Merge cross-source benchmark families; tidy leaderboard panel + table chrome

clean-hierarchy:
* New `groupSameBenchAcrossSources` pass collapses N≥2 single-bench
families that share a bench key but ship disjoint eval_summary_ids
(i.e. independent sources publishing the same conceptual benchmark)
into a single family card. Survivor = most models; other benches
are moved in as siblings with display_name " · <Source>" suffixed
from metrics[0].sources, and a unique key suffix
(mmlu-pro__arcadia-impact) so siblings don't collide. Runs AFTER
decorateHierarchyDerivedTags so the sanitizer's shareToken /
humanizeKey passes don't clobber the suffixed names. Net effect:
one MMLU-Pro card with two source-tagged benches.

eval-detail:
* Drop the standalone bar column (header repeated the eval name, the
bar repeated the score). Inline a thin perf bar under the score
cell with the setup / dataset caption right-aligned beneath.
* Drop the duplicate ApplesToApplesBanner from inside
MultiMetricLeaderboard — the parent already renders it.
* Render the multi-metric SplitPicker INSIDE MultiMetricLeaderboard
(after the section heading + description) so it matches the
position of the single-metric branch's Split picker. Both fibble
and global-mmlu now render the dropdown in the same spot.
* Restore the subtask Split picker for evals like Global MMLU Lite
(1 root metric + N subtasks) but keep it gated by !splitConfig so
fibble (with page-level Split) doesn't double up.
* Wording: rename remaining "Slice" UI strings to "Split" (Slice
breakdown → Split breakdown, "metric spec · ... · slices" →
"splits") to match the rest of the app.
* getMetricChipLabel walks display_name → metric_name → metric_id →
column_key tail and humanises underscores so AgentHarm's null-
display-name metrics surface as "avg full score" etc. instead of
four chips literally labelled "Metric".

family-table:
* "Reported results" column now divides total_results by the family's
bench-level metric count, giving a unique-models-tested
approximation. AgentHarm's 72 metric-rows / 4 metrics → 18 unique
models surfaced instead of the misleading raw 72.
* Drop the dead totalModels prop and unused visibleLeaves local.

score-distribution:
* View toggle (Distribution / Frontier) reverts to ec-pill chips —
same chip style as the metric row below; row labels (VIEW vs
METRIC) carry the differentiation.

models page:
* Default sort changes from Released ↓ to Coverage ↓ (and the Reset
Filters button resets to Coverage ↓).

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

app/evals/page.tsx

@@ -312,7 +312,6 @@ function EvalsPageInner() {
         ) : (
           <FamilyTable
             families={visibleFamilies}
-            totalModels={totalModels}
             evalItems={evalItems}
             benchmarkCards={benchmarkCards}
             categoryFilter={new Set(selectedCategories)}

app/models/page.tsx

@@ -58,7 +58,7 @@ export default function ModelsPage() {
   const [loadingDevelopers, setLoadingDevelopers] = useState(false)
   const [developersReady, setDevelopersReady] = useState(false)
   const [groupByDeveloper, setGroupByDeveloper] = useState(false)
-  const [modelSortBy, setModelSortBy] = useState<ModelSort>("released")
+  const [modelSortBy, setModelSortBy] = useState<ModelSort>("coverage")
   const [modelSortDir, setModelSortDir] = useState<SortDir>("desc")
   const [developerSortBy, setDeveloperSortBy] = useState<DevSort>("models")
   const [developerSortDir, setDeveloperSortDir] = useState<SortDir>("desc")
@@ -383,7 +383,7 @@ export default function ModelsPage() {
               className="btn-ec outline"
               onClick={() => {
                 setSearchQuery("")
-                setModelSortBy("released")
+                setModelSortBy("coverage")
                 setModelSortDir("desc")
                 setDeveloperSortBy("models")
                 setDeveloperSortDir("desc")

components/eval-detail.tsx

@@ -624,7 +624,14 @@ export function EvalDetail({
     return Array.from(seen, ([key, name]) => ({ key, label: name }))
   }, [lb.leaderboard_metrics])
 
-  const hasSlicePicker = !hasMultiMetricLeaderboard && subtaskSlices.length > 1
+  // Suppress the slice picker when a split picker is already in play.
+  // For evals like Fibble Arena both pickers partition the same axis
+  // (each split is one of the per-lie variants; each slice is the
+  // matrix-backfilled subtask for the same per-lie variant), so showing
+  // both reads as a redundant control. The page-level split is more
+  // authoritative — it loads richer per-eval data — so it wins.
+  const hasSlicePicker =
+    !hasMultiMetricLeaderboard && subtaskSlices.length > 1 && !splitConfig
 
   const ALL_SLICE_KEY = "__all__"
   const [activeSlice, setActiveSlice] = useState<string>(ALL_SLICE_KEY)
@@ -870,7 +877,7 @@ export function EvalDetail({
                 className="font-mono text-[10px] uppercase tracking-[0.12em]"
                 style={{ color: "var(--fg-subtle)" }}
               >
-                metric spec · completeness · comparability{summary.subtasks?.length ? " · slices" : ""}
+                metric spec · completeness · comparability{summary.subtasks?.length ? " · splits" : ""}
               </span>
             </div>
             {overviewOpen ? (
@@ -992,7 +999,7 @@ export function EvalDetail({
                       className="font-mono uppercase mb-1"
                       style={{ fontSize: 10, letterSpacing: "0.14em", color: "var(--fg-subtle)" }}
                     >
-                      Slice breakdown · {summary.subtasks.length}
+                      Split breakdown · {summary.subtasks.length}
                     </div>
                     <ul
                       className="flex flex-col"
@@ -1050,10 +1057,11 @@ export function EvalDetail({
             summary={lb.comparability_summary}
             detailsAnchorId="comparability-panel"
           />
-          {splitConfig && (
-            <SplitPicker config={splitConfig} className="mb-4" />
-          )}
-          <MultiMetricLeaderboard summary={lb} isResearchView={isResearchView} />
+          <MultiMetricLeaderboard
+            summary={lb}
+            isResearchView={isResearchView}
+            splitConfig={splitConfig}
+          />
         </section>
       ) : (
         <section>
@@ -1090,11 +1098,16 @@ export function EvalDetail({
             <SplitPicker config={splitConfig} className="mb-4" />
           )}
 
+          {/* Subtask split picker for evals like Global MMLU Lite where
+              the splits live as subtasks of a single eval (not as
+              separate eval IDs the page-level SplitPicker can swap to).
+              Suppressed when a page-level split is already in play —
+              two pickers would partition the same axis (see fibble). */}
           {hasSlicePicker && (
             <SplitPicker
               className="mb-4"
               config={{
-                label: "Slice",
+                label: "Split",
                 activeId: activeSlice,
                 onChange: setActiveSlice,
                 options: [
@@ -1156,12 +1169,7 @@ export function EvalDetail({
                   <th className="hidden lg:table-cell" style={{ minWidth: 160 }}>
                     {isResearchView ? "Developer" : "Provider"}
                   </th>
-                  <th className="hidden md:table-cell" style={{ minWidth: 220 }}>
-                    {lb.composite_benchmark_name && lb.composite_benchmark_name !== lb.evaluation_name
-                      ? `${lb.composite_benchmark_name} · ${lb.evaluation_name}`
-                      : lb.evaluation_name}
-                  </th>
-                  <th className="num" style={{ width: 130 }}>
+                  <th className="num" style={{ minWidth: 200 }}>
                     {lb.metric_config.unit ?? "Score"}
                   </th>
                   <th className="hidden lg:table-cell" style={{ width: 110 }}>Evaluator</th>
@@ -1289,51 +1297,54 @@ export function EvalDetail({
                           </div>
                         </td>
 
-                        <td className="hidden md:table-cell align-top">
-                          {/* Performance bar with shot/setup caption */}
-                          <div className="min-w-[200px] py-0.5">
+                        <td className="num align-top">
+                          {/* Score with inline performance bar so the
+                              previously-dedicated bar column can be
+                              dropped — its only purpose was visualising
+                              this same number. Caption shows shot/CoT
+                              setup or a differing dataset name when
+                              available; otherwise it's omitted. */}
+                          <div className="flex items-baseline justify-end gap-2 tabular-nums" style={{ fontSize: 15, fontWeight: 600 }}>
+                            <span>{formatRawScore(modelResult.score, undefined)}</span>
+                          </div>
+                          <div
+                            className="mt-1 hidden md:block"
+                            style={{
+                              position: "relative",
+                              height: 4,
+                              background: "var(--bg-surface)",
+                              overflow: "hidden",
+                            }}
+                          >
                             <div
                               style={{
-                                position: "relative",
-                                height: 6,
-                                background: "var(--bg-surface)",
-                                overflow: "hidden",
+                                position: "absolute",
+                                inset: 0,
+                                width: `${Math.max(2, normalizedScore * 100)}%`,
+                                background: isTopRank ? "var(--accent)" : "var(--fg-muted)",
+                                opacity: isTopRank ? 1 : 0.55,
                               }}
+                            />
+                          </div>
+                          {setupLabel && (
+                            <div
+                              className="mt-1 font-mono uppercase truncate text-right"
+                              style={{ fontSize: 10, letterSpacing: "0.06em", color: "var(--fg-subtle)" }}
                             >
-                              <div
-                                style={{
-                                  position: "absolute",
-                                  inset: 0,
-                                  width: `${Math.max(2, normalizedScore * 100)}%`,
-                                  background: isTopRank ? "var(--accent)" : "var(--fg-muted)",
-                                  opacity: isTopRank ? 1 : 0.55,
-                                }}
-                              />
+                              {setupLabel}
                             </div>
-                            {setupLabel && (
+                          )}
+                          {!setupLabel &&
+                            datasetName &&
+                            !isResearchView &&
+                            datasetName !== lb.evaluation_name && (
                               <div
-                                className="mt-1 font-mono uppercase truncate"
-                                style={{ fontSize: 10, letterSpacing: "0.06em", color: "var(--fg-subtle)" }}
+                                className="mt-1 font-mono truncate text-right"
+                                style={{ fontSize: 10, color: "var(--fg-subtle)" }}
                               >
-                                {setupLabel}
+                                {datasetName}
                               </div>
                             )}
-                            {!setupLabel &&
-                              datasetName &&
-                              !isResearchView &&
-                              datasetName !== lb.evaluation_name && (
-                                <div
-                                  className="mt-1 font-mono truncate"
-                                  style={{ fontSize: 10, color: "var(--fg-subtle)" }}
-                                >
-                                  {datasetName}
-                                </div>
-                              )}
-                          </div>
-                        </td>
-
-                        <td className="num align-top tabular-nums" style={{ fontSize: 15, fontWeight: 600 }}>
-                          {formatRawScore(modelResult.score, undefined)}
                         </td>
 
                         <td className="hidden lg:table-cell align-top">
@@ -1387,7 +1398,7 @@ export function EvalDetail({
 
                       {isExpanded && (
                         <tr>
-                          <td colSpan={hasAnyUpdatedTimestamp ? 8 : 7} style={{ background: "var(--bg-warm)", padding: 0 }}>
+                          <td colSpan={hasAnyUpdatedTimestamp ? 7 : 6} style={{ background: "var(--bg-warm)", padding: 0 }}>
                             <div className="space-y-5 px-4 py-5 sm:px-6">
                               <div className="grid gap-4 xl:grid-cols-3">
                                 <DetailPanel
@@ -1530,13 +1541,13 @@ export function EvalDetail({
                                     className="font-mono uppercase"
                                     style={{ fontSize: 10, letterSpacing: "0.14em", color: "var(--fg-subtle)" }}
                                   >
-                                    Slice breakdown
+                                    Split breakdown
                                   </div>
                                   <div className="overflow-x-auto" style={{ border: "1px solid var(--border-soft)" }}>
                                     <table className="ec-htable">
                                       <thead>
                                         <tr>
-                                          <th>Slice</th>
+                                          <th>Split</th>
                                           <th className="num">Raw</th>
                                         </tr>
                                       </thead>
@@ -1669,7 +1680,7 @@ export function EvalDetail({
                 })}
                 {leaderboardRows.length === 0 && (
                   <tr>
-                    <td colSpan={hasAnyUpdatedTimestamp ? 8 : 7} style={{ padding: "32px 16px", textAlign: "center", color: "var(--fg-muted)" }}>
+                    <td colSpan={hasAnyUpdatedTimestamp ? 7 : 6} style={{ padding: "32px 16px", textAlign: "center", color: "var(--fg-muted)" }}>
                       No leaderboard entries match the selected parameter range.
                     </td>
                   </tr>
@@ -1706,9 +1717,11 @@ export function EvalDetail({
 function MultiMetricLeaderboard({
   summary,
   isResearchView,
+  splitConfig,
 }: {
   summary: BenchmarkEvalSummary
   isResearchView: boolean
+  splitConfig?: SplitConfig
 }) {
   const [page, setPage] = useState(1)
   // Default sort: the first root-scope metric (the benchmark's overall
@@ -1984,10 +1997,9 @@ function MultiMetricLeaderboard({
 
   return (
     <section>
-      <ApplesToApplesBanner
-        summary={summary.comparability_summary}
-        detailsAnchorId="comparability-panel"
-      />
+      {/* The parent EvalDetail already renders the apples-to-apples
+          banner before this leaderboard section — duplicating it here
+          made the box appear twice on multi-metric evals like fibble. */}
       <div className="section-head">
         <h2>{isResearchView ? "Leaderboard" : "Reporting Comparison"}</h2>
         <span
@@ -2049,6 +2061,10 @@ function MultiMetricLeaderboard({
         </DropdownMenu>
       </div>
 
+      {splitConfig && (
+        <SplitPicker config={splitConfig} className="mb-4" />
+      )}
+
       {/* Distribution panel — one curve, dropdown swaps between metrics */}
       {(() => {
         const distSeries = visibleMetrics
@@ -2086,27 +2102,6 @@ function MultiMetricLeaderboard({
       })()}
 
       <div className="ec-card" style={{ padding: 0, overflow: "hidden" }}>
-        {hasSliceTabs && (
-          <div className="border-b bg-background px-5 py-3 sm:px-6 flex items-center gap-3">
-            <span
-              className="font-mono uppercase tracking-[0.14em] shrink-0"
-              style={{ fontSize: 10, color: "var(--fg-subtle)" }}
-            >
-              Slice
-            </span>
-            <select
-              className="ec-select"
-              value={activeSliceTab}
-              onChange={(e) => setActiveSliceTab(e.target.value)}
-            >
-              <option value="all">Overall</option>
-              {sliceTabs.map((tab) => (
-                <option key={tab.key} value={tab.key}>{tab.label}</option>
-              ))}
-            </select>
-          </div>
-        )}
-
         {hasParameterData && (
           <div className="border-b bg-background px-5 py-4 sm:px-6">
             <ParamRangePicker

components/family-table.tsx

@@ -16,7 +16,6 @@ export type FamilySortCol = "name" | "benchmarks" | "results"
 
 interface FamilyTableProps {
   families: HierarchyFamily[]
-  totalModels: number
   evalItems?: Map<string, BenchmarkEvalListItem>
   benchmarkCards?: Record<string, BenchmarkCard>
   domainFilter?: Set<string> | null
@@ -197,7 +196,6 @@ interface RowData {
 
 export function FamilyTable({
   families,
-  totalModels,
   evalItems,
   benchmarkCards,
   domainFilter,
@@ -299,13 +297,38 @@ export function FamilyTable({
         if (overview) description = overview.length > 140 ? overview.slice(0, 137) + "…" : overview
       }
 
+      // "Reported results" approximates the unique models tested in the
+      // family. provenance_summary.total_results counts every (model,
+      // metric) report row, so families that publish many metrics inflate
+      // the number — agentharm's 4 declared metrics turned 18 unique
+      // models into 72 result rows. Normalise by the family's metric
+      // count (sum of bench-level metric entries; clamp to 1 to avoid
+      // div-by-zero on families whose metrics array is empty).
+      const totalResults =
+        fam.provenance_summary?.total_results ??
+        fam.reproducibility_summary?.results_total ??
+        fam.evals_count ??
+        metricCount
+      const allBenchesForMetricCount: HierarchyBenchmark[] = [
+        ...(fam.benchmarks ?? []),
+        ...(fam.standalone_benchmarks ?? []),
+        ...(fam.composites ?? []).flatMap((c) => c.benchmarks ?? []),
+      ]
+      const familyMetricCount = allBenchesForMetricCount.reduce(
+        (n, b) => n + Math.max((b.metrics ?? []).length, 1),
+        0,
+      )
+      const reportedResults =
+        familyMetricCount > 0
+          ? Math.round(totalResults / familyMetricCount)
+          : totalResults
       out.push({
         key: fam.key,
         name: displayName,
         keySlug: fam.key,
         tags: fam.derivedTags ?? [],
         benchmarks: benchmarkCount,
-        evalsCount: fam.evals_count ?? metricCount,
+        evalsCount: reportedResults,
         leaves: visibleLeafEntries,
         sections,
         description,
@@ -372,9 +395,6 @@ export function FamilyTable({
               ? expanded[row.key] ?? true
               : expanded[row.key] ?? false
             const allLeaves = row.sections.flatMap((s) => s.leaves)
-            const visibleLeaves = isExpanded
-              ? allLeaves.slice(0, LEAVES_INLINE_MAX)
-              : []
             const hiddenLeafCount = isExpanded
               ? Math.max(allLeaves.length - LEAVES_INLINE_MAX, 0)
               : 0
@@ -461,9 +481,6 @@ export function FamilyTable({
                   <td className="num font-mono text-[13px]">{row.benchmarks.toLocaleString()}</td>
                   <td className="num font-mono text-[13px]">
                     {row.evalsCount.toLocaleString()}
-                    {totalModels > 0 && (
-                      <span className="text-[color:var(--fg-subtle)]"> / {totalModels.toLocaleString()}</span>
-                    )}
                   </td>
                   <td>
                     <span className="font-mono text-[10px] tracking-[0.12em] uppercase text-[color:var(--accent)] inline-flex items-center gap-1">

components/score-distribution.tsx

@@ -284,22 +284,32 @@ export function ScoreDistribution({
                 {canShowFrontier ? "View" : "Score distribution"}
               </span>
               {canShowFrontier && (
-                <div className="flex items-center gap-1">
-                  <button
-                    type="button"
-                    className={`ec-pill${effectiveView === "distribution" ? " on" : ""}`}
-                    onClick={() => setView("distribution")}
-                  >
-                    Distribution
-                  </button>
-                  <button
-                    type="button"
-                    className={`ec-pill${effectiveView === "frontier" ? " on" : ""}`}
-                    onClick={() => setView("frontier")}
-                    title="Frontier score over model release dates (cumulative best)."
-                  >
-                    Frontier
-                  </button>
+                <div
+                  role="tablist"
+                  aria-label="Chart view"
+                  className="inline-flex items-center gap-1"
+                >
+                  {(["distribution", "frontier"] as const).map((view) => {
+                    const on = effectiveView === view
+                    const label = view === "distribution" ? "Distribution" : "Frontier"
+                    return (
+                      <button
+                        key={view}
+                        type="button"
+                        role="tab"
+                        aria-selected={on}
+                        onClick={() => setView(view)}
+                        title={
+                          view === "frontier"
+                            ? "Frontier score over model release dates (cumulative best)."
+                            : "Kernel-density distribution of model scores."
+                        }
+                        className={`ec-pill${on ? " on" : ""}`}
+                      >
+                        {label}
+                      </button>
+                    )
+                  })}
                 </div>
               )}
             </div>

lib/clean-hierarchy.ts

@@ -214,6 +214,12 @@ export function cleanHierarchy(
     dedupAggregatorBenchesByScore(h, comparisonIndex)
   }
   decorateHierarchyDerivedTags(h)
+  // Run AFTER the sanitizer so it can't clobber our suffixed bench
+  // display_names (`MMLU-Pro · Arcadia Impact` etc.) — the merged-in
+  // bench keys deliberately use a non-shareToken-matching form to
+  // keep them distinct, which would otherwise trip the humanizeKey
+  // fallback in benchmark-tags.sanitizeName.
+  groupSameBenchAcrossSources(h)
   if (h.benchmark_index) {
     const survivingFamilyKeys = new Set<string>(
       (h.families ?? []).map((f) => f.key),
@@ -891,6 +897,101 @@ function consolidateDedicatedHomeBenchmarks(h: CleanableHierarchy) {
   h.families = allFamilies.filter((fam) => !dropped.has(fam))
 }
 
+/**
+ * Group single-bench families that publish the same conceptual benchmark
+ * from different upstream sources into one merged family card.
+ *
+ * Triggered when ≥2 single-bench families share a bench key but their
+ * bench rows have non-overlapping eval_summary_ids — i.e. independent
+ * sources publishing the same benchmark. The richest family (most
+ * models) keeps its slot; other families contribute their bench under
+ * the survivor as siblings, with each bench's display_name suffixed
+ * with " · <Source>" so the user can tell which run a row came from.
+ *
+ * Runs AFTER decorateHierarchyDerivedTags so the sanitizer's
+ * shareToken / humanizeKey passes don't clobber the suffixed names
+ * (the merged-in bench keys like `mmlu-pro__arcadia` deliberately
+ * don't share tokens with "MMLU-Pro · Arcadia Impact").
+ */
+function groupSameBenchAcrossSources(h: CleanableHierarchy) {
+  const sourceLabel = (
+    bench: HierarchyBenchmark,
+    fallback: string,
+  ): string => {
+    const sources = (bench.metrics ?? []).flatMap((m) => m.sources ?? [])
+    for (const s of sources) {
+      const trimmed = String(s ?? "").trim()
+      if (trimmed) return trimmed
+    }
+    return fallback
+  }
+  const slugifyShort = (s: string) =>
+    s.toLowerCase().replace(/[^a-z0-9]+/g, "-").replace(/^-+|-+$/g, "")
+
+  type Handle = { fam: HierarchyFamily; bench: HierarchyBenchmark }
+  const collectBenches = (fam: HierarchyFamily): HierarchyBenchmark[] => [
+    ...(fam.benchmarks ?? []),
+    ...(fam.standalone_benchmarks ?? []),
+    ...(fam.composites ?? []).flatMap((c) => c.benchmarks ?? []),
+  ]
+  const candidatesByKey = new Map<string, Handle[]>()
+  for (const fam of h.families ?? []) {
+    const benches = collectBenches(fam)
+    if (benches.length !== 1) continue
+    const sole = benches[0]
+    const list = candidatesByKey.get(sole.key) ?? []
+    list.push({ fam, bench: sole })
+    candidatesByKey.set(sole.key, list)
+  }
+
+  const dropped = new Set<HierarchyFamily>()
+  for (const [, group] of candidatesByKey) {
+    if (group.length < 2) continue
+    // eval_ids must be disjoint (otherwise an earlier rule should have
+    // caught them as aliases of the same row).
+    const seenIds = new Set<string>()
+    let disjoint = true
+    for (const entry of group) {
+      for (const id of entry.bench.summary_eval_ids ?? []) {
+        if (seenIds.has(id)) { disjoint = false; break }
+        seenIds.add(id)
+      }
+      if (!disjoint) break
+    }
+    if (!disjoint) continue
+
+    const sortedGroup = [...group].sort((x, y) => {
+      const xModels = x.bench.metrics?.[0]?.models_count ?? 0
+      const yModels = y.bench.metrics?.[0]?.models_count ?? 0
+      if (xModels !== yModels) return yModels - xModels
+      return x.fam.key.localeCompare(y.fam.key)
+    })
+    const survivor = sortedGroup[0]
+    const baseDisplay =
+      survivor.bench.display_name?.trim() ||
+      survivor.fam.display_name?.trim() ||
+      survivor.bench.key
+
+    for (const entry of sortedGroup) {
+      const src = sourceLabel(
+        entry.bench,
+        entry.fam.display_name || entry.fam.key,
+      )
+      entry.bench.display_name = `${baseDisplay} · ${src}`
+      if (entry !== survivor) {
+        entry.bench.key = `${survivor.bench.key}__${slugifyShort(src) || slugifyShort(entry.fam.key)}`
+        survivor.fam.benchmarks = survivor.fam.benchmarks ?? []
+        survivor.fam.benchmarks.push(entry.bench)
+        dropped.add(entry.fam)
+      }
+    }
+    survivor.fam.display_name = baseDisplay
+  }
+
+  if (dropped.size === 0) return
+  h.families = (h.families ?? []).filter((fam) => !dropped.has(fam))
+}
+
 /**
  * AIR-Bench 2024 is a single safety benchmark with a 4-tier taxonomy
  * (314 leaf risk categories per the spec, ~30 of which the HELM