citable: values in declared units + GEO misc fixes

src/app/answers/[slug]/page.tsx

@@ -8,7 +8,7 @@ import { cleanLeftoverTokens } from "@/lib/answers-template";
 import { Breadcrumb } from "@/components/breadcrumb";
 import { Pill } from "@/components/pill";
 import { ProviderLogo } from "@/components/provider-logo";
-import { fmtValue, fmtUnit, unitSuffix } from "@/lib/format";
+import { fmtAsOfUtc, fmtValue, fmtUnit, unitSuffix } from "@/lib/format";
 import { isRegion } from "@/lib/brand";
 import { SITE } from "@/data/site";
 import {
@@ -92,6 +92,7 @@ export default async function AnswerPage({
   // YAML's slug reference) fall through to a neutral fallback via
   // cleanLeftoverTokens so a placeholder string never reaches the SERP.
   const render = (s: string) => cleanLeftoverTokens(renderTemplate(s, bench));
+  const asOfUtc = fmtAsOfUtc(bench.lastRunAt);
   const shortAnswer = render(ans.short_answer);
   const intro = render(ans.intro);
   const methodology = render(ans.methodology);
@@ -187,6 +188,17 @@ export default async function AnswerPage({
 
       <div className="mt-6 max-w-3xl border-y border-rule py-5 text-base sm:text-lg leading-relaxed text-ink">
         {shortAnswer}
+        {/* Dated freshness stamp right next to the quotable sentence,
+            mirroring the "Data as of ... UTC" line on the bench pages.
+            Answer engines quote a claim far more readily when the date
+            of measurement sits beside it. Uses the referenced bench's
+            lastRunAt (real data timestamp), not build time. */}
+        {asOfUtc && (
+          <p className="mt-3 text-[11px] text-ink-faint">
+            Data as of <time dateTime={bench.lastRunAt}>{asOfUtc}</time>,
+            refreshed continuously.
+          </p>
+        )}
       </div>
 
       <p className="mt-6 max-w-3xl text-base leading-relaxed text-ink-soft">

src/app/api/citable/route.ts

@@ -3,6 +3,7 @@ import { getBenchmarks } from "@/data/benchmarks";
 import { SITE } from "@/data/site";
 import { AllBenchmarksDraftError } from "@/lib/spec";
 import { citeBundle, fieldValue, leader, headlineSentence } from "@/lib/citation";
+import { valueInDeclaredUnit } from "@/lib/format";
 import { clientKey, rateLimit, tooManyRequests } from "@/lib/rate-limit";
 
 export const runtime = "nodejs";
@@ -53,19 +54,28 @@ export async function GET(req: Request) {
     // from an undersized field. The headline sentence is rewritten to
     // "insufficient data" by headlineSentence above.
     const insufficient = b.dataConfidence === "insufficient";
+    // `value` and `leader.value` are published in the declared `unit`.
+    // Latency benches with unit "s" store ms internally (fmtUnit
+    // convention); valueInDeclaredUnit converts so the JSON never claims
+    // 645 seconds for a 645 ms head lag.
+    const raw = insufficient ? null : fieldValue(b);
     return {
       slug: b.slug,
       title: b.title,
       category: b.category,
       metric: b.metric,
       unit: b.unit,
       status: b.status,
-      value: insufficient ? null : fieldValue(b),
+      value: raw == null ? null : valueInDeclaredUnit(raw, b.unit),
       leader:
         insufficient
           ? null
           : top
-            ? { name: top.name, slug: top.slug, value: top.value }
+            ? {
+                name: top.name,
+                slug: top.slug,
+                value: valueInDeclaredUnit(top.value, b.unit),
+              }
             : null,
       sampleSize: b.sampleSize,
       expectedN: b.expectedN,

src/app/api/openapi.json/route.ts

@@ -14,8 +14,8 @@ export async function GET() {
     openapi: "3.1.0",
     info: {
       title: "OpenChainBench API",
-      version: "1.0.0",
-      description: SITE.description,
+      version: "1.1.0",
+      description: `${SITE.description} An MCP server (Streamable HTTP, POST only) is also available at ${SITE.url}/api/mcp/mcp exposing list_benchmarks, get_benchmark and query_prom tools; see ${SITE.url}/mcp for install instructions.`,
       license: { name: "CC-BY-4.0", url: "https://creativecommons.org/licenses/by/4.0/" },
     },
     servers: [{ url: SITE.url }],
@@ -67,6 +67,50 @@ export async function GET() {
           },
         },
       },
+      "/api/llm-context": {
+        get: {
+          summary:
+            "All benchmarks, rankings and methodology as one Markdown document, ready to paste into a system prompt.",
+          operationId: "get_llm_context",
+          responses: {
+            "200": {
+              description: "OK",
+              content: { "text/markdown": { schema: { type: "string" } } },
+            },
+            "503": { description: "Benchmarks temporarily unavailable" },
+          },
+        },
+      },
+      "/api/freshness": {
+        get: {
+          summary:
+            "Lightweight freshness probe: last resolved data timestamp (epoch ms) per benchmark slug.",
+          operationId: "get_freshness",
+          responses: {
+            "200": {
+              description: "OK",
+              content: {
+                "application/json": {
+                  schema: { $ref: "#/components/schemas/Freshness" },
+                },
+              },
+            },
+          },
+        },
+      },
+      "/api/search/featured": {
+        get: {
+          summary:
+            "Slim featured-leaders payload (live leaders + trending benches) that feeds the site search dialog.",
+          operationId: "get_featured_leaders",
+          responses: {
+            "200": {
+              description: "OK",
+              content: { "application/json": { schema: { type: "object" } } },
+            },
+          },
+        },
+      },
     },
     components: {
       schemas: {
@@ -85,20 +129,39 @@ export async function GET() {
             title: { type: "string" },
             metric: { type: "string" },
             unit: { type: "string" },
-            value: { type: "number", nullable: true },
+            value: {
+              type: "number",
+              nullable: true,
+              description: "Current leading value, expressed in `unit`.",
+            },
             headline: { type: "string" },
             url: { type: "string", format: "uri" },
             api: { type: "string", format: "uri" },
             ogImage: { type: "string", format: "uri" },
             asOf: { type: "string", format: "date-time" },
           },
         },
+        Freshness: {
+          type: "object",
+          properties: {
+            now: { type: "integer", description: "Server epoch ms at response time." },
+            freshness: {
+              type: "object",
+              additionalProperties: { type: "integer" },
+              description: "Benchmark slug to last data timestamp (epoch ms).",
+            },
+          },
+        },
         Stat: {
           type: "object",
           properties: {
             slug: { type: "string" },
             title: { type: "string" },
-            value: { type: "number", nullable: true },
+            value: {
+              type: "number",
+              nullable: true,
+              description: "Current leading value, expressed in `unit`.",
+            },
             unit: { type: "string" },
             rankings: { type: "array" },
             sparkline: { type: "array", items: { type: "number" } },

src/app/api/stat/[slug]/route.ts

@@ -9,6 +9,7 @@ import {
   leader,
   sparklineFor,
 } from "@/lib/citation";
+import { valueInDeclaredUnit } from "@/lib/format";
 import { clientKey, rateLimit, tooManyRequests } from "@/lib/rate-limit";
 import { SLUG_RE } from "@/lib/slug";
 
@@ -45,6 +46,9 @@ export async function GET(
 
   const top = leader(b);
   const insufficient = b.dataConfidence === "insufficient";
+  // Publish value + leader.value in the declared unit. Unit "s" benches
+  // store ms internally (fmtUnit convention); do not leak that here.
+  const raw = insufficient ? null : fieldValue(b);
   const payload = {
     slug: b.slug,
     title: b.title,
@@ -59,8 +63,11 @@ export async function GET(
     // leader; the headline is rewritten by headlineSentence so the
     // agent / journalist reads "insufficient data" instead of quoting
     // a number drawn from undersized samples.
-    value: insufficient ? null : fieldValue(b),
-    leader: insufficient ? null : top,
+    value: raw == null ? null : valueInDeclaredUnit(raw, b.unit),
+    leader:
+      insufficient || !top
+        ? null
+        : { ...top, value: valueInDeclaredUnit(top.value, b.unit) },
     rankings: b.results
       .filter((r) => r.ms.p50 > 0)
       // Drop "insufficient" rows from the machine-readable ranking too:

src/app/benchmarks/[slug]/page.tsx

@@ -26,6 +26,7 @@ import {
   isInsufficient,
   leader,
 } from "@/lib/citation";
+import { valueInDeclaredUnit } from "@/lib/format";
 import { capDescription } from "@/lib/seo-text";
 import { getBenchCreatedAt } from "@/lib/seo/bench-dates";
 import { SITE } from "@/data/site";
@@ -332,7 +333,9 @@ export default async function BenchmarkPage({
         metric: benchmark.metric,
         metricUnit: benchmark.unit,
         leaderName: currentLeader.name,
-        leaderValue: currentLeader.value,
+        // Observation.measuredValue must be in unitText. Unit "s" benches
+        // store ms internally; convert before publishing to JSON-LD.
+        leaderValue: valueInDeclaredUnit(currentLeader.value, benchmark.unit),
         temporalCoverage: `${getBenchCreatedAt(benchmark.slug).toISOString()}/${benchmark.lastRunAt}`,
         observationDate: benchmark.lastRunAt,
         url: benchmarkUrl,

src/app/llms.txt/route.ts

@@ -46,7 +46,7 @@ export async function GET() {
   lines.push(`- [Citable index (JSON)](${SITE.url}/api/citable): flat list of all benchmarks with current values, ready for one-shot lookup.`);
   lines.push(`- [LLM context (Markdown)](${SITE.url}/api/llm-context): all ${benches.length} benchmarks + rankings + methodology in one Markdown blob, ready to paste into a system prompt.`);
   lines.push(`- [OpenAPI schema](${SITE.url}/api/openapi.json): full description of every endpoint.`);
-  lines.push(`- [MCP server](${SITE.url}/api/mcp/mcp): exposes \`list_benchmarks\`, \`get_benchmark\`, \`query_prom\` tools + \`openchainbench://benchmark/{slug}\` resources over Streamable HTTP. See [/mcp](${SITE.url}/mcp) for install instructions (Claude Desktop, Cursor, generic clients).`);
+  lines.push(`- [MCP server docs](${SITE.url}/mcp): install instructions (Claude Desktop, Cursor, generic clients) for the MCP server at ${SITE.url}/api/mcp/mcp, which exposes \`list_benchmarks\`, \`get_benchmark\`, \`query_prom\` tools + \`openchainbench://benchmark/{slug}\` resources over Streamable HTTP (JSON-RPC via POST; the endpoint is not browsable with GET).`);
   lines.push("");
   lines.push(`## Benchmarks`);
   lines.push("");

src/app/page.tsx

@@ -5,7 +5,7 @@ import { getBenchmarksSafe } from "@/data/benchmarks";
 import { HeroRadar } from "@/components/hero-radar";
 import { HomeBenchTable } from "@/components/home-bench-table";
 import { LiveDashboard } from "@/components/live/dashboard";
-import { GLOBAL_DATASET_JSONLD } from "@/lib/dataset-jsonld";
+import { buildGlobalDatasetJsonLd } from "@/lib/dataset-jsonld";
 import { safeJsonLd } from "@/lib/jsonld";
 
 export const revalidate = 60;
@@ -35,6 +35,15 @@ export const metadata: Metadata = {
 export default async function HomePage() {
   const benchmarks = await getBenchmarksSafe();
 
+  // Same timestamp source the bench pages use for Dataset.dateModified:
+  // the harness lastRunAt (real data timestamp, not build time). The
+  // newest one across the catalog dates the site-wide Dataset node.
+  const latestRunAt = benchmarks
+    .map((b) => b.lastRunAt)
+    .filter((iso) => iso && !Number.isNaN(new Date(iso).getTime()))
+    .sort()
+    .pop();
+
   return (
     <article className="mx-auto max-w-[1400px] px-4 sm:px-6 py-10 sm:py-14 space-y-14 sm:space-y-20">
       {/* Site-wide schema.org/Dataset entry. Points Google Dataset Search,
@@ -46,7 +55,9 @@ export default async function HomePage() {
       <script
         type="application/ld+json"
         // biome-ignore lint/security/noDangerouslySetInnerHtml: serialized via safeJsonLd
-        dangerouslySetInnerHTML={{ __html: safeJsonLd(GLOBAL_DATASET_JSONLD) }}
+        dangerouslySetInnerHTML={{
+          __html: safeJsonLd(buildGlobalDatasetJsonLd(latestRunAt)),
+        }}
       />
 
       {/* Hero */}

src/components/benchmark-body.tsx

@@ -24,7 +24,7 @@
 import { CountLeaderboard } from "@/components/count-leaderboard";
 import { SummaryStat } from "@/components/summary-stat";
 import { ViewSwitcher } from "@/components/view-switcher";
-import { fmtUnit } from "@/lib/format";
+import { fmtAsOfUtc, fmtUnit } from "@/lib/format";
 import { computeFieldStats } from "@/lib/stats";
 import { defaultViewFor, viewsForBenchmark } from "@/lib/views";
 import { useViewPreference } from "@/hooks/use-view-preference";
@@ -62,18 +62,6 @@
   );
 }
 
-/** "2026-07-06 14:00 UTC" from the bench's lastRunAt ISO string. Fixed
- *  UTC rendering so server and client HTML agree (no hydration drift)
- *  and answer engines get an unambiguous freshness stamp. */
-function fmtAsOfUtc(iso: string): string | null {
-  const d = new Date(iso);
-  if (Number.isNaN(d.getTime())) return null;
-  const pad = (n: number) => String(n).padStart(2, "0");
-  return `${d.getUTCFullYear()}-${pad(d.getUTCMonth() + 1)}-${pad(
-    d.getUTCDate(),
-  )} ${pad(d.getUTCHours())}:${pad(d.getUTCMinutes())} UTC`;
-}
-
 /** Mutate `url.searchParams` to keep one dimension param in sync.
  *  Removes the param when the value is the first option (the implicit
  *  default) so canonical URLs stay short. */
@@ -540,7 +528,7 @@
   // not only on the timeseries view. Providers the panel has no value
   // for (book could not fill the tier) drop out of the ranking, which
   // is the skipped-not-extrapolated rule made visible.
   const panelViewBenchmark = useMemo(() => {
     if (!activePanel) return viewBenchmark;
     const vals = activePanel.values ?? {};
     return {

src/lib/dataset-jsonld.ts

@@ -29,11 +29,13 @@ export const HF_DATASET_URL =
 export const ZENODO_CONCEPT_DOI = "10.5281/zenodo.20800311";
 export const ZENODO_CONCEPT_URL = `https://doi.org/${ZENODO_CONCEPT_DOI}`;
 
-/** Latest parquet headlines snapshot. `LATEST` is the convention used by
- *  the HF publisher pipeline so downstream consumers can pin to a date or
- *  follow head. */
+/** Latest parquet headlines snapshot via Hugging Face's auto-converted
+ *  parquet API. The repo itself only holds dated partitions
+ *  (headlines/snapshot_date=YYYY-MM-DD/); the previously linked
+ *  snapshot_date=LATEST path never existed and 404'd. This endpoint
+ *  always tracks the head of main, so it stays valid without rebuilds. */
 export const HF_HEADLINES_LATEST =
-  "https://huggingface.co/datasets/OpenChainBench/benchmarks/resolve/main/headlines/snapshot_date=LATEST/part-0.parquet";
+  "https://huggingface.co/api/datasets/OpenChainBench/benchmarks/parquet/headlines/train/0.parquet";
 
 /** CC-BY-4.0 license URL. Matches the per-row license surfaced in
  *  /api/citable and the footer on every page. */
@@ -68,7 +70,20 @@ const CREATOR_PUBLISHER = {
  * Site-wide Dataset entry. Emitted on the home page so Google Dataset
  * Search and Perplexity have a single canonical record pointing at both
  * the live JSON index (/api/citable) and the parquet mirror on HF.
+ *
+ * `buildGlobalDatasetJsonLd` folds in `dateModified` from the newest
+ * bench lastRunAt so the home Dataset carries the same freshness signal
+ * the per-bench Dataset nodes already emit.
  */
+export function buildGlobalDatasetJsonLd(
+  dateModified?: string,
+): Record<string, unknown> {
+  return {
+    ...GLOBAL_DATASET_JSONLD,
+    ...(dateModified ? { dateModified } : {}),
+  };
+}
+
 export const GLOBAL_DATASET_JSONLD = {
   "@context": "https://schema.org",
   "@type": "Dataset",

src/lib/format.ts

@@ -1,3 +1,27 @@
+/** Convert a stored value into the bench's declared unit for machine
+ *  readable payloads. Latency benches with unit "s" store milliseconds
+ *  by convention (see fmtUnit below, which divides by 1000 before
+ *  rendering). That convention must not leak into public JSON where
+ *  agents read `value` + `unit` literally: aggregator-head-lag shipped
+ *  {"value":645,"unit":"s"} while the headline said "0.6 s". Every
+ *  other unit stores exactly what it declares. */
+export function valueInDeclaredUnit(value: number, unit: string): number {
+  return unit === "s" ? value / 1000 : value;
+}
+
+/** "2026-07-06 14:00 UTC" from an ISO timestamp. Fixed UTC rendering so
+ *  server and client HTML agree (no hydration drift) and answer engines
+ *  get an unambiguous freshness stamp. Shared by benchmark-body and the
+ *  answers pages so the dated line reads identically everywhere. */
+export function fmtAsOfUtc(iso: string): string | null {
+  const d = new Date(iso);
+  if (Number.isNaN(d.getTime())) return null;
+  const pad = (n: number) => String(n).padStart(2, "0");
+  return `${d.getUTCFullYear()}-${pad(d.getUTCMonth() + 1)}-${pad(
+    d.getUTCDate(),
+  )} ${pad(d.getUTCHours())}:${pad(d.getUTCMinutes())} UTC`;
+}
+
 export function fmtUnit(value: number, unit: string) {
   if (!Number.isFinite(value)) return "-";
   if (unit === "pct") return formatPercent(value);