Build results-first D3 benchmark explorer

tests/test_publication_claims.py

@@ -49,13 +49,15 @@ def test_public_claim_surfaces_use_the_frozen_lane_and_current_gate() -> None:
 
 
 def test_employer_facing_site_links_current_evidence_and_metadata() -> None:
-    hero = Path("web/src/components/Hero.tsx").read_text()
+    results = Path("web/src/components/ResultsExplorer.tsx").read_text()
     footer = Path("web/src/components/Footer.tsx").read_text()
     index = Path("web/index.html").read_text()
 
-    assert "0 beat the scripted bar" in hero
-    assert "sota-v2-phase-one-2026-07-19" in hero
-    assert "Built by Ned Cutler" in footer
+    assert "Above scripted bar" in results
+    assert "modelsAboveBar" in results
+    assert "Holm-adjusted family test" in results
+    assert "sota-v2-phase-one-2026-07-19" in footer
+    assert "Ned Cutler" in footer
     assert "REPRODUCING_SOTA_V2_RELEASE.md" in footer
     assert 'href="%BASE_URL%favicon.svg"' in index
     assert 'rel="canonical" href="https://nedcut.github.io/gm-bench/"' in index

web/index.html

@@ -4,31 +4,31 @@
     <meta charset="UTF-8" />
     <link rel="icon" type="image/svg+xml" href="%BASE_URL%favicon.svg" />
     <meta name="viewport" content="width=device-width, initial-scale=1.0" />
-    <meta name="theme-color" content="#F4F7F9" />
+    <meta name="theme-color" content="#EEF2F5" />
     <meta
       name="description"
-      content="GM-Bench is a deterministic, pre-registered test of long-horizon LLM planning. Eight models were evaluated under one frozen lane; none beat the strongest scripted heuristic."
+      content="GM-Bench: a deterministic front-office benchmark for LLM agents. Phase one — 0 models beat the strongest scripted heuristic under a frozen compute lane."
     />
     <!-- Canonical metadata is not a fetched subresource, so SRI does not apply. -->
     <link rel="canonical" href="https://nedcut.github.io/gm-bench/" />
     <meta property="og:type" content="website" />
     <meta property="og:url" content="https://nedcut.github.io/gm-bench/" />
-    <meta property="og:title" content="GM-Bench — Can a language model run a front office?" />
+    <meta property="og:title" content="GM-Bench — phase one standings" />
     <meta
       property="og:description"
-      content="Eight models were evaluated in a deterministic, pre-registered front-office benchmark. None beat the strongest scripted heuristic."
+      content="Phase one: 0 models beat the scripted bar under one frozen lane. Scores, tiers, and the red line."
     />
     <meta name="twitter:card" content="summary" />
-    <meta name="twitter:title" content="GM-Bench — Can a language model run a front office?" />
+    <meta name="twitter:title" content="GM-Bench — phase one standings" />
     <meta
       name="twitter:description"
-      content="Eight models were evaluated under one frozen lane. None beat the strongest scripted heuristic."
+      content="Phase one: 0 models beat the scripted bar under one frozen lane."
     />
-    <title>GM-Bench — Can a language model run a front office?</title>
+    <title>GM-Bench — phase one standings</title>
     <link rel="preconnect" href="https://fonts.googleapis.com" />
     <link rel="preconnect" href="https://fonts.gstatic.com" crossorigin />
     <link
-      href="https://fonts.googleapis.com/css2?family=Barlow+Condensed:wght@500;600;700&family=Source+Serif+4:ital,opsz,wght@0,8..60,400;0,8..60,600;1,8..60,400&family=IBM+Plex+Mono:wght@400;500;600&display=swap"
+      href="https://fonts.googleapis.com/css2?family=Barlow+Condensed:wght@600;700&family=IBM+Plex+Mono:wght@400;500;600&family=IBM+Plex+Sans:ital,wght@0,400;0,500;0,600;1,400&family=Newsreader:opsz,wght@6..72,500;6..72,600&display=swap"
       rel="stylesheet"
     />
   </head>

web/package.json

@@ -5,15 +5,20 @@
   "type": "module",
   "scripts": {
     "dev": "vite",
-    "build": "tsc -b && vite build",
+    "validate:data": "bun scripts/validate_results_data.ts",
+    "build": "bun run validate:data && tsc -b && vite build",
     "lint": "oxlint",
     "preview": "vite preview"
   },
   "dependencies": {
+    "d3-array": "^3.2.4",
+    "d3-scale": "^4.0.2",
     "react": "^19.2.7",
     "react-dom": "^19.2.7"
   },
   "devDependencies": {
+    "@types/d3-array": "^3.2.2",
+    "@types/d3-scale": "^4.0.9",
     "@types/node": "^26.1.1",
     "@types/react": "^19.2.17",
     "@types/react-dom": "^19.2.3",

web/scripts/validate_results_data.ts

@@ -0,0 +1,18 @@
+import leaderboardData from "../src/data/leaderboard.json";
+import { buildBenchmarkView } from "../src/benchmarkData";
+import type { Leaderboard } from "../src/types";
+
+const leaderboard = leaderboardData as Leaderboard;
+const benchmark = buildBenchmarkView(leaderboard);
+
+if (benchmark.modelCount !== leaderboard.publication.eligible_headline_models) {
+  throw new Error(
+    `Results UI has ${benchmark.modelCount} rows, but publication metadata declares ` +
+      `${leaderboard.publication.eligible_headline_models} eligible headline models`,
+  );
+}
+
+console.log(
+  `Validated ${benchmark.modelCount} result rows from one leaderboard source ` +
+    `(${benchmark.repeats} repeats; ${benchmark.modelsAboveBar} above scripted bar).`,
+);

web/src/App.tsx

@@ -1,30 +1,40 @@
+import { useState } from "react";
 import snapshotData from "./data/snapshot.json";
 import leaderboardData from "./data/leaderboard.json";
 import type { Leaderboard as LeaderboardData, Snapshot } from "./types";
+import { buildBenchmarkView } from "./benchmarkData";
 import Nav from "./components/Nav";
-import Hero from "./components/Hero";
-import Leaderboard from "./components/Leaderboard";
-import Integrity from "./components/Integrity";
-import Results from "./components/Results";
+import ResultsExplorer from "./components/ResultsExplorer";
+import Analysis from "./components/Analysis";
 import HowItWorks from "./components/HowItWorks";
-import Adapters from "./components/Adapters";
 import Quickstart from "./components/Quickstart";
 import Footer from "./components/Footer";
 
 const snapshot = snapshotData as Snapshot;
 const leaderboard = leaderboardData as LeaderboardData;
+const benchmark = buildBenchmarkView(leaderboard);
 
 export default function App() {
+  const [selectedModelId, setSelectedModelId] = useState(
+    benchmark.models[0]?.id ?? "",
+  );
+
   return (
     <>
       <Nav />
       <main>
-        <Hero data={leaderboard} />
-        <Leaderboard data={leaderboard} />
-        <Integrity data={leaderboard} />
+        <ResultsExplorer
+          data={leaderboard}
+          benchmark={benchmark}
+          selectedModelId={selectedModelId}
+          onSelectModel={setSelectedModelId}
+        />
+        <Analysis
+          benchmark={benchmark}
+          selectedModelId={selectedModelId}
+          onSelectModel={setSelectedModelId}
+        />
         <HowItWorks snapshot={snapshot} />
-        <Results snapshot={snapshot} />
-        <Adapters />
         <Quickstart />
       </main>
       <Footer data={leaderboard} />

web/src/benchmarkData.ts

@@ -0,0 +1,136 @@
+import type {
+  Leaderboard,
+  LeaderboardModel,
+  TieredLeaderboardModel,
+} from "./types";
+
+export type ResultModel = TieredLeaderboardModel & {
+  paired_lift: number;
+  ci95: [number, number];
+  cost_per_episode_usd: number;
+};
+
+export const MECHANICS = [
+  ["cap_free_agency", "Cap & FA"],
+  ["draft", "Draft"],
+  ["information_memory", "Information"],
+  ["lineup", "Lineup"],
+  ["trades", "Trades"],
+] as const;
+
+export interface BenchmarkView {
+  models: ResultModel[];
+  modelCount: number;
+  modelsAboveBar: number;
+  repeats: number;
+  scriptedBar: number;
+  oracle: number;
+}
+
+function finite(value: unknown): value is number {
+  return typeof value === "number" && Number.isFinite(value);
+}
+
+function assertResultModel(model: TieredLeaderboardModel): asserts model is ResultModel {
+  if (!finite(model.mean_score)) {
+    throw new Error(`Leaderboard row ${model.id} is missing a finite mean_score`);
+  }
+  if (!finite(model.paired_lift)) {
+    throw new Error(`Leaderboard row ${model.id} is missing a finite paired_lift`);
+  }
+  if (
+    !Array.isArray(model.ci95) ||
+    model.ci95.length !== 2 ||
+    !finite(model.ci95[0]) ||
+    !finite(model.ci95[1]) ||
+    model.ci95[0] > model.ci95[1]
+  ) {
+    throw new Error(`Leaderboard row ${model.id} has an invalid ci95`);
+  }
+  if (!finite(model.cost_per_episode_usd) || model.cost_per_episode_usd < 0) {
+    throw new Error(`Leaderboard row ${model.id} has an invalid cost_per_episode_usd`);
+  }
+  for (const [key] of MECHANICS) {
+    const outcome = model.mechanic_breakdown[key];
+    if (
+      outcome === undefined ||
+      !finite(outcome.accepted) ||
+      !finite(outcome.rejected) ||
+      outcome.accepted < 0 ||
+      outcome.rejected < 0
+    ) {
+      throw new Error(`Leaderboard row ${model.id} has invalid ${key} outcomes`);
+    }
+  }
+}
+
+export function buildBenchmarkView(data: Leaderboard): BenchmarkView {
+  const ids = new Set<string>();
+  const completeModels: ResultModel[] = data.models.map((model) => {
+    if (ids.has(model.id)) {
+      throw new Error(`Leaderboard contains duplicate model id ${model.id}`);
+    }
+    ids.add(model.id);
+    assertResultModel(model);
+    return model;
+  });
+
+  if (data.models.length === 0 && data.publication.publishable_ranking) {
+    throw new Error("Publishable leaderboard has no model rows");
+  }
+
+  const models = [...completeModels].sort((a, b) => b.mean_score - a.mean_score);
+  const scriptedBar =
+    data.baselines.find((baseline) => baseline.agent === "pick-trader")?.mean_score ??
+    data.headroom.pick_trader;
+  if (!finite(scriptedBar) || !finite(data.headroom.oracle)) {
+    throw new Error("Leaderboard is missing a finite scripted bar or oracle ceiling");
+  }
+
+  const decisionPoints = models[0]?.decision_points ?? 0;
+  const denominator = data.preset.seeds.length * data.preset.decision_points_per_episode;
+  const repeats = denominator > 0 ? decisionPoints / denominator : 0;
+  if (!Number.isInteger(repeats) || repeats < 1) {
+    throw new Error("Leaderboard decision counts do not yield a whole repeat count");
+  }
+  for (const model of models) {
+    if (model.decision_points !== decisionPoints) {
+      throw new Error("Leaderboard rows disagree on decision_points");
+    }
+  }
+
+  return {
+    models,
+    modelCount: models.length,
+    modelsAboveBar: models.filter((model) => model.mean_score > scriptedBar).length,
+    repeats,
+    scriptedBar,
+    oracle: data.headroom.oracle,
+  };
+}
+
+export function shortModelName(model: string): string {
+  return model.split("/").pop() ?? model;
+}
+
+export function issueLabel(issue: string): string {
+  if (issue.includes("illegal actions")) return "Illegal actions";
+  if (issue.includes("fallback")) return "Adapter fallback";
+  if (issue.includes("failed queries")) return "Query errors";
+  if (issue.includes("strongest scripted baseline")) return "Below bar";
+  return "Protocol note";
+}
+
+export function issueLabels(model: LeaderboardModel): string[] {
+  return (model.sota_v2_issues ?? []).map(issueLabel);
+}
+
+export function rejectionRate(
+  model: LeaderboardModel,
+  mechanic: (typeof MECHANICS)[number][0],
+): number {
+  const outcome = model.mechanic_breakdown[mechanic];
+  if (!outcome) return 0;
+  const total = outcome.accepted + outcome.rejected;
+  return total === 0 ? 0 : outcome.rejected / total;
+}