Merge corpus dashboard into home as paper-aligned landing

The home page now renders the four interpretive signals (paper §4.2.1)
as a compact 4-tile strip alongside corpus stats from the rollout
hierarchy (paper §3.2). The standalone /corpus page redirects to home
and the Corpus tab is removed from navigation. Reproducibility headline
is inverted (% documented rather than % with gaps) for impact while the
paper-aligned gap rate stays in the detail line.

About page is rewritten to mirror the paper's structure: six-level
rollout hierarchy, four interpretive signals, two reader modes, and
composition with EEE + Auto-BenchmarkCards.

CorpusDashboard now accepts an `embedded` prop so callers can host the
full drill-down without its self-contained masthead. Manifest type is
extended with the count fields the upstream pipeline already emits so
the home page reads counts directly instead of pulling model-cards-lite.

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

app/about/page.tsx

@@ -1,5 +1,15 @@
 import Link from "next/link"
-import { ArrowRight, BookOpenText, Database, Scale, Search } from "lucide-react"
+import {
+  ArrowRight,
+  BookOpenText,
+  ClipboardCheck,
+  Database,
+  GitCompareArrows,
+  Layers,
+  Scale,
+  ShieldCheck,
+  Users,
+} from "lucide-react"
 
 import { Navigation } from "@/components/navigation"
 import { PageHeader } from "@/components/page-header"
@@ -13,148 +23,212 @@ export default function AboutPage() {
       <PageHeader
         eyebrow="About"
         title="About Eval Cards"
-        description="Why public AI evaluation reporting needs shared infrastructure, and what this project builds toward it."
+        description="An interpretative integration layer for AI evaluation reporting. Eval Cards composes existing evaluation infrastructure into a single reading surface, organizes reported scores through a six-level rollout hierarchy, and surfaces four interpretive signals that help readers decide whether to trust a reported result."
         size="wide"
         metaItems={[
-          { label: "Modes", value: "Research + Policy" },
-          { label: "Core", value: "Reporting + Schema + Platform" },
+          { label: "Reader modes", value: "Research + Policy" },
+          { label: "Signal layers", value: "Reproducibility · Completeness · Provenance · Comparability" },
         ]}
       />
 
-      <main className="mx-auto w-full max-w-[92rem] px-4 py-10 sm:px-6 lg:px-8">
-        <div className="space-y-16">
+      <main className="mx-auto w-full max-w-[88rem] px-4 py-12 sm:px-6 lg:px-8">
+        <div className="space-y-14">
+          {/* The framing problem */}
           <section className="grid gap-8 border-b border-border/60 pb-12 lg:grid-cols-[minmax(220px,0.34fr)_minmax(0,1fr)] lg:gap-12">
             <div className="space-y-3">
-              <SectionLabel label="Why this exists" />
-              <h2 className="max-w-sm text-3xl font-semibold tracking-[-0.03em] text-foreground sm:text-[2.2rem]">
-                Evaluation scores travel faster than the context needed to interpret them.
+              <SectionLabel label="The problem" />
+              <h2 className="max-w-sm text-2xl font-semibold tracking-[-0.03em] text-foreground sm:text-[1.75rem]">
+                Evaluation results travel faster than the context needed to interpret them.
               </h2>
             </div>
 
-            <div className="grid gap-6 xl:grid-cols-[minmax(0,1.2fr)_minmax(260px,0.8fr)]">
-              <div className="space-y-4 text-base leading-8 text-muted-foreground">
-                <p>
-                  AI evaluations shape discussions about capability, safety, and deployment. Public reporting arrives through papers, leaderboards, blog posts, and framework outputs that rarely line up.
-                </p>
-                <p>
-                  Eval Cards treats reporting as infrastructure. Beyond collecting scores, it surfaces who reported a result, when it was run, what slice it covers, and whether the setup makes comparison reasonable.
-                </p>
-              </div>
-
-              <div className="rounded-[1.75rem] border border-border/70 bg-muted/20 p-5">
-                <div className="text-[11px] font-semibold uppercase tracking-[0.18em] text-muted-foreground">
-                  Reading stance
-                </div>
-                <div className="mt-4 space-y-4">
-                  <SignalLine icon={<Search className="h-4 w-4" />} label="Inspect provenance alongside rankings." tone="sky" />
-                  <SignalLine icon={<Scale className="h-4 w-4" />} label="Surface comparability caveats early." tone="amber" />
-                  <SignalLine icon={<BookOpenText className="h-4 w-4" />} label="Support research and policy use without duplicating the record." tone="emerald" />
-                </div>
-              </div>
+            <div className="space-y-4 text-base leading-7 text-muted-foreground">
+              <p>
+                AI evaluations are produced at scale and reported through papers, leaderboards, model cards, and
+                framework outputs that rarely share conventions. The cost of fragmentation is paid at the point
+                of interpretation: provenance gets lost, configurations differ silently, and incomplete reporting
+                is treated indistinguishably from complete reporting.
+              </p>
+              <p>
+                Eval Cards composes three existing sources into a single record — Auto-BenchmarkCards for
+                benchmark metadata, the EEE schema for evaluation run data, and voluntary developer disclosure
+                for fields neither captures — then exposes interpretive signals over the result.
+              </p>
             </div>
           </section>
 
-          <section className="space-y-5 border-b border-border/60 pb-12">
-            <div className="grid gap-4 lg:grid-cols-[minmax(220px,0.34fr)_minmax(0,1fr)] lg:items-start">
-              <div className="space-y-3">
-                <SectionLabel label="Core contribution" />
-                <h2 className="text-3xl font-semibold tracking-[-0.03em] text-foreground">
-                  Three layers, one public interface.
-                </h2>
-              </div>
-              <p className="max-w-3xl text-base leading-8 text-muted-foreground">
-                A reporting framework, a standardized data layer, and an interface for exploring the evidence across models, evaluations, and benchmarks.
+          {/* Six-level rollout hierarchy */}
+          <section className="grid gap-8 border-b border-border/60 pb-12 lg:grid-cols-[minmax(220px,0.34fr)_minmax(0,1fr)] lg:gap-12">
+            <div className="space-y-3">
+              <SectionLabel label="Rollout hierarchy" />
+              <h2 className="text-2xl font-semibold tracking-[-0.03em] text-foreground sm:text-[1.75rem]">
+                Six levels from a benchmark family down to a single metric.
+              </h2>
+              <p className="max-w-md text-sm leading-7 text-muted-foreground">
+                Reports are not flat (model, benchmark, score) triples. Every score resolves to an explicit
+                path through this hierarchy, which makes drill-down and apples-to-apples comparison possible.
               </p>
             </div>
 
-            <div className="grid gap-4 xl:grid-cols-3">
-              <AboutPanel
-                icon={<BookOpenText className="h-4 w-4" />}
-                tone="amber"
-                title="Reporting framework"
-                body="A structured view of which evaluation details matter for interpretation, reproducibility, and public accountability."
+            <ol className="space-y-3">
+              <RollupRow
+                index="1"
+                title="Family"
+                body="A related collection sharing a common object of measurement or methodological lineage (e.g., the SWE-bench family, the MMLU family)."
               />
-              <AboutPanel
-                icon={<Database className="h-4 w-4" />}
-                tone="sky"
-                title="Standardized records"
-                body="Evaluation outputs and benchmark metadata are shaped into a consistent layer that can be queried and compared across sources."
+              <RollupRow
+                index="2"
+                title="Suite"
+                body="A named composite reporting unit that aggregates multiple benchmarks under a unified presentation (e.g., Open LLM Leaderboard v2, HELM Instruct)."
               />
-              <AboutPanel
-                icon={<Scale className="h-4 w-4" />}
-                tone="emerald"
-                title="Public reading interface"
-                body="The interface keeps setup differences, provenance, and evidence gaps visible so interpretation does not detach from reporting conditions."
+              <RollupRow
+                index="3"
+                title="Single benchmark"
+                body="An individual evaluation with a defined dataset and scoring method (e.g., GSM8K, IFEval, MMLU-Pro)."
               />
-            </div>
+              <RollupRow
+                index="4"
+                title="Split"
+                body="A named partition of a benchmark's item set (e.g., test, validation, or a language-specific subset)."
+              />
+              <RollupRow
+                index="5"
+                title="Subtask"
+                body="A capability- or construct-level decomposition within a benchmark (e.g., algebra within MATH)."
+              />
+              <RollupRow
+                index="6"
+                title="Metric"
+                body="The specific scoring rule attached to a result (e.g., pass@1, accuracy, F1)."
+              />
+            </ol>
           </section>
 
-          <section className="grid gap-10 border-b border-border/60 pb-12 lg:grid-cols-2">
-            <div className="space-y-5">
-              <SectionLabel label="What the platform supports" />
-              <div className="space-y-4 text-base leading-8 text-muted-foreground">
-                <p>
-                  Move from a model to its reported benchmarks, from an evaluation to the models that ran on it, and from a benchmark family down to its slices and metrics.
-                </p>
-                <p>
-                  Scores are useful only when paired with benchmark scope, evaluator identity, generation setup, and the places where reporting is incomplete.
-                </p>
-              </div>
+          {/* Four interpretive signals */}
+          <section className="grid gap-8 border-b border-border/60 pb-12 lg:grid-cols-[minmax(220px,0.34fr)_minmax(0,1fr)] lg:gap-12">
+            <div className="space-y-3">
+              <SectionLabel label="Interpretive signals" />
+              <h2 className="text-2xl font-semibold tracking-[-0.03em] text-foreground sm:text-[1.75rem]">
+                Four signals that help readers decide whether to trust a reported score.
+              </h2>
+              <p className="max-w-md text-sm leading-7 text-muted-foreground">
+                Each signal is computed at the record level and rolled up at the corpus level. Per-record
+                instances appear on every model and benchmark page; rollups appear on the home page.
+              </p>
             </div>
 
-            <div className="space-y-5">
-              <SectionLabel label="Who it is for" />
-              <div className="grid gap-4 sm:grid-cols-2">
-                <AudienceCard
-                  title="Researchers"
-                  body="Need reproducibility cues, benchmark decomposition, and apples-to-apples comparison support before drawing substantive conclusions."
-                />
-                <AudienceCard
-                  title="Policy readers"
-                  body="Need evaluator independence, benchmark purpose, source accountability, and clearly stated limitations before using reported results."
-                />
-              </div>
-              <p className="text-sm leading-7 text-muted-foreground">
-                Audience modes change emphasis, not the underlying record.
+            <div className="grid gap-3 sm:grid-cols-2">
+              <SignalCard
+                icon={<ShieldCheck className="h-4 w-4" />}
+                label="Reproducibility"
+                question="Can this be reproduced?"
+                body="Flags reported scores whose generation and prompting setup is too underspecified to re-run independently."
+              />
+              <SignalCard
+                icon={<ClipboardCheck className="h-4 w-4" />}
+                label="Reporting completeness"
+                question="Is the documentation complete?"
+                body="Measures the fraction of operationalized framework fields that are populated for a benchmark."
+              />
+              <SignalCard
+                icon={<Users className="h-4 w-4" />}
+                label="Provenance"
+                question="Who reported this, and what risks does the benchmark carry?"
+                body="Distinguishes first-party, third-party, and collaborative reporting and surfaces multi-source coverage."
+              />
+              <SignalCard
+                icon={<GitCompareArrows className="h-4 w-4" />}
+                label="Comparability"
+                question="Are these scores really comparable?"
+                body="Flags score divergence across setup variants and across reporting parties for the same (model, benchmark, metric)."
+              />
+            </div>
+          </section>
+
+          {/* Two reader modes */}
+          <section className="grid gap-8 border-b border-border/60 pb-12 lg:grid-cols-[minmax(220px,0.34fr)_minmax(0,1fr)] lg:gap-12">
+            <div className="space-y-3">
+              <SectionLabel label="Reader modes" />
+              <h2 className="text-2xl font-semibold tracking-[-0.03em] text-foreground sm:text-[1.75rem]">
+                Two renderings of the same record.
+              </h2>
+              <p className="max-w-md text-sm leading-7 text-muted-foreground">
+                The two modes operate on the same underlying data. The difference is which fields are
+                surfaced, which are compressed, and which framing is used.
               </p>
             </div>
+
+            <div className="grid gap-3 sm:grid-cols-2">
+              <ModeCard
+                title="Research mode"
+                body="Foregrounds methodology and configuration. Reproducibility gaps list specific missing fields. Comparability surfaces the underlying setup differences. Default for technical evaluators, benchmark developers, and meta-analysis."
+              />
+              <ModeCard
+                title="Policy mode"
+                body="Foregrounds accountability and plain-language interpretation. The same signals render with narrative caveats and compressed metric configuration. Default for regulators, standards bodies, and non-technical readers."
+              />
+            </div>
           </section>
 
-          <section className="grid gap-10 lg:grid-cols-[minmax(0,1.12fr)_minmax(280px,0.72fr)]">
-            <div className="space-y-5">
-              <SectionLabel label="Current scope" />
-              <div className="space-y-4 text-base leading-8 text-muted-foreground">
-                <p>
-                  Read this version as a public beta. It tries to make structured evaluation reporting legible, while testing which fields, comparison flows, and abstractions hold up in practice.
-                </p>
-                <p>
-                  The larger goal: make AI evaluation evidence easier to inspect and harder to over-interpret.
-                </p>
-              </div>
+          {/* Composition + sources */}
+          <section className="grid gap-8 lg:grid-cols-[minmax(220px,0.34fr)_minmax(0,1fr)] lg:gap-12">
+            <div className="space-y-3">
+              <SectionLabel label="Composition" />
+              <h2 className="text-2xl font-semibold tracking-[-0.03em] text-foreground sm:text-[1.75rem]">
+                Eval Cards does not collect evaluation data directly.
+              </h2>
+              <p className="max-w-md text-sm leading-7 text-muted-foreground">
+                It composes three existing sources and applies a canonicalization layer that maps surface
+                identifiers to nodes in the rollout hierarchy.
+              </p>
             </div>
 
-            <div className="space-y-4 rounded-[1.75rem] border border-border/70 bg-[linear-gradient(180deg,rgba(71,129,177,0.08),rgba(71,129,177,0.02))] p-5">
-              <SectionLabel label="Next steps" />
-              <div className="space-y-3 text-sm leading-7 text-muted-foreground">
-                <p>The homepage is the shortest path into the data.</p>
-                <p>The survey collects missing fields, confusing labels, and misleading comparisons.</p>
-              </div>
-              <div className="flex flex-wrap gap-3 pt-2">
-                <Link href="/">
-                  <Button className="gap-2 rounded-full px-5">
-                    Back to homepage
-                    <ArrowRight className="h-4 w-4" />
-                  </Button>
-                </Link>
-                <Link href="/survey">
-                  <Button variant="outline" className="gap-2 rounded-full px-5">
-                    Leave feedback
-                    <ArrowRight className="h-4 w-4" />
-                  </Button>
-                </Link>
-              </div>
+            <div className="space-y-3">
+              <SourceRow
+                icon={<BookOpenText className="h-4 w-4" />}
+                title="Auto-BenchmarkCards"
+                body="Benchmark metadata: design intent, scoring methodology, data licensing, risk annotations."
+              />
+              <SourceRow
+                icon={<Database className="h-4 w-4" />}
+                title="EEE (evaluation run data)"
+                body="Generation configuration, evaluator relationships, and per-instance results from major frameworks."
+              />
+              <SourceRow
+                icon={<Scale className="h-4 w-4" />}
+                title="Voluntary disclosure"
+                body="Two reserved fields accept developer-supplied disclosure: preregistration links and lifecycle status."
+              />
             </div>
           </section>
+
+          {/* CTAs */}
+          <section className="flex flex-wrap gap-3 border-t border-border/60 pt-10">
+            <Link href="/">
+              <Button className="gap-2 rounded-full px-5">
+                Back to home
+                <ArrowRight className="h-4 w-4" />
+              </Button>
+            </Link>
+            <Link href="/models">
+              <Button variant="outline" className="gap-2 rounded-full px-5">
+                Browse models
+                <ArrowRight className="h-4 w-4" />
+              </Button>
+            </Link>
+            <Link href="/evals">
+              <Button variant="outline" className="gap-2 rounded-full px-5">
+                Browse evaluations
+                <ArrowRight className="h-4 w-4" />
+              </Button>
+            </Link>
+            <Link href="/survey">
+              <Button variant="ghost" className="gap-2 rounded-full px-4 text-foreground">
+                Leave feedback
+                <ArrowRight className="h-4 w-4" />
+              </Button>
+            </Link>
+          </section>
         </div>
       </main>
     </div>
@@ -169,83 +243,60 @@ function SectionLabel({ label }: { label: string }) {
   )
 }
 
-function AboutRow({
-  title,
-  body,
-}: {
-  title: string
-  body: string
-}) {
+function RollupRow({ index, title, body }: { index: string; title: string; body: string }) {
   return (
-    <div className="border-t border-border/50 pt-5 first:border-t-0 first:pt-0">
-      <div className="text-sm font-semibold text-foreground">{title}</div>
-      <p className="mt-2 text-sm leading-7 text-muted-foreground">{body}</p>
-    </div>
+    <li className="flex gap-4 rounded-2xl border border-border/70 bg-card p-4">
+      <div className="flex h-7 w-7 shrink-0 items-center justify-center rounded-full bg-muted/70 text-xs font-semibold text-muted-foreground">
+        {index}
+      </div>
+      <div className="min-w-0">
+        <div className="text-sm font-semibold text-foreground">{title}</div>
+        <p className="mt-1 text-sm leading-6 text-muted-foreground">{body}</p>
+      </div>
+    </li>
   )
 }
 
-function AboutPanel({
+function SignalCard({
   icon,
-  title,
+  label,
+  question,
   body,
-  tone,
 }: {
   icon: React.ReactNode
-  title: string
+  label: string
+  question: string
   body: string
-  tone: "amber" | "emerald" | "sky"
 }) {
-  const toneClasses = {
-    amber: "bg-amber-50/70 dark:bg-amber-950/20",
-    emerald: "bg-emerald-50/70 dark:bg-emerald-950/20",
-    sky: "bg-sky-50/70 dark:bg-sky-950/20",
-  }[tone]
-
   return (
-    <div className={`rounded-[1.5rem] border border-border/70 p-5 ${toneClasses}`}>
-      <div className="flex items-center gap-2 text-sm font-semibold text-foreground">
-        <span className="rounded-full bg-background/80 p-2 text-muted-foreground">{icon}</span>
-        {title}
+    <div className="flex flex-col gap-2 rounded-2xl border border-border/70 bg-card p-4">
+      <div className="flex items-center gap-2">
+        <span className="rounded-full bg-muted/60 p-1.5 text-muted-foreground">{icon}</span>
+        <div className="text-[10px] font-semibold uppercase tracking-[0.22em] text-muted-foreground">{label}</div>
       </div>
-      <p className="mt-3 text-sm leading-7 text-muted-foreground">{body}</p>
+      <div className="text-sm font-semibold text-foreground">{question}</div>
+      <p className="text-sm leading-6 text-muted-foreground">{body}</p>
     </div>
   )
 }
 
-function AudienceCard({
-  title,
-  body,
-}: {
-  title: string
-  body: string
-}) {
+function ModeCard({ title, body }: { title: string; body: string }) {
   return (
-    <div className="rounded-[1.35rem] border border-border/70 bg-background/80 p-4">
+    <div className="rounded-2xl border border-border/70 bg-card p-4">
       <div className="text-sm font-semibold text-foreground">{title}</div>
-      <p className="mt-2 text-sm leading-7 text-muted-foreground">{body}</p>
+      <p className="mt-2 text-sm leading-6 text-muted-foreground">{body}</p>
     </div>
   )
 }
 
-function SignalLine({
-  icon,
-  label,
-  tone,
-}: {
-  icon: React.ReactNode
-  label: string
-  tone: "amber" | "emerald" | "sky"
-}) {
-  const toneClasses = {
-    amber: "bg-amber-50 text-amber-800 dark:bg-amber-950/30 dark:text-amber-300",
-    emerald: "bg-emerald-50 text-emerald-800 dark:bg-emerald-950/30 dark:text-emerald-300",
-    sky: "bg-sky-50 text-sky-800 dark:bg-sky-950/30 dark:text-sky-300",
-  }[tone]
-
+function SourceRow({ icon, title, body }: { icon: React.ReactNode; title: string; body: string }) {
   return (
-    <div className="flex items-start gap-3">
-      <div className={`rounded-full p-2 ${toneClasses}`}>{icon}</div>
-      <div className="text-sm leading-6 text-muted-foreground">{label}</div>
+    <div className="flex items-start gap-3 rounded-2xl border border-border/70 bg-card p-4">
+      <span className="rounded-full bg-muted/60 p-1.5 text-muted-foreground">{icon}</span>
+      <div className="min-w-0">
+        <div className="text-sm font-semibold text-foreground">{title}</div>
+        <p className="mt-1 text-sm leading-6 text-muted-foreground">{body}</p>
+      </div>
     </div>
   )
 }

app/corpus/page.tsx

@@ -1,36 +1,7 @@
-import { CorpusDashboard } from "@/components/signals/corpus-dashboard"
-import { Navigation } from "@/components/navigation"
-import { fetchCorpusAggregates, fetchEvalListLite } from "@/lib/hf-data"
+import { redirect } from "next/navigation"
 
-export default async function CorpusPage() {
-  const [aggregates, evalList] = await Promise.all([
-    fetchCorpusAggregates(),
-    fetchEvalListLite().catch(() => ({ evals: [] })),
-  ])
-
-  const completenessScores = evalList.evals
-    .map((entry) => entry.evalcards?.annotations?.reporting_completeness?.completeness_score)
-    .filter((score): score is number => typeof score === "number" && Number.isFinite(score))
-
-  return (
-    <div className="min-h-screen bg-background">
-      <Navigation />
-      <main className="container mx-auto px-4 py-8">
-        {aggregates ? (
-          <CorpusDashboard aggregates={aggregates} completenessScores={completenessScores} />
-        ) : (
-          <section className="rounded-2xl border border-dashed border-border/70 bg-card p-8 text-center">
-            <div className="text-[11px] font-semibold uppercase tracking-[0.22em] text-muted-foreground">
-              Interpretive signals
-            </div>
-            <h1 className="mt-2 text-2xl font-semibold tracking-tight">Corpus aggregates are not available yet</h1>
-            <p className="mx-auto mt-3 max-w-2xl text-sm leading-6 text-muted-foreground">
-              The frontend is ready for `corpus-aggregates.json`, but this cached backend snapshot does not include it yet.
-              Once the dataset ships the file, this page will render reproducibility, completeness, provenance, and comparability rollups.
-            </p>
-          </section>
-        )}
-      </main>
-    </div>
-  )
+// Corpus signals were folded into the home page. Preserve any external links
+// to /corpus by sending readers to the same content at the root.
+export default function CorpusRedirect() {
+  redirect("/")
 }

app/page.tsx

@@ -1,256 +1,213 @@
 import Link from "next/link"
+import { ArrowRight } from "lucide-react"
+
 import { Button } from "@/components/ui/button"
-import { ArrowRight, BookOpenText, Database, MessageSquare, Scale } from "lucide-react"
+import { Badge } from "@/components/ui/badge"
 import { HomeModeLabel } from "@/components/home-mode-label"
 import { Navigation } from "@/components/navigation"
-import { getBackendManifestData, getEvalListLiteData, getModelCardsLite } from "@/lib/data-backend"
+import { CorpusSignalsStrip } from "@/components/signals/corpus-signals-strip"
+import { getDeveloperList } from "@/lib/data-backend"
+import {
+  fetchBackendManifest,
+  fetchCorpusAggregates,
+  fetchEvalHierarchy,
+} from "@/lib/hf-data"
 
 function formatGeneratedAt(value: string | null | undefined) {
-  if (!value) return "Unknown"
-
+  if (!value) return null
   try {
-    return new Date(value).toLocaleString("en-US", {
-      month: "short",
+    const date = new Date(value)
+    if (Number.isNaN(date.getTime())) return value
+    return date.toLocaleDateString("en-US", {
+      year: "numeric",
+      month: "long",
       day: "numeric",
-      hour: "numeric",
-      minute: "2-digit",
     })
   } catch {
     return value
   }
 }
 
+function formatNumber(value: number | undefined | null): string {
+  if (value == null || !Number.isFinite(value)) return "—"
+  return value.toLocaleString("en-US")
+}
+
 export default async function HomePage() {
-  const [models, evalList, manifest] = await Promise.all([
-    getModelCardsLite(),
-    getEvalListLiteData(),
-    getBackendManifestData(),
+  const [aggregates, manifest, hierarchy, developers] = await Promise.all([
+    fetchCorpusAggregates(),
+    fetchBackendManifest(),
+    fetchEvalHierarchy(),
+    getDeveloperList().catch(() => []),
   ])
 
-  const evalSummaries = evalList.evals
-  const developerCount = new Set(models.map((entry) => entry.developer).filter(Boolean)).size
-  const avgBenchmarksPerModel =
-    models.length > 0
-      ? models.reduce((sum, entry) => sum + entry.benchmarks_count, 0) / models.length
-      : 0
-  const totalReportedResults = models.reduce((sum, entry) => sum + entry.evaluations_count, 0)
+  const stats = hierarchy.stats
+  const familyCount = stats?.family_count ?? hierarchy.families.length
+  const compositeCount = stats?.composite_count ?? 0
+  const singleBenchmarkCount = stats?.single_benchmark_count ?? 0
+  const standaloneBenchmarkCount = stats?.standalone_benchmark_count ?? 0
+  const benchmarkLeafCount = singleBenchmarkCount + standaloneBenchmarkCount
+  const sliceCount = stats?.slice_count ?? 0
+  const metricCount = stats?.metric_count ?? 0
+  const tripleCount = stats?.metric_rows_scanned ?? 0
+  const modelCount = manifest?.model_count ?? 0
+  const developerCount = developers.length
+  const generatedAt = formatGeneratedAt(manifest?.generated_at)
 
   return (
     <div className="min-h-screen bg-background">
       <Navigation />
-      <main className="relative overflow-hidden">
-        <div className="absolute inset-x-0 top-0 -z-10 h-[24rem] bg-[radial-gradient(circle_at_top,rgba(196,167,96,0.12),transparent_60%)]" />
-        <div className="absolute inset-y-0 right-0 -z-10 hidden w-[36rem] bg-[radial-gradient(circle_at_center,rgba(71,129,177,0.08),transparent_66%)] lg:block" />
 
-        <section className="mx-auto flex min-h-[calc(100vh-4.25rem)] w-full max-w-[92rem] flex-col gap-10 px-4 pb-10 pt-12 sm:px-6 sm:pt-16 lg:px-8 lg:pb-12">
-          <div className="flex flex-col gap-8 xl:flex-row xl:items-stretch">
-            <div className="flex flex-1 flex-col gap-6">
-              <div className="flex flex-wrap items-center gap-x-4 gap-y-2 text-[11px] font-semibold uppercase tracking-[0.24em] text-muted-foreground">
-                <span className="rounded-full border border-border/70 bg-background px-3 py-1">Beta preview</span>
-                <HomeModeLabel />
-                {manifest ? <span>Updated {formatGeneratedAt(manifest.generated_at)}</span> : null}
-              </div>
+      <main className="mx-auto w-full max-w-[88rem] px-4 py-12 sm:px-6 lg:px-8 lg:py-16">
+        {/* Masthead */}
+        <section className="border-b border-border/60 pb-10">
+          <div className="flex flex-wrap items-center gap-x-4 gap-y-2 text-[11px] font-semibold uppercase tracking-[0.24em] text-muted-foreground">
+            <HomeModeLabel />
+            {generatedAt && <span>Snapshot · {generatedAt}</span>}
+            {aggregates && (
+              <Badge variant="outline" className="font-normal tracking-normal">
+                Signals v{aggregates.signal_version}
+              </Badge>
+            )}
+          </div>
 
-              <div className="space-y-4">
-                <h1 className="max-w-4xl text-balance text-4xl font-semibold tracking-[-0.05em] text-foreground sm:text-5xl lg:text-[4.25rem] lg:leading-[1.02]">
-                  Public reporting for AI evaluations
-                </h1>
-                <p className="max-w-2xl text-base leading-7 text-muted-foreground">
-                  Explore which models and benchmarks are reported, where coverage is sparse, and what details are missing.
-                </p>
-              </div>
+          <h1 className="mt-5 max-w-4xl text-balance text-4xl font-semibold tracking-[-0.04em] text-foreground sm:text-5xl lg:text-[3.75rem] lg:leading-[1.05]">
+            Eval Cards
+          </h1>
+          <p className="mt-3 max-w-3xl text-base leading-7 text-muted-foreground sm:text-lg sm:leading-8">
+            An interpretative integration layer for AI evaluation reporting. Eval Cards composes
+            benchmark metadata, evaluation run data, and provenance into a single reading surface,
+            and surfaces four interpretive signals — reproducibility, reporting completeness,
+            provenance, and comparability — over a public corpus of reported scores.
+          </p>
+
+          <div className="mt-6 flex flex-wrap gap-3">
+            <Link href="/models">
+              <Button size="lg" className="gap-2 rounded-full px-6">
+                Browse models
+                <ArrowRight className="h-4 w-4" />
+              </Button>
+            </Link>
+            <Link href="/evals">
+              <Button size="lg" variant="outline" className="gap-2 rounded-full px-6">
+                Browse evaluations
+                <ArrowRight className="h-4 w-4" />
+              </Button>
+            </Link>
+            <Link href="/about">
+              <Button size="lg" variant="ghost" className="gap-2 rounded-full px-4 text-foreground">
+                About this project
+                <ArrowRight className="h-4 w-4" />
+              </Button>
+            </Link>
+          </div>
+        </section>
 
-              <div className="flex flex-wrap gap-3 pt-1">
-                <Link href="/models">
-                  <Button size="lg" className="gap-2 rounded-full px-6">
-                    Browse models
-                    <ArrowRight className="h-4 w-4" />
-                  </Button>
-                </Link>
-                <Link href="/evals">
-                  <Button size="lg" variant="outline" className="gap-2 rounded-full px-6">
-                    Browse evaluations
-                    <ArrowRight className="h-4 w-4" />
-                  </Button>
-                </Link>
-                <Link href="/survey">
-                  <Button size="lg" variant="ghost" className="gap-2 rounded-full px-4 text-foreground">
-                    Leave feedback
-                    <ArrowRight className="h-4 w-4" />
-                  </Button>
-                </Link>
+        {/* Corpus stats — paper §5.1 */}
+        <section className="border-b border-border/60 py-10">
+          <div className="grid gap-6 lg:grid-cols-[minmax(220px,0.34fr)_minmax(0,1fr)] lg:gap-12">
+            <div>
+              <div className="text-[11px] font-semibold uppercase tracking-[0.22em] text-muted-foreground">
+                Corpus
               </div>
+              <h2 className="mt-2 text-2xl font-semibold tracking-tight text-foreground sm:text-[1.75rem]">
+                What this snapshot covers
+              </h2>
+              <p className="mt-3 max-w-md text-sm leading-7 text-muted-foreground">
+                Reports are organized through a six-level rollout hierarchy
+                (family → suite → single benchmark → split → subtask → metric) so that
+                aggregate claims can be drilled down to the evidence supporting them.
+              </p>
             </div>
 
-            <aside className="flex min-w-[300px] flex-col rounded-[2rem] border border-border/70 bg-card/80 p-5 shadow-[0_18px_50px_rgba(15,23,42,0.06)] xl:max-w-[22rem]">
-              <div className="grid grid-cols-2 gap-3">
-                <QuietStat label="Models" value={models.length.toString()} detail="Tracked" tone="amber" />
-                <QuietStat label="Evaluations" value={evalSummaries.length.toString()} detail="Benchmarks" tone="sky" />
-                <QuietStat label="Developers" value={developerCount.toString()} detail="Organizations" tone="emerald" />
-                <QuietStat
-                  label="Results"
-                  value={totalReportedResults.toLocaleString()}
-                  detail={`${avgBenchmarksPerModel.toFixed(1)} avg per model`}
-                  tone="slate"
-                />
-              </div>
-            </aside>
+            <div className="grid grid-cols-2 gap-3 sm:grid-cols-3">
+              <CorpusStat
+                label="Models"
+                value={formatNumber(modelCount)}
+                detail="Tracked across reporting sources"
+              />
+              <CorpusStat
+                label="Reported results"
+                value={formatNumber(tripleCount)}
+                detail="(model, benchmark, metric) triples"
+              />
+              <CorpusStat
+                label="Reporting organizations"
+                value={formatNumber(developerCount)}
+                detail="Developers and third-party evaluators"
+              />
+              <CorpusStat
+                label="Benchmark families"
+                value={formatNumber(familyCount)}
+                detail="Top of the rollout hierarchy"
+              />
+              <CorpusStat
+                label="Suites"
+                value={formatNumber(compositeCount)}
+                detail="Composite reporting units"
+              />
+              <CorpusStat
+                label="Single benchmarks"
+                value={formatNumber(benchmarkLeafCount)}
+                detail={`${formatNumber(sliceCount)} slices · ${formatNumber(metricCount)} metrics`}
+              />
+            </div>
           </div>
+        </section>
 
-          <div className="grid grid-cols-1 gap-4 sm:grid-cols-3">
-            <SignalCard
-              icon={<Database className="h-4 w-4" />}
-              title="Reported benchmarks"
-              body="Which benchmarks, settings, and sources are documented."
-              tone="sky"
-            />
-            <SignalCard
-              icon={<Scale className="h-4 w-4" />}
-              title="Comparison context"
-              body="Evaluator relationships and config gaps attached to each record."
-              tone="amber"
-            />
-            <SignalCard
-              icon={<BookOpenText className="h-4 w-4" />}
-              title="Reader modes"
-              body="Research and policy views highlight different fields."
-              tone="emerald"
-            />
+        {/* Four interpretive signals */}
+        <section className="py-10">
+          <div className="mb-6 grid gap-4 lg:grid-cols-[minmax(220px,0.34fr)_minmax(0,1fr)] lg:gap-12">
+            <div>
+              <div className="text-[11px] font-semibold uppercase tracking-[0.22em] text-muted-foreground">
+                Interpretive signals
+              </div>
+              <h2 className="mt-2 text-2xl font-semibold tracking-tight text-foreground sm:text-[1.75rem]">
+                Reproducibility, completeness, provenance, comparability
+              </h2>
+            </div>
+            <p className="max-w-3xl text-sm leading-7 text-muted-foreground">
+              Each signal answers a question a reader brings to a reported score. Per-record
+              instances appear on every model and benchmark page. Corpus-level rollups below
+              show how reporting practice looks across the public record as a whole.
+            </p>
           </div>
 
-          <div className="grid grid-cols-1 gap-4 border-t border-border/60 pt-8 sm:grid-cols-2 lg:grid-cols-4">
-            <RoutePanel
-              href="/models"
-              icon={<Database className="h-4 w-4" />}
-              title="Model records"
-              body="Reported benchmarks per model, plus what's missing."
-            />
-            <RoutePanel
-              href="/evals"
-              icon={<BookOpenText className="h-4 w-4" />}
-              title="Benchmark records"
-              body="How a benchmark is reported across models."
-            />
-            <RoutePanel
-              href="/about"
-              icon={<Scale className="h-4 w-4" />}
-              title="Project notes"
-              body="Why this reporting format exists."
-            />
-            <Link
-              href="/survey"
-              className="group flex h-full flex-col rounded-[1.5rem] border border-border/70 bg-muted/20 p-5 transition-colors hover:bg-muted/30"
-            >
-              <div className="flex items-center gap-2 text-sm font-semibold text-foreground">
-                <span className="rounded-full bg-rose-50 p-2 text-rose-600 dark:bg-rose-950/30">
-                  <MessageSquare className="h-4 w-4" />
-                </span>
-                Feedback
-              </div>
-              <p className="mt-3 text-sm leading-7 text-muted-foreground">
-                Flag missing fields or unclear labels.
+          {aggregates ? (
+            <CorpusSignalsStrip aggregates={aggregates} />
+          ) : (
+            <div className="rounded-2xl border border-dashed border-border/70 bg-muted/10 p-8 text-center">
+              <h3 className="text-lg font-semibold tracking-tight">Corpus aggregates unavailable</h3>
+              <p className="mx-auto mt-2 max-w-2xl text-sm leading-6 text-muted-foreground">
+                The current backend snapshot does not include <code>corpus-aggregates.json</code>.
+                When it does, this section will render the four corpus-level rollups.
               </p>
-              <div className="mt-auto pt-4 inline-flex items-center gap-2 text-sm font-semibold text-foreground">
-                Open survey
-                <ArrowRight className="h-4 w-4" />
-              </div>
-            </Link>
-          </div>
+            </div>
+          )}
         </section>
       </main>
     </div>
   )
 }
 
-function QuietStat({
+function CorpusStat({
   label,
   value,
   detail,
-  tone,
 }: {
   label: string
   value: string
   detail: string
-  tone: "amber" | "emerald" | "rose" | "sky" | "slate"
 }) {
-  const toneClasses = {
-    amber: "border-amber-200/80 bg-amber-50/70 dark:border-amber-900/50 dark:bg-amber-950/20",
-    emerald: "border-emerald-200/80 bg-emerald-50/70 dark:border-emerald-900/50 dark:bg-emerald-950/20",
-    rose: "border-rose-200/80 bg-rose-50/70 dark:border-rose-900/50 dark:bg-rose-950/20",
-    sky: "border-sky-200/80 bg-sky-50/70 dark:border-sky-900/50 dark:bg-sky-950/20",
-    slate: "border-border/70 bg-muted/25",
-  }[tone]
-
   return (
-    <div className={`flex h-full flex-col rounded-[1.35rem] border p-4 ${toneClasses}`}>
-      <div className="text-[10px] font-semibold uppercase tracking-[0.18em] text-muted-foreground">{label}</div>
-      <div className="mt-2 text-2xl font-semibold tracking-tight text-foreground tabular-nums">{value}</div>
-      <p className="text-sm leading-6 text-muted-foreground">{detail}</p>
-    </div>
-  )
-}
-
-function SignalCard({
-  icon,
-  title,
-  body,
-  tone,
-}: {
-  icon: React.ReactNode
-  title: string
-  body: string
-  tone: "amber" | "emerald" | "sky"
-}) {
-  const toneClasses = {
-    amber: "bg-amber-50 text-amber-800 dark:bg-amber-950/30 dark:text-amber-300",
-    emerald: "bg-emerald-50 text-emerald-800 dark:bg-emerald-950/30 dark:text-emerald-300",
-    sky: "bg-sky-50 text-sky-800 dark:bg-sky-950/30 dark:text-sky-300",
-  }[tone]
-
-  const surfaceClasses = {
-    amber: "border-amber-200/70 bg-amber-50/40 dark:border-amber-900/40 dark:bg-amber-950/15",
-    emerald: "border-emerald-200/70 bg-emerald-50/40 dark:border-emerald-900/40 dark:bg-emerald-950/15",
-    sky: "border-sky-200/70 bg-sky-50/40 dark:border-sky-900/40 dark:bg-sky-950/15",
-  }[tone]
-
-  return (
-    <div className={`flex h-full flex-col gap-3 rounded-[1.5rem] border p-5 ${surfaceClasses}`}>
-      <div className="flex items-center gap-2 text-sm font-semibold text-foreground">
-        <span className={`rounded-full p-2 ${toneClasses}`}>{icon}</span>
-        {title}
+    <div className="rounded-2xl border border-border/70 bg-card p-4">
+      <div className="text-[10px] font-semibold uppercase tracking-[0.22em] text-muted-foreground">
+        {label}
       </div>
-      <p className="text-sm leading-6 text-muted-foreground">{body}</p>
-    </div>
-  )
-}
-
-function RoutePanel({
-  href,
-  icon,
-  title,
-  body,
-}: {
-  href: string
-  icon: React.ReactNode
-  title: string
-  body: string
-}) {
-  return (
-    <Link
-      href={href}
-      className="group flex h-full flex-col rounded-[1.5rem] border border-border/70 bg-background/80 p-5 transition-colors hover:bg-muted/20"
-    >
-      <div className="flex items-center gap-2 text-sm font-semibold text-foreground">
-        <span className="rounded-full bg-muted/60 p-2 text-muted-foreground transition-colors group-hover:text-foreground">
-          {icon}
-        </span>
-        {title}
+      <div className="mt-2 text-2xl font-semibold tracking-tight text-foreground tabular-nums sm:text-[1.625rem]">
+        {value}
       </div>
-      <p className="mt-3 text-sm leading-7 text-muted-foreground">{body}</p>
-      <div className="mt-auto pt-4 inline-flex items-center gap-2 text-sm font-semibold text-foreground">
-        Open
-        <ArrowRight className="h-4 w-4" />
-      </div>
-    </Link>
+      <p className="mt-1 text-xs leading-5 text-muted-foreground">{detail}</p>
+    </div>
   )
 }

components/navigation.tsx

@@ -38,12 +38,6 @@ export function Navigation() {
       icon: BarChart3,
       isActive: pathname === "/evals" || pathname?.startsWith("/evals/")
     },
-    {
-      href: "/corpus",
-      label: "Corpus",
-      icon: FlaskConical,
-      isActive: pathname === "/corpus" || pathname?.startsWith("/corpus/")
-    },
     {
       href: "/survey",
       label: "Survey",

components/signals/corpus-dashboard.tsx

@@ -32,9 +32,16 @@ const SOURCE_COLORS: Record<string, string> = {
 export function CorpusDashboard({
   aggregates,
   completenessScores,
+  embedded = false,
 }: {
   aggregates: CorpusAggregates
   completenessScores: number[]
+  /**
+   * When true, hides the dashboard's own masthead card. The host page is
+   * expected to provide its own title and generated-date badge. The view
+   * toggle is preserved inline above the signal sections.
+   */
+  embedded?: boolean
 }) {
   const { mode } = useAudienceMode()
   const [view, setView] = useState<"overall" | "category">("overall")
@@ -56,44 +63,72 @@ export function CorpusDashboard({
 
   return (
     <div className="space-y-6">
-      <section className="rounded-2xl border border-border/70 bg-card p-5 shadow-sm">
-        <div className="flex flex-col gap-4 lg:flex-row lg:items-start lg:justify-between">
-          <div>
-            <div className="text-[11px] font-semibold uppercase tracking-[0.22em] text-muted-foreground">
-              Interpretive signals
-            </div>
-            <h1 className="mt-2 text-3xl font-semibold tracking-tight">Corpus Dashboard</h1>
-            <p className="mt-2 max-w-3xl text-sm leading-6 text-muted-foreground">
-              Corpus-level rollups for reproducibility, documentation completeness, source provenance, and comparability.
-            </p>
+      {embedded ? (
+        <div className="flex flex-wrap items-center justify-between gap-3">
+          <div className="text-[11px] font-semibold uppercase tracking-[0.22em] text-muted-foreground">
+            Interpretive signals
+          </div>
+          <div className="inline-flex rounded-full border bg-muted/20 p-1">
+            <Button
+              type="button"
+              size="sm"
+              variant={view === "overall" ? "default" : "ghost"}
+              className="h-8 rounded-full"
+              onClick={() => setView("overall")}
+            >
+              Overall
+            </Button>
+            <Button
+              type="button"
+              size="sm"
+              variant={view === "category" ? "default" : "ghost"}
+              className="h-8 rounded-full"
+              onClick={() => setView("category")}
+            >
+              By category
+            </Button>
           </div>
+        </div>
+      ) : (
+        <section className="rounded-2xl border border-border/70 bg-card p-5 shadow-sm">
+          <div className="flex flex-col gap-4 lg:flex-row lg:items-start lg:justify-between">
+            <div>
+              <div className="text-[11px] font-semibold uppercase tracking-[0.22em] text-muted-foreground">
+                Interpretive signals
+              </div>
+              <h1 className="mt-2 text-3xl font-semibold tracking-tight">Corpus Dashboard</h1>
+              <p className="mt-2 max-w-3xl text-sm leading-6 text-muted-foreground">
+                Corpus-level rollups for reproducibility, reporting completeness, provenance, and comparability.
+              </p>
+            </div>
 
-          <div className="flex flex-wrap items-center gap-2">
-            <Badge variant="outline">Signals v{aggregates.signal_version}</Badge>
-            <Badge variant="outline">Generated {formatGeneratedDate(aggregates.generated_at)}</Badge>
-            <div className="inline-flex rounded-full border bg-muted/20 p-1">
-              <Button
-                type="button"
-                size="sm"
-                variant={view === "overall" ? "default" : "ghost"}
-                className="h-8 rounded-full"
-                onClick={() => setView("overall")}
-              >
-                Overall
-              </Button>
-              <Button
-                type="button"
-                size="sm"
-                variant={view === "category" ? "default" : "ghost"}
-                className="h-8 rounded-full"
-                onClick={() => setView("category")}
-              >
-                By category
-              </Button>
+            <div className="flex flex-wrap items-center gap-2">
+              <Badge variant="outline">Signals v{aggregates.signal_version}</Badge>
+              <Badge variant="outline">Generated {formatGeneratedDate(aggregates.generated_at)}</Badge>
+              <div className="inline-flex rounded-full border bg-muted/20 p-1">
+                <Button
+                  type="button"
+                  size="sm"
+                  variant={view === "overall" ? "default" : "ghost"}
+                  className="h-8 rounded-full"
+                  onClick={() => setView("overall")}
+                >
+                  Overall
+                </Button>
+                <Button
+                  type="button"
+                  size="sm"
+                  variant={view === "category" ? "default" : "ghost"}
+                  className="h-8 rounded-full"
+                  onClick={() => setView("category")}
+                >
+                  By category
+                </Button>
+              </div>
             </div>
           </div>
-        </div>
-      </section>
+        </section>
+      )}
 
       {view === "overall" ? (
         <div className="grid gap-6">

components/signals/corpus-signals-strip.tsx

@@ -0,0 +1,146 @@
+import type { CorpusAggregates } from "@/lib/backend-artifacts"
+
+/**
+ * Compact 4-tile rollup of the four interpretive signals (paper §4.2.1).
+ * Designed for the home page: one headline per signal, one caption, one
+ * sub-detail line. Drill-down lives on per-record pages.
+ */
+export function CorpusSignalsStrip({
+  aggregates,
+}: {
+  aggregates: CorpusAggregates
+}) {
+  const repro = aggregates.reproducibility.overall
+  const comp = aggregates.completeness.overall
+  const prov = aggregates.provenance.overall
+  const cmp = aggregates.comparability.overall
+
+  // Invert the gap rate so the headline reads as "how many are documented",
+  // not "how many have gaps". The inverse is more impactful and lower-is-worse,
+  // which matches the user's intuition. The gap rate stays in the detail line.
+  const reproDocumentedRate =
+    repro.reproducibility_gap_rate == null
+      ? null
+      : Math.max(0, 1 - repro.reproducibility_gap_rate)
+  const reproDocumentedRateText = formatPct(reproDocumentedRate)
+  const reproGapRateText = formatPct(repro.reproducibility_gap_rate)
+  const reproDetail = topMissingFields(repro.per_field_missingness, 2)
+
+  const compMean = formatPct(comp.completeness_score_mean)
+  const compMedian = formatPct(comp.completeness_score_median)
+
+  const provMulti = formatPct(prov.multi_source_rate)
+  const totalReports = prov.total_triples
+  const tpShare = totalReports > 0 ? prov.source_type_distribution.third_party / totalReports : 0
+  const fpShare = totalReports > 0 ? prov.source_type_distribution.first_party / totalReports : 0
+
+  const cmpRate = formatPct(cmp.variant_divergence_rate)
+  const crossPartyAvailable = cmp.cross_party_eligible_groups > 0
+
+  return (
+    <div className="grid gap-3 sm:grid-cols-2 xl:grid-cols-4">
+      <SignalTile
+        label="Reproducibility"
+        value={reproDocumentedRateText}
+        caption="of reported scores have a complete setup recorded — the rest cannot be independently re-run"
+        detail={
+          reproDetail
+            ? `${reproGapRateText} have at least one undocumented field. Most often missing: ${reproDetail}`
+            : `${reproGapRateText} have at least one undocumented field`
+        }
+      />
+
+      <SignalTile
+        label="Reporting completeness"
+        value={compMean}
+        caption={`mean across ${comp.total_benchmarks.toLocaleString()} benchmarks (median ${compMedian})`}
+        detail="Source-provenance fields populate fully; preregistration fields are unmet"
+      />
+
+      <SignalTile
+        label="Provenance"
+        value={provMulti}
+        caption="of (model, benchmark) groups have reports from more than one party"
+        detail={`${formatPct(tpShare)} third-party, ${formatPct(fpShare)} first-party of ${totalReports.toLocaleString()} results`}
+      />
+
+      <SignalTile
+        label="Comparability"
+        value={cmpRate}
+        caption={`of setup-eligible groups diverge across variants (${cmp.variant_divergent_groups.toLocaleString()} of ${cmp.variant_eligible_groups.toLocaleString()})`}
+        detail={
+          crossPartyAvailable
+            ? `Cross-party divergence: ${formatPct(cmp.cross_party_divergence_rate)}`
+            : "Cross-party divergence not yet computable — too few multi-org reports"
+        }
+      />
+    </div>
+  )
+}
+
+function SignalTile({
+  label,
+  value,
+  caption,
+  detail,
+}: {
+  label: string
+  value: string
+  caption: string
+  detail: string
+}) {
+  return (
+    <div className="flex flex-col gap-2 rounded-2xl border border-border/70 bg-card p-4">
+      <div className="text-[10px] font-semibold uppercase tracking-[0.22em] text-muted-foreground">
+        {label}
+      </div>
+      <div className="flex items-baseline gap-2">
+        <span className="text-3xl font-semibold tabular-nums tracking-tight text-foreground">
+          {value}
+        </span>
+      </div>
+      <p className="text-sm leading-snug text-muted-foreground">{caption}.</p>
+      <p className="mt-auto text-xs leading-snug text-muted-foreground/80">{detail}.</p>
+    </div>
+  )
+}
+
+function formatPct(value: number | null | undefined): string {
+  if (value == null || !Number.isFinite(value)) return "—"
+  if (value === 0) return "0%"
+  if (value > 0 && value < 0.01) return "<1%"
+  return `${Math.round(value * 100)}%`
+}
+
+const FIELD_LABELS: Record<string, string> = {
+  temperature: "temperature",
+  max_tokens: "max tokens",
+  top_p: "top-p",
+  prompt_template: "prompt template",
+  eval_plan: "eval plan",
+  eval_limits: "eval limits",
+}
+
+/**
+ * Returns a short comma-separated list of the top-N missing fields, ranked by
+ * missing count. We rank by count rather than rate so cross-denominator fields
+ * (agentic-only vs all-triples) don't get artificially boosted.
+ */
+function topMissingFields(
+  perField: Record<
+    string,
+    { missing_count: number; missing_rate: number | null; denominator: string }
+  >,
+  n: number,
+): string {
+  const entries = Object.entries(perField)
+    .filter(([, v]) => v.missing_count > 0)
+    .sort((a, b) => b[1].missing_count - a[1].missing_count)
+    .slice(0, n)
+
+  if (entries.length === 0) return ""
+
+  return entries
+    .map(([key, v]) => `${FIELD_LABELS[key] ?? key.replace(/_/g, " ")} (${formatPct(v.missing_rate)})`)
+    .join(", ")
+}

lib/backend-artifacts.ts

@@ -2,6 +2,14 @@ export interface BackendManifest {
   generated_at: string
   config_version: number
   skipped_configs: string[]
+  // Counts the upstream pipeline already records in manifest.json. The home
+  // page reads these directly so it does not have to load model-cards-lite
+  // (~20MB) just to display headline numbers.
+  model_count?: number
+  eval_count?: number
+  metric_eval_count?: number
+  source_config_count?: number
+  skipped_config_count?: number
   summary_artifacts?: {
     corpus_aggregates?: string
     [key: string]: string | undefined