide undisclosed fields"
+            : `Show all ${totalFields} checked fields (${totalFields - disclosedFields} not disclosed)`}
+        </button>
+      )}
+
+      {promptTemplate && (
+        <details className="mt-4 rounded-xl border bg-muted/10">
+          <summary className="cursor-pointer px-3 py-2 text-xs font-semibold uppercase tracking-[0.16em] text-muted-foreground">
+            Prompt template
+          </summary>
+          <pre className="max-h-[18rem] overflow-auto whitespace-pre-wrap break-words border-t bg-background/60 px-3 py-3 text-xs leading-5">
+            {promptTemplate}
+          </pre>
+        </details>
+      )}
+      {!promptTemplate && !isCompact && (
+        <div className="mt-4 flex items-center gap-1.5 rounded-xl border border-dashed border-amber-300/60 bg-amber-50/40 px-3 py-2 text-xs text-amber-800 dark:border-amber-900/60 dark:bg-amber-950/20 dark:text-amber-200">
+          <AlertTriangle className="h-3.5 w-3.5" />
+          <span>
+            Prompt template not disclosed by the source.
+            <SignalTooltip content="Without the prompt, scores can't be reliably reproduced because phrasing changes results.">
+              <span className="ml-1 underline decoration-dotted underline-offset-4 cursor-help">Why this matters</span>
+            </SignalTooltip>
+          </span>
+        </div>
+      )}
+
+      {modelResult.source_metadata.source_url && (
+        <div className="mt-3 text-xs">
+          <a
+            href={modelResult.source_metadata.source_url}
+            target="_blank"
+            rel="noreferrer"
+            className="inline-flex items-center gap-1 text-primary underline-offset-4 hover:underline"
+          >
+            View original source <ExternalLink className="h-3 w-3" />
+          </a>
+        </div>
+      )}
+    </section>
+  )
+}

components/signals/comparability-panel.tsx

@@ -1,11 +1,12 @@
 "use client"
 
 import type { ReactNode } from "react"
-import { ChevronDown, GitCompareArrows, UsersRound } from "lucide-react"
+import { ChevronDown, GitCompareArrows, Info, UsersRound } from "lucide-react"
 
 import { useAudienceMode } from "@/components/audience-mode-provider"
 import { Badge } from "@/components/ui/badge"
 import { Collapsible, CollapsibleContent, CollapsibleTrigger } from "@/components/ui/collapsible"
+import { SignalTooltip } from "@/components/signals/signal-tooltip"
 import type { BenchmarkComparability, ComparabilitySummary, DifferingSetupField } from "@/lib/backend-artifacts"
 import {
   formatFieldLabel,
@@ -31,7 +32,10 @@ export function ComparabilityPanel({
   }
 
   return (
-    <section className="rounded-2xl border border-border/70 bg-background/70 p-4 sm:p-5">
+    <section
+      id="comparability-panel"
+      className="rounded-2xl border border-border/70 bg-background/70 p-4 sm:p-5 scroll-mt-24"
+    >
       <div className="flex flex-col gap-2 sm:flex-row sm:items-start sm:justify-between">
         <div className="space-y-1">
           <div className="flex items-center gap-2">
@@ -60,44 +64,54 @@ export function ComparabilityPanel({
         </div>
       )}
 
-      <div className="mt-4 grid gap-3 lg:grid-cols-2">
-        {variantGroups.length > 0 && (
-          <GroupList
-            icon="variant"
-            title="Variant divergence"
-            count={variantGroups.length}
-          >
-            {variantGroups.slice(0, 8).map((group) => (
-              <DivergenceGroupItem
-                key={group.group_id}
-                modelRouteId={group.model_route_id}
-                magnitude={group.divergence_magnitude}
-                threshold={group.threshold_used}
-                fields={group.differing_setup_fields}
-              />
-            ))}
-          </GroupList>
-        )}
+      {(() => {
+        const onlyOne =
+          (variantGroups.length > 0 ? 1 : 0) + (crossPartyGroups.length > 0 ? 1 : 0) === 1
+        const sectionClass = onlyOne ? "" : "lg:grid lg:grid-cols-2 lg:gap-3"
+        const itemsClass = onlyOne ? "grid gap-2 md:grid-cols-2" : "space-y-2"
+        return (
+          <div className={`mt-4 ${sectionClass}`}>
+            {variantGroups.length > 0 && (
+              <GroupList
+                icon="variant"
+                title="Variant divergence"
+                count={variantGroups.length}
+                itemsClassName={itemsClass}
+              >
+                {variantGroups.slice(0, 8).map((group) => (
+                  <DivergenceGroupItem
+                    key={group.group_id}
+                    modelRouteId={group.model_route_id}
+                    magnitude={group.divergence_magnitude}
+                    threshold={group.threshold_used}
+                    fields={group.differing_setup_fields}
+                  />
+                ))}
+              </GroupList>
+            )}
 
-        {crossPartyGroups.length > 0 && (
-          <GroupList
-            icon="cross-party"
-            title="Cross-party divergence"
-            count={crossPartyGroups.length}
-          >
-            {crossPartyGroups.slice(0, 8).map((group) => (
-              <DivergenceGroupItem
-                key={group.group_id}
-                modelRouteId={group.model_route_id}
-                magnitude={group.divergence_magnitude}
-                threshold={group.threshold_used}
-                fields={group.differing_setup_fields}
-                scoresByOrganization={group.scores_by_organization}
-              />
-            ))}
-          </GroupList>
-        )}
-      </div>
+            {crossPartyGroups.length > 0 && (
+              <GroupList
+                icon="cross-party"
+                title="Cross-party divergence"
+                count={crossPartyGroups.length}
+                itemsClassName={itemsClass}
+              >
+                {crossPartyGroups.slice(0, 8).map((group) => (
+                  <DivergenceGroupItem
+                    key={group.group_id}
+                    modelRouteId={group.model_route_id}
+                    magnitude={group.divergence_magnitude}
+                    threshold={group.threshold_used}
+                    fields={group.differing_setup_fields}
+                    scoresByOrganization={group.scores_by_organization}
+                  />
+                ))}
+              </GroupList>
+            )}
+          </div>
+        )
+      })()}
     </section>
   )
 }
@@ -107,11 +121,13 @@ function GroupList({
   title,
   count,
   children,
+  itemsClassName = "space-y-2",
 }: {
   icon: "variant" | "cross-party"
   title: string
   count: number
   children: ReactNode
+  itemsClassName?: string
 }) {
   const Icon = icon === "variant" ? GitCompareArrows : UsersRound
 
@@ -130,13 +146,58 @@ function GroupList({
           <ChevronDown className="h-4 w-4 text-muted-foreground" />
         </button>
       </CollapsibleTrigger>
-      <CollapsibleContent className="mt-2 space-y-2">
+      <CollapsibleContent className={`mt-2 ${itemsClassName}`}>
         {children}
       </CollapsibleContent>
     </Collapsible>
   )
 }
 
+/**
+ * Try to extract a human-readable label from a structured setup-field value.
+ * Common shape from agentic evals: { additional_details: { agent_name, agent_framework } }.
+ * Falls back to picking the first short string property, or null when the
+ * value can't be summarized cleanly.
+ */
+function extractFriendlyLabel(value: unknown): string | null {
+  if (value == null) return null
+  if (typeof value === "string") return value.length > 60 ? null : value
+  if (typeof value === "number" || typeof value === "boolean") return String(value)
+  if (typeof value !== "object") return null
+
+  const obj = value as Record<string, unknown>
+  const details = (obj.additional_details && typeof obj.additional_details === "object")
+    ? (obj.additional_details as Record<string, unknown>)
+    : obj
+
+  const agentName = typeof details.agent_name === "string" ? details.agent_name : null
+  const agentFramework = typeof details.agent_framework === "string" ? details.agent_framework : null
+  if (agentName) {
+    return agentFramework && agentFramework !== agentName ? `${agentName} (${agentFramework})` : agentName
+  }
+
+  for (const key of ["name", "label", "id", "title"]) {
+    const v = details[key]
+    if (typeof v === "string" && v.length <= 60) return v
+  }
+
+  return null
+}
+
+function chipsForFieldValues(values: unknown[]): { label: string; raw: unknown }[] {
+  const seen = new Set<string>()
+  const result: { label: string; raw: unknown }[] = []
+  for (const v of values) {
+    const friendly = extractFriendlyLabel(v)
+    const label = friendly ?? formatSignalValue(v)
+    const truncated = label.length > 80 ? label.slice(0, 80) + "…" : label
+    if (seen.has(truncated)) continue
+    seen.add(truncated)
+    result.push({ label: truncated, raw: v })
+  }
+  return result
+}
+
 function DivergenceGroupItem({
   modelRouteId,
   magnitude,
@@ -150,31 +211,80 @@ function DivergenceGroupItem({
   fields: DifferingSetupField[]
   scoresByOrganization?: Record<string, number>
 }) {
+  const fieldLabels = fields.map((f) => formatFieldLabel(f.field))
+  const summarySentence =
+    fieldLabels.length === 0
+      ? `Reported scores diverge by ${formatSignalNumber(magnitude)}, above the ${formatSignalNumber(threshold)} threshold. The setup difference is not labelled.`
+      : `Reported scores diverge by ${formatSignalNumber(magnitude)} (threshold ${formatSignalNumber(threshold)}) because the runs differ on ${
+          fieldLabels.length === 1
+            ? fieldLabels[0]
+            : fieldLabels.slice(0, -1).join(", ") + " and " + fieldLabels[fieldLabels.length - 1]
+        }. The chips below show each variant.`
+
   return (
     <a
       href={`#row-${modelRouteId}`}
-      className="block rounded-xl border border-border/60 bg-background px-3 py-2 text-sm transition-colors hover:bg-muted/20"
+      className="block rounded-xl border border-border/60 bg-background px-3 py-2.5 text-sm transition-colors hover:bg-muted/20"
     >
       <div className="flex items-start justify-between gap-3">
         <div className="min-w-0">
-          <div className="font-medium">{modelRouteId}</div>
+          <div className="flex items-center gap-1.5">
+            <span className="font-medium">{modelRouteId}</span>
+            <SignalTooltip content={summarySentence}>
+              <Info className="h-3.5 w-3.5 shrink-0 cursor-help text-muted-foreground" />
+            </SignalTooltip>
+          </div>
           <div className="mt-1 text-xs text-muted-foreground">
-            Divergence {formatSignalNumber(magnitude)}; threshold {formatSignalNumber(threshold)}
+            Diverges by {formatSignalNumber(magnitude)} (threshold {formatSignalNumber(threshold)})
           </div>
         </div>
         <span className="shrink-0 text-xs font-medium text-primary">Jump to row</span>
       </div>
 
-      {fields.length > 0 && (
-        <div className="mt-2 space-y-1 text-xs text-muted-foreground">
-          {fields.slice(0, 3).map((field) => (
-            <div key={field.field}>
-              <span className="font-medium text-foreground">{formatFieldLabel(field.field)}:</span>{" "}
-              {field.values.map(formatSignalValue).join(", ")}
+      {fields.slice(0, 3).map((field) => {
+        const chips = chipsForFieldValues(field.values).slice(0, 6)
+        const overflow = field.values.length - chips.length
+        return (
+          <div key={field.field} className="mt-2">
+            <div className="mb-1 text-[11px] font-semibold uppercase tracking-[0.14em] text-muted-foreground">
+              Differs by {formatFieldLabel(field.field)}
             </div>
-          ))}
-        </div>
-      )}
+            <div className="flex flex-wrap gap-1.5">
+              {chips.map((chip, idx) => {
+                const friendly = extractFriendlyLabel(chip.raw)
+                const tooltipBody = friendly ? formatSignalValue(chip.raw) : null
+                const pill = (
+                  <span
+                    className="inline-flex max-w-[18rem] items-center rounded-full border border-border/60 bg-muted/20 px-2 py-0.5 text-[11px] text-foreground/90 truncate"
+                    title={!tooltipBody ? chip.label : undefined}
+                  >
+                    {chip.label}
+                  </span>
+                )
+                return tooltipBody ? (
+                  <SignalTooltip
+                    key={`${field.field}-${idx}`}
+                    content={
+                      <span className="block max-w-[24rem] break-all font-mono text-[10px] leading-snug">
+                        {tooltipBody}
+                      </span>
+                    }
+                  >
+                    {pill}
+                  </SignalTooltip>
+                ) : (
+                  <span key={`${field.field}-${idx}`}>{pill}</span>
+                )
+              })}
+              {overflow > 0 && (
+                <span className="inline-flex items-center rounded-full border border-dashed border-border/60 px-2 py-0.5 text-[11px] text-muted-foreground">
+                  +{overflow} more
+                </span>
+              )}
+            </div>
+          </div>
+        )
+      })}
 
       {scoresByOrganization && Object.keys(scoresByOrganization).length > 0 && (
         <div className="mt-2 flex flex-wrap gap-1.5">

components/signals/reproducibility-panel.tsx

@@ -22,7 +22,7 @@ export function ReproducibilityPanel({
     <div className="rounded-2xl border bg-background/70 p-4">
       <div className="mb-4 flex items-start gap-2">
         <AlertTriangle className="mt-0.5 h-4 w-4 shrink-0 text-amber-600 dark:text-amber-300" />
-        <div>
+        <div className="min-w-0 flex-1">
           <div className="font-semibold">
             {isResearchView ? "Reproducibility" : "Re-runnability"}
           </div>

components/signals/row-signals-compact.tsx

@@ -0,0 +1,98 @@
+"use client"
+
+import { AlertTriangle, ShieldCheck } from "lucide-react"
+import { useAudienceMode } from "@/components/audience-mode-provider"
+import type { RowAnnotations } from "@/lib/backend-artifacts"
+import { cn } from "@/lib/utils"
+import { SignalTooltip } from "./signal-tooltip"
+import { formatMissingField } from "./signal-utils"
+import { getRelationshipShortLabel } from "./provenance-badge"
+
+/**
+ * Compact, single-icon row signal indicator. Shows a warning icon when any
+ * row-level concern fires (reproducibility gap, first-party-only reporting),
+ * with a tooltip listing the specifics. Replaces the two full coloured badges
+ * that were taking up vertical space under every model name.
+ *
+ * Returns null when there are no concerns, so green rows stay quiet.
+ */
+export function RowSignalsCompact({
+  annotations,
+  className,
+  showWhenClean = false,
+}: {
+  annotations?: RowAnnotations | null
+  className?: string
+  showWhenClean?: boolean
+}) {
+  const { mode } = useAudienceMode()
+
+  if (!annotations) return null
+
+  const reproGap = annotations.reproducibility_gap
+  const provenance = annotations.provenance
+  const hasReproGap = reproGap?.has_reproducibility_gap === true
+  const firstPartyOnly = provenance?.first_party_only === true
+
+  const concerns: { title: string; detail: string }[] = []
+
+  if (hasReproGap && reproGap) {
+    const missing = reproGap.missing_fields.map(formatMissingField)
+    concerns.push({
+      title: mode === "policy" ? "Setup not documented" : "Reproducibility gap",
+      detail:
+        mode === "policy"
+          ? `${reproGap.populated_field_count} of ${reproGap.required_field_count} setup fields recorded. This score may be hard to re-run.`
+          : `Missing: ${missing.join(", ") || "none listed"}. ${reproGap.populated_field_count} of ${reproGap.required_field_count} setup fields recorded.`,
+    })
+  }
+
+  if (firstPartyOnly) {
+    concerns.push({
+      title:
+        mode === "policy" ? "Only model developer reported" : `${getRelationshipShortLabel("first_party", mode)} only`,
+      detail:
+        mode === "policy"
+          ? "Only the model developer reported this score; no independent replication is recorded."
+          : "First-party only — no independent replication is recorded for this group.",
+    })
+  }
+
+  if (concerns.length === 0) {
+    if (!showWhenClean) return null
+    return (
+      <SignalTooltip content="No row-level concerns flagged for this score.">
+        <span className={cn("inline-flex h-5 w-5 items-center justify-center", className)}>
+          <ShieldCheck className="h-3.5 w-3.5 text-emerald-600 dark:text-emerald-400" />
+        </span>
+      </SignalTooltip>
+    )
+  }
+
+  return (
+    <SignalTooltip
+      content={
+        <span className="block space-y-1.5">
+          <span className="block font-semibold">{concerns.length === 1 ? "1 concern" : `${concerns.length} concerns`}</span>
+          {concerns.map((c) => (
+            <span key={c.title} className="block">
+              <span className="font-medium">{c.title}.</span>{" "}
+              <span className="text-muted-foreground">{c.detail}</span>
+            </span>
+          ))}
+        </span>
+      }
+    >
+      <span
+        className={cn(
+          "inline-flex h-5 items-center gap-1 rounded-full bg-amber-50 px-1.5 text-amber-700 dark:bg-amber-950/40 dark:text-amber-200 cursor-help",
+          className
+        )}
+        aria-label={`${concerns.length} reporting concern${concerns.length === 1 ? "" : "s"} for this row`}
+      >
+        <AlertTriangle className="h-3 w-3" />
+        {concerns.length > 1 && <span className="text-[10px] font-semibold leading-none">{concerns.length}</span>}
+      </span>
+    </SignalTooltip>
+  )
+}

components/term.tsx

@@ -0,0 +1,47 @@
+"use client"
+
+import type { ReactNode } from "react"
+import { lookupTerm } from "@/lib/glossary"
+import { SignalTooltip } from "@/components/signals/signal-tooltip"
+
+interface TermProps {
+  /**
+   * Glossary key to look up. Defaults to the visible text.
+   */
+  term?: string
+  children: ReactNode
+  /**
+   * Override the tooltip body. If omitted, glossary entry is used.
+   */
+  explain?: ReactNode
+  className?: string
+}
+
+export function Term({ term, children, explain, className }: TermProps) {
+  const key = term ?? (typeof children === "string" ? children : undefined)
+  const entry = key ? lookupTerm(key) : undefined
+
+  const content = explain ?? (entry ? (
+    <span className="block space-y-1">
+      <span className="block">{entry.short}</span>
+      {entry.long ? <span className="block text-muted-foreground">{entry.long}</span> : null}
+    </span>
+  ) : null)
+
+  if (!content) {
+    return <span className={className}>{children}</span>
+  }
+
+  return (
+    <SignalTooltip content={content}>
+      <span
+        className={
+          "underline decoration-dotted decoration-muted-foreground/60 underline-offset-4 cursor-help " +
+          (className ?? "")
+        }
+      >
+        {children}
+      </span>
+    </SignalTooltip>
+  )
+}

lib/eval-processing.ts

@@ -117,6 +117,8 @@ export interface ModelResultForBenchmark {
   source_metadata: SourceMetadata
   source_data: BenchmarkEvaluation['source_data']
   result: EvaluationResult
+  /** URL to the underlying record JSON in the upstream HF dataset, when known. */
+  source_record_url?: string
   aggregate_components?: Array<{
     evaluation_id: string
     composite_benchmark_key: string

lib/glossary.ts

@@ -0,0 +1,136 @@
+export interface GlossaryEntry {
+  short: string
+  long?: string
+}
+
+const ENTRIES: Record<string, GlossaryEntry> = {
+  "0-shot": {
+    short: "Model answers without seeing any worked examples first.",
+    long: "The benchmark gives the model only the question. Tests how the model performs cold, without in-context demonstrations.",
+  },
+  "zero-shot": {
+    short: "Model answers without seeing any worked examples first.",
+  },
+  "1-shot": {
+    short: "Model sees one worked example before answering.",
+  },
+  "5-shot": {
+    short: "Model sees five worked examples before answering.",
+    long: "The five examples are included in the prompt as demonstrations. More shots usually raise scores; comparing a 0-shot result to a 5-shot result is not apples-to-apples.",
+  },
+  "few-shot": {
+    short: "Model sees a small number of worked examples before answering.",
+  },
+  "n-shot": {
+    short: "Number of worked examples shown to the model in the prompt.",
+  },
+  "pass@1": {
+    short: "The model gets one attempt; counted correct only if that single attempt passes.",
+    long: "Common in code benchmarks. Stricter than pass@k for k>1, since the model cannot retry.",
+  },
+  "pass@16": {
+    short: "The model gets 16 attempts; counted correct if any one of them passes.",
+    long: "Higher numbers tend to be inflated relative to pass@1 because the model has many tries.",
+  },
+  "pass@k": {
+    short: "The model gets k attempts; counted correct if any one of them passes.",
+  },
+  "majority voting": {
+    short: "The model answers many times; the most common answer is taken as the final answer.",
+    long: "Also called self-consistency. Tends to raise scores on reasoning tasks but costs more compute per question.",
+  },
+  "self-consistency": {
+    short: "The model answers many times; the most common answer is taken as the final answer.",
+  },
+  "best-of-n": {
+    short: "The model produces N answers; the best one (by a scorer or oracle) is reported.",
+    long: "Inflates scores compared to a single attempt. Worth flagging when comparing models.",
+  },
+  "exact match": {
+    short: "The model's answer is counted correct only if it matches the reference string exactly.",
+    long: "Strict — small formatting differences fail. Often used for short-answer tasks.",
+  },
+  "llm-as-judge": {
+    short: "Another language model grades the answers.",
+    long: "Faster than human grading but the judge model can have biases (length, style, self-preference). Worth knowing the judge identity.",
+  },
+  "llm as judge": {
+    short: "Another language model grades the answers.",
+  },
+  "human-in-the-loop": {
+    short: "Human graders score or check the model's answers.",
+  },
+  temperature: {
+    short: "Controls randomness in the model's output.",
+    long: "Lower (e.g. 0) is deterministic; higher (e.g. 1) is more varied. Affects scores — different temperatures make scores hard to compare.",
+  },
+  "top-p": {
+    short: "Limits the model to sampling from the most likely next tokens (nucleus sampling).",
+  },
+  "top-k": {
+    short: "Limits the model to sampling from the K most likely next tokens.",
+  },
+  perplexity: {
+    short: "How surprised the model is by the text. Lower is better.",
+  },
+  "average normalized score": {
+    short: "Each benchmark's score is rescaled to a 0–1 range, then averaged.",
+    long: "Lets you combine benchmarks that use different scales. The exact rescaling rule matters — check the methodology.",
+  },
+  agentic: {
+    short: "The model uses tools, takes multiple steps, and acts on an environment to complete a task.",
+  },
+  "agent budget": {
+    short: "The maximum compute, tokens, or steps the agent is allowed before it must stop.",
+  },
+  "f1 score": {
+    short: "A single number combining precision and recall. Higher is better.",
+  },
+  bleu: {
+    short: "A score for how close generated text is to a reference, by overlapping word sequences.",
+  },
+  rouge: {
+    short: "A score for how much a generated summary overlaps with a reference summary.",
+  },
+  benchmark: {
+    short: "A standardized test used to compare models on a specific capability.",
+  },
+}
+
+const KEY_ALIASES: Record<string, string> = {
+  "0 shot": "0-shot",
+  "1 shot": "1-shot",
+  "5 shot": "5-shot",
+  "few shot": "few-shot",
+  "n shot": "n-shot",
+  "zero shot": "zero-shot",
+  "best of n": "best-of-n",
+  "best-of-N": "best-of-n",
+  "LLM-as-judge": "llm-as-judge",
+  "LLM as judge": "llm as judge",
+  "self consistency": "self-consistency",
+  "Pass@1": "pass@1",
+  "Pass@16": "pass@16",
+  "Pass@K": "pass@k",
+  "F1": "f1 score",
+  "F1 score": "f1 score",
+  "BLEU": "bleu",
+  "ROUGE": "rouge",
+  "Agentic": "agentic",
+  "Temperature": "temperature",
+  "Perplexity": "perplexity",
+}
+
+export function lookupTerm(term: string): GlossaryEntry | undefined {
+  const trimmed = term.trim()
+  if (!trimmed) return undefined
+  const direct = ENTRIES[trimmed] ?? ENTRIES[trimmed.toLowerCase()]
+  if (direct) return direct
+  const aliased = KEY_ALIASES[trimmed] ?? KEY_ALIASES[trimmed.toLowerCase()]
+  if (aliased) return ENTRIES[aliased]
+  return undefined
+}
+
+export function hasTerm(term: string): boolean {
+  return lookupTerm(term) !== undefined
+}

lib/known-issues.ts

@@ -0,0 +1,47 @@
+import knownIssuesData from "@/metadata/benchmark_known_issues.json"
+
+export interface KnownIssue {
+  title: string
+  summary: string
+  severity: "info" | "warning" | "critical"
+  source_url?: string
+  published?: string
+}
+
+const RAW_ISSUES = (knownIssuesData as { issues?: Record<string, KnownIssue[]> }).issues ?? {}
+
+function normalizeKey(value: string): string {
+  return value
+    .toLowerCase()
+    .trim()
+    .replace(/[\s_\-/]+/g, " ")
+    .replace(/\s+/g, " ")
+}
+
+const NORMALIZED_LOOKUP = new Map<string, KnownIssue[]>()
+for (const [rawKey, issues] of Object.entries(RAW_ISSUES)) {
+  if (!Array.isArray(issues) || issues.length === 0) continue
+  NORMALIZED_LOOKUP.set(normalizeKey(rawKey), issues)
+  // Also index a hyphenated variant since some benchmark keys use dashes
+  NORMALIZED_LOOKUP.set(normalizeKey(rawKey).replace(/\s+/g, "-"), issues)
+}
+
+/**
+ * Look up curated known issues for a benchmark by trying any of the supplied
+ * names/keys (e.g. evaluation_name, composite_benchmark_key, family_key).
+ * Returns the first matching list, or an empty array if nothing is recorded.
+ */
+export function getKnownIssues(...candidates: Array<string | undefined | null>): KnownIssue[] {
+  for (const candidate of candidates) {
+    if (!candidate) continue
+    const key = normalizeKey(candidate)
+    const direct = NORMALIZED_LOOKUP.get(key)
+    if (direct) return direct
+    // Try collapsed (no spaces/dashes) form too, since some sources mash names together
+    const collapsed = key.replace(/\s+/g, "")
+    for (const [registered, issues] of NORMALIZED_LOOKUP) {
+      if (registered.replace(/\s+/g, "") === collapsed) return issues
+    }
+  }
+  return []
+}

lib/model-data.ts

@@ -749,6 +749,7 @@ function toModelResultsForMetric(
       },
       source_data: detail.source_data ?? { dataset_name: benchmarkKey },
       result: evaluationResult,
+      source_record_url: mr.source_record_url,
     }
   })
 }

metadata/benchmark_known_issues.json

@@ -0,0 +1,84 @@
+{
+  "_doc": "Curated known issues for specific benchmarks. Keyed by lowercased benchmark identifier. Severity is info | warning | critical. Each entry: title, summary, severity, optional source_url, optional published date (YYYY-MM-DD). This registry is for benchmark-specific concerns documented in academic work or by maintainers — not generic AI risks. Lookup is case-insensitive and tolerant of separator differences (spaces, dashes, underscores).",
+  "issues": {
+    "mmlu": [
+      {
+        "title": "Test-set contamination in pretraining corpora",
+        "summary": "Multiple analyses find MMLU questions and exact answers in publicly scraped pretraining data, inflating scores for models trained on that data.",
+        "severity": "warning",
+        "source_url": "https://arxiv.org/abs/2310.16787",
+        "published": "2023-10-25"
+      },
+      {
+        "title": "Errors and ambiguities in test items",
+        "summary": "Independent audits found mislabeled answers, ambiguous wording, and duplicated questions across MMLU subjects, capping reliable accuracy below 100%.",
+        "severity": "warning",
+        "source_url": "https://arxiv.org/abs/2406.04127"
+      }
+    ],
+    "mmlu-pro": [
+      {
+        "title": "Inherits MMLU contamination concerns for some subjects",
+        "summary": "MMLU-Pro reuses items from MMLU and other public sources; subjects overlapping with the MMLU origin may exhibit similar pretraining-data leakage.",
+        "severity": "info",
+        "source_url": "https://arxiv.org/abs/2406.01574"
+      }
+    ],
+    "gsm8k": [
+      {
+        "title": "Memorization on the test split",
+        "summary": "Studies show large models reproduce verbatim GSM8K test problems, suggesting training-time exposure rather than reasoning. Independent replication on perturbed variants (GSM-Symbolic, GSM-1k) gives lower scores.",
+        "severity": "warning",
+        "source_url": "https://arxiv.org/abs/2410.05229"
+      }
+    ],
+    "hellaswag": [
+      {
+        "title": "Annotation noise and outdated cultural references",
+        "summary": "A non-trivial fraction of HellaSwag items are mislabeled or rely on now-stale references, capping ceiling performance and complicating cross-year comparisons.",
+        "severity": "info"
+      }
+    ],
+    "humaneval": [
+      {
+        "title": "Limited and saturated test set",
+        "summary": "Only 164 problems; top models exceed 90% pass@1 and the benchmark is widely viewed as saturated. Use newer code benchmarks (LiveCodeBench, SWE-bench) for current capability claims.",
+        "severity": "warning"
+      },
+      {
+        "title": "Test-case adequacy",
+        "summary": "HumanEval+ found that the original tests miss many incorrect solutions; reported scores can overstate real correctness.",
+        "severity": "info",
+        "source_url": "https://arxiv.org/abs/2305.01210"
+      }
+    ],
+    "narrativeqa": [
+      {
+        "title": "Largely deprecated in current evaluations",
+        "summary": "Few recent leading models report NarrativeQA. Treat scores as historical context rather than current capability evidence.",
+        "severity": "info"
+      }
+    ],
+    "boolq": [
+      {
+        "title": "Largely saturated",
+        "summary": "Top models score above 90%; small differences are noise. Use as a sanity check, not a discriminator.",
+        "severity": "info"
+      }
+    ],
+    "ifeval": [
+      {
+        "title": "Narrow instruction taxonomy",
+        "summary": "IFEval covers a fixed set of verifiable instruction types. High scores do not generalise to open-ended instruction following.",
+        "severity": "info"
+      }
+    ],
+    "humanity's last exam": [
+      {
+        "title": "Marketing framing vs. methodological scope",
+        "summary": "The name implies a definitive ceiling for AI capability; the underlying benchmark is a 3,000-item exam in selected academic disciplines and does not measure general intelligence or safety. Read scores in context.",
+        "severity": "warning"
+      }
+    ]
+  }
+}