Add Neon leaderboard dashboard (single-viewport, Docker static)

.gitignore

@@ -0,0 +1,4 @@
+.shots/
+__pycache__/
+*.pyc
+.DS_Store

Dockerfile

@@ -0,0 +1,14 @@
+# Static dashboard served on the port Hugging Face Spaces expects (7860).
+# The app is a single HTML page + precomputed data; no backend logic is needed,
+# so a tiny static file server is all that's required.
+FROM python:3.12-slim
+
+WORKDIR /app
+
+# Only the files the browser actually needs.
+COPY index.html app.js data.js data.json ./
+
+EXPOSE 7860
+
+# Serve from /app on 0.0.0.0:7860. index.html is returned for "/".
+CMD ["python", "-m", "http.server", "7860", "--bind", "0.0.0.0"]

README.md

@@ -1,11 +1,100 @@
----
-title: Leaderboard
-emoji: 🔥
-colorFrom: gray
-colorTo: indigo
-sdk: docker
-pinned: false
-short_description: Model performance comparison leaderboard
----
-
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+---
+title: Neon Leaderboard
+emoji: 📊
+colorFrom: green
+colorTo: blue
+sdk: docker
+app_port: 7860
+pinned: false
+short_description: Single-viewport dashboard for the Neon model leaderboard
+---
+
+# Neon Leaderboard
+
+A single-page, single-viewport dashboard for exploring the
+[`neon-router`](https://github.com/NeonGeckoCom/neon-router) weekly leaderboard —
+Neon.ai models and retrieval pipelines evaluated across seven QA benchmarks.
+
+Everything is **static**: no backend, no build step at serve time. The page reads
+one precomputed file (`data.js`) and renders entirely in the browser, so it deploys
+identically to a Hugging Face *Static* Space or GitHub Pages.
+
+## What's inside
+
+| Tab | What it shows |
+|---|---|
+| **Compare** | The full results grid. Pick a benchmark and see every pipeline ranked, or switch to **By model** to follow one pipeline across all benchmarks. Searchable benchmark / model dropdowns, sort by any metric (MRR, speed, cost…), filter to a single model, toggle Neon-only / disqualified rows, and relative color heat per column. |
+| **Picks** | The router's single recommended pipeline per client **profile** × benchmark, with a statistical verdict (winner clear / inconclusive / all disqualified). Hover a profile header for its hard gates. |
+| **Models** | A summary page for any model: pipeline, repo, mean MRR / gen / latency, Pareto and DQ counts, and its metrics across every benchmark. |
+| **Metrics** | All models × all metrics — either **mean scores** (color-coded) or **wins / losses** per metric. Sortable by any metric. |
+| **Guide** | Full metrics glossary, every client profile's hard gates, and run provenance. |
+
+Hover any column header anywhere for a plain-language definition (e.g. *P50 = 50th
+percentile latency*, *MRR = Mean Reciprocal Rank*). A light/dark toggle (◐) sits in
+the top-right; append `?theme=light` to the URL to share a light-mode link, and
+`#picks` / `#models` / `#metrics` / `#guide` to deep-link a tab.
+
+## Faithfulness to the source
+
+`data.js` is generated from the leaderboard's own source of truth — `run.json` plus
+the profile YAMLs — and the picks, Pareto front, disqualifications, verdicts, and
+win/loss tallies are recomputed with logic ported directly from
+`router/pareto.py` and `router/reports/leaderboard.py`. Spot-checked to match
+`reports/dashboard/leaderboard.md` exactly.
+
+## Regenerating the data
+
+When the leaderboard is re-run, rebuild `data.js` from a `neon-router` checkout:
+
+```bash
+python scripts/build_data.py \
+  --run-json   /path/to/neon-router/reports/dashboard/raw/run.json \
+  --profiles-dir /path/to/neon-router/configs/profiles \
+  --out data.js
+```
+
+With a `neon-router` clone sitting next to this folder, the defaults already point
+at it, so `python scripts/build_data.py` is enough. The script needs only
+`pyyaml` (`pip install pyyaml`).
+
+To change which models get the **NEON** badge, edit the `NEON_MODELS` set near the
+top of `scripts/build_data.py` and re-run it.
+
+## Deploying
+
+### Hugging Face Static Space (BrainForge)
+
+1. Create a new Space → SDK **Static** under the `BrainForge` org.
+2. Push `index.html`, `app.js`, `data.js`, and this `README.md` (its front-matter
+   above configures the Space). That's the whole site.
+
+```bash
+# from this folder, with the Space repo as the remote
+git lfs install --skip-repo
+git add index.html app.js data.js README.md scripts/build_data.py
+git commit -m "Neon leaderboard dashboard"
+git push
+```
+
+### GitHub Pages (NeonGeckoCom)
+
+1. Put these files at the repo root (or in `/docs`).
+2. Settings → Pages → deploy from branch, root (or `/docs`).
+No build step is required — it's plain static files.
+
+## Local preview
+
+```bash
+python -m http.server 8000
+# open http://localhost:8000/
+```
+
+Files:
+
+```
+index.html            # markup + styles (single viewport shell)
+app.js                # all rendering / interactions (vanilla JS, no deps)
+data.js               # precomputed dataset (window.LEADERBOARD_DATA)
+data.json             # same data as plain JSON (for non-browser consumers)
+scripts/build_data.py # regenerates data.js / data.json from run.json + profiles
+```

app.js

@@ -0,0 +1,569 @@
+/* Neon Leaderboard - static dashboard.
+   All data is precomputed in data.js (window.LEADERBOARD_DATA) by
+   scripts/build_data.py, faithful to the neon-router report logic. */
+(function () {
+  "use strict";
+  const D = window.LEADERBOARD_DATA;
+  if (!D) { document.body.innerHTML = "<p style='padding:30px'>data.js failed to load.</p>"; return; }
+
+  const $ = (s, r) => (r || document).querySelector(s);
+  const $$ = (s, r) => Array.from((r || document).querySelectorAll(s));
+  const el = (tag, attrs, kids) => {
+    const n = document.createElement(tag);
+    if (attrs) for (const k in attrs) {
+      if (k === "class") n.className = attrs[k];
+      else if (k === "html") n.innerHTML = attrs[k];
+      else if (k === "text") n.textContent = attrs[k];
+      else if (k.startsWith("on") && typeof attrs[k] === "function") n.addEventListener(k.slice(2), attrs[k]);
+      else if (attrs[k] != null) n.setAttribute(k, attrs[k]);
+    }
+    if (kids) (Array.isArray(kids) ? kids : [kids]).forEach(c => c != null && n.appendChild(typeof c === "string" ? document.createTextNode(c) : c));
+    return n;
+  };
+
+  const GLOSS = D.glossary || {};
+  const METRICS = D.metrics_def;            // [{key,label,field,higher,fmt,tip}]
+  const BENCH = D.benchmarks;
+  const AGG = D.aggregates;
+  const CANDS = D.candidates;               // map id -> meta
+
+  const aggBy = {};                         // benchmark -> [agg]
+  const aggByCand = {};                     // candidate_id -> [agg] (across benchmarks)
+  AGG.forEach(a => {
+    (aggBy[a.benchmark] = aggBy[a.benchmark] || []).push(a);
+    (aggByCand[a.candidate_id] = aggByCand[a.candidate_id] || []).push(a);
+  });
+  const CAND_IDS = Object.keys(CANDS);
+  const MODEL_NAMES = Array.from(new Set(CAND_IDS.map(id => CANDS[id].model)));
+  // Prefer a Neon model with generation data as the default detail selection.
+  const DEFAULT_CAND = CAND_IDS.find(id => CANDS[id].is_neon) || CAND_IDS[0];
+
+  // ---- formatting ----
+  function num(v, code) {
+    if (v == null || v === "") return "";
+    const d = { f0: 0, f1: 1, f2: 2, f3: 3, f5: 5 }[code];
+    return Number(v).toFixed(d == null ? 3 : d);
+  }
+  function pipeLabel(c) {
+    const r = c.reranker && c.reranker !== "none" ? " · " + c.reranker : "";
+    return c.model + " · " + c.retriever + r;
+  }
+  function benchShort(b) { return b.replace(/^dataset-/, "").replace(/-qa$/, ""); }
+
+  // ---- tooltip ----
+  const tip = $("#tip");
+  function showTip(html, x, y) {
+    tip.innerHTML = html; tip.style.opacity = "1";
+    const r = tip.getBoundingClientRect();
+    let left = x + 14, top = y + 16;
+    if (left + r.width > innerWidth - 8) left = x - r.width - 12;
+    if (top + r.height > innerHeight - 8) top = y - r.height - 12;
+    tip.style.left = Math.max(6, left) + "px"; tip.style.top = Math.max(6, top) + "px";
+  }
+  function hideTip() { tip.style.opacity = "0"; }
+  function bindTip(node, html) {
+    if (!html) return node;
+    node.addEventListener("mousemove", e => showTip(html, e.clientX, e.clientY));
+    node.addEventListener("mouseleave", hideTip);
+    return node;
+  }
+  function glossTip(term, extra) {
+    const def = GLOSS[term];
+    if (!def && !extra) return null;
+    return "<b>" + term + "</b><br>" + (def || "") + (extra ? "<br><br>" + extra : "");
+  }
+
+  // ---- color heat (diverging, low-glare) ----
+  function heatStyle(v, min, max, higher) {
+    if (v == null || v === "" || max === min) return "";
+    let t = (v - min) / (max - min);            // 0..1, 1 = numerically max
+    if (!higher) t = 1 - t;                       // 1 = good
+    const intensity = Math.abs(t - 0.5) * 2;      // 0 mid .. 1 extreme
+    const hue = t >= 0.5 ? 168 : 28;
+    const dark = document.documentElement.getAttribute("data-theme") !== "light";
+    const light = dark ? 46 : 60;
+    const sat = dark ? 52 : 62;
+    const alpha = (0.05 + intensity * 0.30).toFixed(3);
+    return "background:hsla(" + hue + "," + sat + "%," + light + "%," + alpha + ")";
+  }
+
+  // ---- searchable combobox (single select) ----
+  function Combo(host, opts) {
+    // opts: {options:[{value,label,sub,neon}], value, placeholder, onChange, width}
+    let value = opts.value;
+    let open = false, kbd = -1, filtered = opts.options;
+    host.className = "combo";
+    const btn = el("button", { class: "field", style: "text-align:left;min-width:" + (opts.width || 180) + "px;display:flex;align-items:center;gap:6px" });
+    const menu = el("div", { class: "menu" });
+    const search = el("input", { class: "field", placeholder: "Type to search…", style: "width:100%;margin-bottom:5px" });
+    const list = el("div");
+    menu.appendChild(search); menu.appendChild(list);
+    host.appendChild(btn); host.appendChild(menu);
+
+    function curLabel() { const o = opts.options.find(o => o.value === value); return o ? o.label : (opts.placeholder || "Select…"); }
+    function paintBtn() {
+      btn.innerHTML = "";
+      const o = opts.options.find(o => o.value === value);
+      if (o && o.neon) btn.appendChild(el("span", { class: "badge neon", text: "NEON" }));
+      btn.appendChild(el("span", { text: curLabel(), style: "flex:1;overflow:hidden;text-overflow:ellipsis;white-space:nowrap" }));
+      btn.appendChild(el("span", { text: "▾", style: "color:var(--text-faint);font-size:10px" }));
+    }
+    function paintList() {
+      list.innerHTML = "";
+      filtered.forEach((o, i) => {
+        const row = el("div", { class: "opt" + (o.value === value ? " sel" : "") + (i === kbd ? " kbd" : "") });
+        if (o.neon) row.appendChild(el("span", { class: "badge neon", text: "NEON" }));
+        row.appendChild(el("span", { text: o.label }));
+        if (o.sub) row.appendChild(el("small", { text: o.sub }));
+        row.addEventListener("click", () => { value = o.value; close(); opts.onChange && opts.onChange(value); });
+        list.appendChild(row);
+      });
+      if (!filtered.length) list.appendChild(el("div", { class: "opt", style: "color:var(--text-faint)", text: "No matches" }));
+    }
+    function doFilter() {
+      const q = search.value.trim().toLowerCase();
+      filtered = !q ? opts.options : opts.options.filter(o =>
+        (o.label + " " + (o.sub || "")).toLowerCase().includes(q));
+      kbd = filtered.length ? 0 : -1; paintList();
+    }
+    function openMenu() { open = true; host.classList.add("open"); search.value = ""; doFilter(); setTimeout(() => search.focus(), 0); }
+    function close() { open = false; host.classList.remove("open"); paintBtn(); }
+    btn.addEventListener("click", e => { e.stopPropagation(); open ? close() : openMenu(); });
+    search.addEventListener("input", doFilter);
+    search.addEventListener("keydown", e => {
+      if (e.key === "ArrowDown") { kbd = Math.min(filtered.length - 1, kbd + 1); paintList(); e.preventDefault(); }
+      else if (e.key === "ArrowUp") { kbd = Math.max(0, kbd - 1); paintList(); e.preventDefault(); }
+      else if (e.key === "Enter" && filtered[kbd]) { value = filtered[kbd].value; close(); opts.onChange && opts.onChange(value); }
+      else if (e.key === "Escape") close();
+    });
+    document.addEventListener("click", e => { if (open && !host.contains(e.target)) close(); });
+    paintBtn();
+    return {
+      get value() { return value; },
+      set value(v) { value = v; paintBtn(); },
+      setOptions(o, v) { opts.options = o; if (v !== undefined) value = v; paintBtn(); }
+    };
+  }
+
+  // ===================================================================
+  // State
+  // ===================================================================
+  const S = {
+    tab: "compare",
+    cmpMode: "byBench",
+    bench: BENCH[0],
+    cand: null,            // selected pipeline id (byModel)
+    filterModel: "__all__",
+    sortKey: "mrr",
+    sortDir: -1,           // -1 desc, 1 asc
+    neonOnly: false,
+    showDQ: true,
+    heat: true,
+    metricMode: "scores",
+    metricSort: "mrr",
+    metricNeonOnly: false,
+    metricHeat: true,
+    pageCand: null,
+  };
+
+  // ===================================================================
+  // Header / footer
+  // ===================================================================
+  $("#headerMeta").textContent = (D.date || "") + " · " + D.n_candidates + " pipelines · " +
+    BENCH.length + " benchmarks · " + (D.n_queries || "?") + " queries";
+  const prov = D.provenance || {};
+  $("#provFoot").textContent = (prov.weekly ? "weekly roster sweep" : "sweep") + (prov.merged_at ? " · merged " + prov.merged_at.slice(0, 19).replace("T", " ") + "Z" : "");
+  function paintLegendSwatches() {
+    $("#lgGood").style.cssText = "width:11px;height:11px;border-radius:3px;display:inline-block;" + heatStyle(1, 0, 1, true);
+    $("#lgBad").style.cssText = "width:11px;height:11px;border-radius:3px;display:inline-block;" + heatStyle(0, 0, 1, true);
+  }
+
+  // ===================================================================
+  // Tabs
+  // ===================================================================
+  const CTLS = { compare: "#ctl-compare", picks: "#ctl-picks", models: "#ctl-models", metrics: "#ctl-metrics", guide: "#ctl-guide" };
+  $$("#tabs button").forEach(b => b.addEventListener("click", () => setTab(b.dataset.tab)));
+  function setTab(t) {
+    S.tab = t;
+    try { history.replaceState(null, "", "#" + t); } catch (e) {}
+    $$("#tabs button").forEach(b => b.classList.toggle("active", b.dataset.tab === t));
+    $$(".view").forEach(v => v.classList.toggle("active", v.id === "view-" + t));
+    for (const k in CTLS) $(CTLS[k]).style.display = (k === t ? "flex" : "none");
+    render();
+  }
+
+  // ===================================================================
+  // Compare controls
+  // ===================================================================
+  const benchCombo = Combo($("#benchCombo"), {
+    options: BENCH.map(b => ({ value: b, label: benchShort(b), sub: b })),
+    value: S.bench, width: 170, onChange: v => { S.bench = v; render(); }
+  });
+  const candOptions = () => CAND_IDS.map(id => ({ value: id, label: pipeLabel(CANDS[id]), neon: CANDS[id].is_neon, sub: CANDS[id].is_neon ? "Neon" : (CANDS[id].is_retrieval_only ? "retrieval" : "") }));
+  const modelCombo = Combo($("#modelCombo"), { options: candOptions(), value: DEFAULT_CAND, width: 230, onChange: v => { S.cand = v; render(); } });
+  S.cand = DEFAULT_CAND;
+  const filterCombo = Combo($("#filterCombo"), {
+    options: [{ value: "__all__", label: "All models" }].concat(MODEL_NAMES.map(m => ({ value: m, label: m, neon: (CANDS[CAND_IDS.find(id => CANDS[id].model === m)] || {}).is_neon }))),
+    value: "__all__", width: 160, onChange: v => { S.filterModel = v; render(); }
+  });
+  const sortSel = $("#sortSel");
+  function fillSortSel() {
+    sortSel.innerHTML = "";
+    sortSel.appendChild(el("option", { value: "pipe", text: "Pipeline (A–Z)" }));
+    METRICS.forEach(m => sortSel.appendChild(el("option", { value: m.field, text: m.label })));
+    sortSel.value = S.sortKey;
+  }
+  fillSortSel();
+  sortSel.addEventListener("change", () => { S.sortKey = sortSel.value; render(); });
+  $("#sortDir").addEventListener("click", () => { S.sortDir *= -1; $("#sortDir").textContent = S.sortDir < 0 ? "▾" : "▴"; render(); });
+  $$("#cmpModeSeg button").forEach(b => b.addEventListener("click", () => {
+    S.cmpMode = b.dataset.mode;
+    $$("#cmpModeSeg button").forEach(x => x.classList.toggle("active", x === b));
+    $("#benchPickWrap").style.display = S.cmpMode === "byBench" ? "flex" : "none";
+    $("#modelPickWrap").style.display = S.cmpMode === "byModel" ? "flex" : "none";
+    render();
+  }));
+  $("#neonOnly").addEventListener("change", e => { S.neonOnly = e.target.checked; render(); });
+  $("#showDQ").addEventListener("change", e => { S.showDQ = e.target.checked; render(); });
+  $("#heat").addEventListener("change", e => { S.heat = e.target.checked; render(); });
+
+  // ===================================================================
+  // Compare render
+  // ===================================================================
+  function activeMetricCols(rows) {
+    return METRICS.filter(m => rows.some(r => r[m.field] != null && r[m.field] !== ""));
+  }
+  function renderCompare() {
+    const table = $("#cmpTable");
+    table.innerHTML = "";
+    let rows, flagFor, leftHead, leftCell;
+
+    if (S.cmpMode === "byBench") {
+      const bt = D.bench_tables[S.bench] || { order: [], pareto: [], dq: {} };
+      const paretoSet = new Set(bt.pareto);
+      rows = (aggBy[S.bench] || []).slice();
+      if (S.filterModel !== "__all__") rows = rows.filter(r => r.model === S.filterModel);
+      if (S.neonOnly) rows = rows.filter(r => r.is_neon);
+      if (!S.showDQ) rows = rows.filter(r => !bt.dq[r.candidate_id]);
+      flagFor = r => ({ pareto: paretoSet.has(r.candidate_id), dq: bt.dq[r.candidate_id] });
+      leftHead = "pipeline"; leftCell = r => pipeCell(r, flagFor(r));
+    } else {
+      rows = (aggByCand[S.cand] || []).slice();
+      flagFor = () => ({});
+      leftHead = "benchmark";
+      leftCell = r => bindTip(el("td", { class: "txt" }, el("span", { class: "pipe", text: benchShort(r.benchmark) })), "<b>" + r.benchmark + "</b>");
+    }
+
+    const cols = activeMetricCols(rows);
+    // sort
+    rows.sort((a, b) => {
+      if (S.sortKey === "pipe") {
+        const av = S.cmpMode === "byBench" ? pipeLabel(CANDS[a.candidate_id]) : a.benchmark;
+        const bv = S.cmpMode === "byBench" ? pipeLabel(CANDS[b.candidate_id]) : b.benchmark;
+        return av < bv ? S.sortDir : av > bv ? -S.sortDir : 0;
+      }
+      const av = a[S.sortKey], bv = b[S.sortKey];
+      if (av == null && bv == null) return 0;
+      if (av == null) return 1; if (bv == null) return -1;
+      return (av - bv) * S.sortDir;
+    });
+
+    // column ranges for heat
+    const ranges = {};
+    cols.forEach(m => {
+      const vs = rows.map(r => r[m.field]).filter(v => v != null && v !== "");
+      ranges[m.field] = vs.length ? { min: Math.min(...vs), max: Math.max(...vs) } : null;
+    });
+
+    // head
+    const thead = el("thead");
+    const htr = el("tr");
+    htr.appendChild(bindTip(el("th", { class: "txt lbl sortable", onclick: () => sortByCol("pipe"), html: leftHead + (S.sortKey === "pipe" ? " <span class='arrow'>" + (S.sortDir < 0 ? "▼" : "▲") + "</span>" : "") }),
+      S.cmpMode === "byBench" ? glossTip("pipeline", "model · retriever · reranker") : "<b>benchmark</b><br>The evaluation dataset for this row."));
+    cols.forEach(m => {
+      const arrow = S.sortKey === m.field ? " <span class='arrow'>" + (S.sortDir < 0 ? "▼" : "▲") + "</span>" : "";
+      htr.appendChild(bindTip(el("th", { class: "sortable", onclick: () => sortByCol(m.field), html: m.label + arrow }),
+        glossTip(m.tip, (m.higher ? "Higher is better." : "Lower is better."))));
+    });
+    thead.appendChild(htr); table.appendChild(thead);
+
+    // body
+    const tb = el("tbody");
+    if (!rows.length) { tb.appendChild(el("tr", null, el("td", { colspan: cols.length + 1, class: "empty", text: "No rows match the current filters." }))); }
+    rows.forEach(r => {
+      const tr = el("tr", { class: r.is_neon ? "neon" : "" });
+      tr.appendChild(leftCell(r));
+      cols.forEach(m => {
+        const v = r[m.field];
+        const td = el("td", { text: num(v, m.fmt) });
+        if (S.heat && ranges[m.field]) td.style.cssText = heatStyle(v, ranges[m.field].min, ranges[m.field].max, m.higher);
+        tr.appendChild(td);
+      });
+      tb.appendChild(tr);
+    });
+    table.appendChild(tb);
+  }
+  function sortByCol(k) { if (S.sortKey === k) S.sortDir *= -1; else { S.sortKey = k; S.sortDir = (k === "pipe") ? 1 : -1; } sortSel.value = k; $("#sortDir").textContent = S.sortDir < 0 ? "▾" : "▴"; render(); }
+
+  function pipeCell(r, flag) {
+    const c = CANDS[r.candidate_id];
+    const inner = el("span", { class: "pipe" });
+    if (flag.pareto) inner.appendChild(el("span", { class: "star", text: "★ " }));
+    inner.appendChild(el("b", { text: c.model }));
+    inner.appendChild(el("small", { text: "  " + c.retriever + (c.reranker !== "none" ? " · " + c.reranker : "") }));
+    const td = el("td", { class: "txt" });
+    if (c.is_neon) td.appendChild(el("span", { class: "badge neon", text: "NEON", style: "margin-right:6px" }));
+    td.appendChild(inner);
+    if (flag.dq) td.appendChild(el("span", { class: "badge dq", text: "DQ:" + flag.dq, style: "margin-left:7px" }));
+    return bindTip(td, "<b>" + pipeLabel(c) + "</b>" +
+      (c.model_repo ? "<br>repo: " + c.model_repo : "") +
+      (flag.dq ? "<br><br>Disqualified by a profile's <b>" + flag.dq + "</b> gate." : "") +
+      (flag.pareto ? "<br><br>★ On the Pareto front (not dominated on accuracy / latency / build-cost)." : ""));
+  }
+
+  // ===================================================================
+  // Picks render
+  // ===================================================================
+  const VCLASS = { "winner clear": "v-clear", "inconclusive": "v-inconclusive", "all disqualified": "v-dq", "no winner": "v-none", "no signal": "v-none" };
+  function renderPicks() {
+    const table = $("#picksTable"); table.innerHTML = "";
+    const profs = D.profiles;
+    const thead = el("thead"); const htr = el("tr");
+    htr.appendChild(el("th", { class: "txt", text: "benchmark", style: "text-align:left" }));
+    profs.forEach(p => {
+      const extra = "Hard gates: " + (p.gates.join(" · ") || "—") + "<br>Ranks on: " + p.metric +
+        (p.corpus_profile ? "<br>Corpus: " + p.corpus_profile : "") +
+        (p.deployment_constraint ? "<br>Deploy: " + p.deployment_constraint : "");
+      htr.appendChild(bindTip(el("th", { html: p.name.replace(/_/g, " ") }), glossTip("profile", extra)));
+    });
+    thead.appendChild(htr); table.appendChild(thead);
+
+    const tb = el("tbody");
+    BENCH.forEach(b => {
+      const tr = el("tr");
+      tr.appendChild(bindTip(el("th", { text: benchShort(b) }), "<b>" + b + "</b>"));
+      const row = D.picks[b] || {};
+      profs.forEach(p => {
+        const cell = row[p.name];
+        const td = el("td");
+        if (cell && cell.pick) {
+          const c = CANDS[cell.pick];
+          const wrap = el("div");
+          const line = el("div", { style: "font-family:var(--mono);font-size:11.5px;margin-bottom:4px" });
+          if (c && c.is_neon) line.appendChild(el("span", { class: "badge neon", text: "NEON", style: "margin-right:5px" }));
+          line.appendChild(el("span", { text: cell.model + " · " + cell.retriever }));
+          wrap.appendChild(line);
+          const mv = cell.metric === "generation_acc" ? "gen " + num(cell.metric_val, "f2") : "H@1 " + num(cell.metric_val, "f2");
+          wrap.appendChild(bindTip(el("span", { class: "verdict " + (VCLASS[cell.verdict] || "v-none"), text: cell.verdict }),
+            glossTip("verdict") ));
+          wrap.appendChild(el("span", { class: "cell-sub", text: "MRR " + num(cell.mrr, "f3") + " · " + mv + " · p50 " + num(cell.p50, "f0") + "ms" }));
+          td.appendChild(wrap);
+        } else {
+          td.appendChild(el("span", { class: "verdict " + (VCLASS[(cell && cell.verdict)] || "v-dq"), text: (cell && cell.verdict) || "—" }));
+        }
+        tr.appendChild(td);
+      });
+      tb.appendChild(tr);
+    });
+    table.appendChild(tb);
+  }
+
+  // ===================================================================
+  // Models render
+  // ===================================================================
+  const modelPageCombo = Combo($("#modelPageCombo"), { options: candOptions(), value: DEFAULT_CAND, width: 240, onChange: v => { S.pageCand = v; render(); } });
+  S.pageCand = DEFAULT_CAND;
+  function renderModels() {
+    const host = $("#modelPage"); host.innerHTML = "";
+    const id = S.pageCand; const c = CANDS[id];
+    const rows = (aggByCand[id] || []).slice().sort((a, b) => benchShort(a.benchmark) < benchShort(b.benchmark) ? -1 : 1);
+    $("#modelPageBadge").innerHTML = c.is_neon ? "<span class='badge neon'>NEON.ai model</span>" : (c.is_retrieval_only ? "<span class='badge pareto'>retrieval baseline</span>" : "");
+
+    // aggregate stats
+    const mean = f => { const vs = rows.map(r => r[f]).filter(v => v != null); return vs.length ? vs.reduce((a, b) => a + b, 0) / vs.length : null; };
+    let paretoCount = 0; BENCH.forEach(b => { const bt = D.bench_tables[b]; if (bt && bt.pareto.includes(id)) paretoCount++; });
+    let dqCount = 0; BENCH.forEach(b => { const bt = D.bench_tables[b]; if (bt && bt.dq[id]) dqCount++; });
+
+    // left card
+    const left = el("div", { class: "card" });
+    left.appendChild(el("h3", { style: "margin:0 0 10px;font-size:14px", text: c.model }));
+    const kvs = [
+      ["pipeline", c.retriever + (c.reranker !== "none" ? " · " + c.reranker : "")],
+      ["model repo", c.model_repo || "—"],
+      ["serving", c.model_adapter || "—"],
+      ["benchmarks", String(rows.length)],
+      ["type", c.is_neon ? "Neon.ai model" : (c.is_retrieval_only ? "retrieval-only" : "external")],
+    ];
+    kvs.forEach(([k, v]) => left.appendChild(el("div", { class: "kv" }, [el("span", { class: "k", text: k }), el("span", { class: "v", text: v })])));
+    const stats = el("div", { class: "stat-row", style: "margin-top:14px" });
+    const stat = (n, l, tipTerm) => { const s = el("div", { class: "stat" }, [el("div", { class: "n", text: n }), bindTip(el("div", { class: "l", text: l }), glossTip(tipTerm))]); stats.appendChild(s); };
+    stat(num(mean("mrr"), "f3"), "mean MRR", "MRR");
+    const gm = mean("generation_acc"); stat(gm == null ? "—" : num(gm, "f2"), "mean gen", "gen");
+    stat(num(mean("p50_retrieval_latency_ms"), "f0"), "p50 ms", "P50");
+    stat(String(paretoCount), "Pareto ★", "Pareto");
+    stat(String(dqCount), "DQ count", "DQ");
+    left.appendChild(stats);
+    host.appendChild(left);
+
+    // right card: metrics across benchmarks
+    const right = el("div", { class: "card scroll" });
+    const cols = activeMetricCols(rows);
+    const table = el("table");
+    const thead = el("thead"); const htr = el("tr");
+    htr.appendChild(el("th", { class: "txt", text: "benchmark", style: "text-align:left;position:sticky;top:0;background:var(--bg-elev)" }));
+    cols.forEach(m => htr.appendChild(bindTip(el("th", { html: m.label, style: "position:sticky;top:0;background:var(--bg-elev)" }), glossTip(m.tip, m.higher ? "Higher is better." : "Lower is better."))));
+    thead.appendChild(htr); table.appendChild(thead);
+    const ranges = {}; cols.forEach(m => { const vs = rows.map(r => r[m.field]).filter(v => v != null); ranges[m.field] = vs.length ? { min: Math.min(...vs), max: Math.max(...vs) } : null; });
+    const tb = el("tbody");
+    rows.forEach(r => {
+      const bt = D.bench_tables[r.benchmark];
+      const tr = el("tr");
+      const star = bt && bt.pareto.includes(id) ? el("span", { class: "star", text: "★ " }) : null;
+      const dq = bt && bt.dq[id] ? el("span", { class: "badge dq", text: "DQ:" + bt.dq[id], style: "margin-left:6px" }) : null;
+      const tdL = el("td", { class: "txt" }); if (star) tdL.appendChild(star); tdL.appendChild(el("span", { text: benchShort(r.benchmark) })); if (dq) tdL.appendChild(dq);
+      tr.appendChild(tdL);
+      cols.forEach(m => { const v = r[m.field]; const td = el("td", { text: num(v, m.fmt) }); if (ranges[m.field]) td.style.cssText = heatStyle(v, ranges[m.field].min, ranges[m.field].max, m.higher); tr.appendChild(td); });
+      tb.appendChild(tr);
+    });
+    table.appendChild(tb); right.appendChild(table); host.appendChild(right);
+  }
+
+  // ===================================================================
+  // Metrics render
+  // ===================================================================
+  const metricSortSel = $("#metricSortSel");
+  function fillMetricSort() {
+    metricSortSel.innerHTML = "";
+    metricSortSel.appendChild(el("option", { value: "name", text: "Pipeline (A–Z)" }));
+    METRICS.forEach(m => metricSortSel.appendChild(el("option", { value: m.key, text: m.label })));
+    metricSortSel.value = S.metricSort;
+  }
+  fillMetricSort();
+  metricSortSel.addEventListener("change", () => { S.metricSort = metricSortSel.value; render(); });
+  $$("#metricModeSeg button").forEach(b => b.addEventListener("click", () => { S.metricMode = b.dataset.mode; $$("#metricModeSeg button").forEach(x => x.classList.toggle("active", x === b)); render(); }));
+  $("#metricNeonOnly").addEventListener("change", e => { S.metricNeonOnly = e.target.checked; render(); });
+  $("#metricHeat").addEventListener("change", e => { S.metricHeat = e.target.checked; render(); });
+
+  function renderMetrics() {
+    const table = $("#metricsTable"); table.innerHTML = "";
+    const stats = D.metric_stats;          // key -> cid -> {wins,losses,mean,count}
+    let ids = CAND_IDS.slice();
+    if (S.metricNeonOnly) ids = ids.filter(id => CANDS[id].is_neon);
+
+    // sort rows
+    if (S.metricSort === "name") ids.sort((a, b) => pipeLabel(CANDS[a]) < pipeLabel(CANDS[b]) ? -1 : 1);
+    else {
+      const mdef = METRICS.find(m => m.key === S.metricSort);
+      const st = stats[S.metricSort] || {};
+      ids.sort((a, b) => {
+        const sa = st[a], sb = st[b];
+        const va = S.metricMode === "winloss" ? (sa ? sa.wins - sa.losses : -1e9) : (sa && sa.mean != null ? sa.mean : (mdef.higher ? -1e9 : 1e9));
+        const vb = S.metricMode === "winloss" ? (sb ? sb.wins - sb.losses : -1e9) : (sb && sb.mean != null ? sb.mean : (mdef.higher ? -1e9 : 1e9));
+        return S.metricMode === "winloss" ? (vb - va) : ((vb - va) * (mdef.higher ? 1 : -1));
+      });
+    }
+
+    const thead = el("thead"); const htr = el("tr");
+    htr.appendChild(el("th", { class: "txt", text: "pipeline", style: "text-align:left;position:sticky;left:0;z-index:6" }));
+    METRICS.forEach(m => htr.appendChild(bindTip(el("th", { html: m.label }), glossTip(m.tip, m.higher ? "Higher is better." : "Lower is better."))));
+    thead.appendChild(htr); table.appendChild(thead);
+
+    // ranges for mean heat per metric
+    const ranges = {};
+    METRICS.forEach(m => { const vs = ids.map(id => (stats[m.key][id] || {}).mean).filter(v => v != null); ranges[m.key] = vs.length ? { min: Math.min(...vs), max: Math.max(...vs) } : null; });
+
+    const tb = el("tbody");
+    ids.forEach(id => {
+      const c = CANDS[id];
+      const tr = el("tr", { class: c.is_neon ? "neon" : "" });
+      const th = el("th", { style: "font-family:var(--mono);font-size:11.5px" });
+      if (c.is_neon) th.appendChild(el("span", { class: "badge neon", text: "NEON", style: "margin-right:5px" }));
+      th.appendChild(el("span", { text: pipeLabel(c) }));
+      bindTip(th, "<b>" + pipeLabel(c) + "</b>" + (c.model_repo ? "<br>repo: " + c.model_repo : ""));
+      tr.appendChild(th);
+      METRICS.forEach(m => {
+        const s = (stats[m.key] || {})[id];
+        const td = el("td");
+        if (!s) { td.textContent = ""; }
+        else if (S.metricMode === "winloss") {
+          td.innerHTML = "<span style='color:var(--neon)'>" + s.wins + "</span> – <span style='color:#ff9b7d'>" + s.losses + "</span>";
+          const net = s.wins - s.losses;
+          if (S.metricHeat) td.style.cssText = heatStyle(net, -3, 3, true);
+        } else {
+          td.textContent = num(s.mean, m.fmt);
+          if (S.metricHeat && ranges[m.key]) td.style.cssText = heatStyle(s.mean, ranges[m.key].min, ranges[m.key].max, m.higher);
+        }
+        tr.appendChild(td);
+      });
+      tb.appendChild(tr);
+    });
+    table.appendChild(tb);
+  }
+
+  // ===================================================================
+  // Guide render
+  // ===================================================================
+  function renderGuide() {
+    const host = $("#guideGrid"); host.innerHTML = "";
+    // left: glossary
+    const left = el("div", { class: "guide-col" });
+    left.appendChild(el("h3", { text: "Metrics glossary" }));
+    const gt = el("table", { class: "gloss" });
+    const order = ["MRR", "H@1", "H@5", "nDCG", "P50", "P95", "build", "gen", "gen ms", "tok/s", "gen $", "$/q", "par", "shift", "n", "Pareto", "DQ", "verdict"];
+    order.forEach(k => { if (GLOSS[k]) gt.appendChild(el("tr", null, [el("td", { class: "term", text: k }), el("td", { class: "def", text: GLOSS[k] })])); });
+    left.appendChild(gt); host.appendChild(left);
+    // right: profiles + provenance
+    const right = el("div", { class: "guide-col" });
+    right.appendChild(el("h3", { text: "Client profiles (hard gates)" }));
+    D.profiles.forEach(p => {
+      const box = el("div", { class: "prof" });
+      box.appendChild(el("h4", { text: p.name.replace(/_/g, " ") + "  —  ranks on " + p.metric }));
+      const gates = el("div", { class: "gates" });
+      (p.gates.length ? p.gates : ["no hard gates"]).forEach(g => gates.appendChild(el("span", { class: "gate", text: g })));
+      box.appendChild(gates);
+      const sub = [];
+      if (p.corpus_profile) sub.push("corpus: " + p.corpus_profile);
+      if (p.deployment_constraint) sub.push("deploy: " + p.deployment_constraint);
+      if (sub.length) box.appendChild(el("div", { style: "color:var(--text-faint);font-size:11px;margin-top:6px", text: sub.join("  ·  ") }));
+      right.appendChild(box);
+    });
+    right.appendChild(el("h3", { text: "Provenance", style: "margin-top:10px" }));
+    const pv = el("div", { class: "prof" });
+    const sources = (prov.merged_from || []);
+    pv.appendChild(el("div", { style: "font-size:11.5px;color:var(--text-dim)", text: "Type: " + (prov.weekly ? "weekly roster sweep" : "sweep") }));
+    if (prov.merged_at) pv.appendChild(el("div", { style: "font-size:11.5px;color:var(--text-dim)", text: "Merged at: " + prov.merged_at }));
+    sources.forEach(s => pv.appendChild(el("div", { style: "font-family:var(--mono);font-size:10.5px;color:var(--text-faint);margin-top:3px", text: s })));
+    right.appendChild(pv);
+    host.appendChild(right);
+  }
+
+  // ===================================================================
+  // Theme
+  // ===================================================================
+  function applyTheme(t) {
+    document.documentElement.setAttribute("data-theme", t);
+    try { localStorage.setItem("nlb-theme", t); } catch (e) {}
+    paintLegendSwatches(); render();
+  }
+  $("#themeBtn").addEventListener("click", () => applyTheme(document.documentElement.getAttribute("data-theme") === "light" ? "dark" : "light"));
+  let savedTheme = "dark"; try { savedTheme = localStorage.getItem("nlb-theme") || "dark"; } catch (e) {}
+  const qTheme = new URLSearchParams(location.search).get("theme");
+  if (qTheme === "light" || qTheme === "dark") savedTheme = qTheme;
+  document.documentElement.setAttribute("data-theme", savedTheme);
+
+  // ===================================================================
+  // Master render
+  // ===================================================================
+  function render() {
+    if (S.tab === "compare") renderCompare();
+    else if (S.tab === "picks") renderPicks();
+    else if (S.tab === "models") renderModels();
+    else if (S.tab === "metrics") renderMetrics();
+    else if (S.tab === "guide") renderGuide();
+  }
+
+  paintLegendSwatches();
+  const initTab = (location.hash || "").replace("#", "");
+  if (["compare", "picks", "models", "metrics", "guide"].includes(initTab)) setTab(initTab);
+  else render();
+})();

index.html

@@ -0,0 +1,313 @@
+<!doctype html>
+<html lang="en">
+<head>
+<meta charset="utf-8" />
+<meta name="viewport" content="width=device-width, initial-scale=1, viewport-fit=cover" />
+<title>Neon Leaderboard</title>
+<style>
+  :root {
+    --bg: #0f1419;
+    --bg-elev: #161c24;
+    --bg-elev2: #1c242e;
+    --panel: #131a22;
+    --border: #283441;
+    --border-soft: #1f2a35;
+    --text: #d7dee7;
+    --text-dim: #93a1b0;
+    --text-faint: #64727f;
+    --accent: #3fb6a8;          /* neon teal */
+    --accent-dim: #2c7a72;
+    --neon: #5ad1c2;
+    --good-h: 168;              /* teal-green */
+    --bad-h: 28;                /* warm amber */
+    --chip-bg: #1d2731;
+    --shadow: 0 4px 18px rgba(0,0,0,.35);
+    --radius: 9px;
+    --row-h: 26px;
+    --font: -apple-system, BlinkMacSystemFont, "Segoe UI", Roboto, Helvetica, Arial, sans-serif;
+    --mono: ui-monospace, "SF Mono", "Cascadia Code", "Roboto Mono", Menlo, Consolas, monospace;
+  }
+  html[data-theme="light"] {
+    --bg: #f3f5f8;
+    --bg-elev: #ffffff;
+    --bg-elev2: #eef1f5;
+    --panel: #ffffff;
+    --border: #d4dbe3;
+    --border-soft: #e4e9ef;
+    --text: #1f2933;
+    --text-dim: #56636f;
+    --text-faint: #8a97a3;
+    --accent: #1f8f82;
+    --accent-dim: #5fb5ab;
+    --neon: #178a7c;
+    --chip-bg: #eef2f6;
+    --shadow: 0 4px 18px rgba(40,55,75,.12);
+  }
+
+  * { box-sizing: border-box; }
+  html, body { height: 100%; margin: 0; }
+  body {
+    background: var(--bg);
+    color: var(--text);
+    font-family: var(--font);
+    font-size: 13px;
+    line-height: 1.35;
+    -webkit-font-smoothing: antialiased;
+    overflow: hidden;            /* single viewport, no page scroll */
+  }
+  button { font-family: inherit; color: inherit; cursor: pointer; }
+  select, input { font-family: inherit; font-size: 12.5px; }
+
+  .app { display: flex; flex-direction: column; height: 100vh; height: 100dvh; }
+
+  /* ---------- Header ---------- */
+  header.top {
+    display: flex; align-items: center; gap: 18px;
+    padding: 0 16px; height: 50px; flex: 0 0 auto;
+    background: var(--bg-elev); border-bottom: 1px solid var(--border);
+  }
+  .brand { display: flex; align-items: baseline; gap: 9px; white-space: nowrap; }
+  .brand .dot { width: 9px; height: 9px; border-radius: 50%; background: var(--accent);
+    box-shadow: 0 0 9px var(--accent); align-self: center; }
+  .brand h1 { font-size: 15px; font-weight: 650; margin: 0; letter-spacing: .2px; }
+  .brand .meta { color: var(--text-faint); font-size: 11px; }
+
+  nav.tabs { display: flex; gap: 2px; margin-left: 4px; }
+  nav.tabs button {
+    background: transparent; border: 0; padding: 8px 14px; border-radius: 7px;
+    color: var(--text-dim); font-size: 13px; font-weight: 550; letter-spacing: .2px;
+    transition: background .12s, color .12s;
+  }
+  nav.tabs button:hover { background: var(--bg-elev2); color: var(--text); }
+  nav.tabs button.active { background: var(--accent); color: #06231f; }
+  html[data-theme="light"] nav.tabs button.active { color: #fff; }
+
+  .spacer { flex: 1 1 auto; }
+  .icon-btn {
+    background: var(--bg-elev2); border: 1px solid var(--border); border-radius: 7px;
+    width: 32px; height: 32px; display: grid; place-items: center; color: var(--text-dim);
+    font-size: 15px;
+  }
+  .icon-btn:hover { color: var(--text); border-color: var(--accent-dim); }
+
+  /* ---------- Controls bar ---------- */
+  .controls {
+    display: flex; align-items: center; gap: 14px; flex-wrap: wrap;
+    padding: 9px 16px; flex: 0 0 auto;
+    background: var(--panel); border-bottom: 1px solid var(--border-soft);
+    min-height: 46px;
+  }
+  .ctl { display: flex; align-items: center; gap: 7px; }
+  .ctl label { color: var(--text-faint); font-size: 11px; text-transform: uppercase; letter-spacing: .6px; }
+  .field {
+    background: var(--bg-elev2); border: 1px solid var(--border); color: var(--text);
+    border-radius: 7px; padding: 6px 9px; min-width: 90px; outline: none;
+  }
+  .field:focus { border-color: var(--accent-dim); }
+  select.field { padding-right: 26px; appearance: none;
+    background-image: linear-gradient(45deg, transparent 50%, var(--text-faint) 50%),
+                      linear-gradient(135deg, var(--text-faint) 50%, transparent 50%);
+    background-position: calc(100% - 14px) 50%, calc(100% - 9px) 50%;
+    background-size: 5px 5px, 5px 5px; background-repeat: no-repeat; }
+  .seg { display: inline-flex; background: var(--bg-elev2); border: 1px solid var(--border);
+    border-radius: 7px; overflow: hidden; }
+  .seg button { background: transparent; border: 0; padding: 6px 11px; color: var(--text-dim); font-size: 12px; }
+  .seg button.active { background: var(--accent); color: #06231f; }
+  html[data-theme="light"] .seg button.active { color: #fff; }
+  .toggle { display: inline-flex; align-items: center; gap: 6px; color: var(--text-dim); font-size: 12px; user-select: none; }
+  .toggle input { accent-color: var(--accent); }
+
+  /* combobox (searchable dropdown) */
+  .combo { position: relative; }
+  .combo .menu {
+    position: absolute; z-index: 40; top: calc(100% + 4px); left: 0; min-width: 240px; max-width: 360px;
+    max-height: 320px; overflow: auto; background: var(--bg-elev); border: 1px solid var(--border);
+    border-radius: 8px; box-shadow: var(--shadow); padding: 5px; display: none;
+  }
+  .combo.open .menu { display: block; }
+  .combo .opt { padding: 6px 9px; border-radius: 6px; cursor: pointer; display: flex; align-items: center; gap: 8px; white-space: nowrap; }
+  .combo .opt:hover, .combo .opt.kbd { background: var(--bg-elev2); }
+  .combo .opt.sel { color: var(--accent); }
+  .combo .opt small { color: var(--text-faint); margin-left: auto; }
+
+  /* ---------- Main ---------- */
+  main { flex: 1 1 auto; min-height: 0; display: flex; position: relative; }
+  .view { display: none; flex: 1; min-height: 0; flex-direction: column; padding: 12px 16px 8px; }
+  .view.active { display: flex; }
+
+  .table-wrap { flex: 1 1 auto; min-height: 0; overflow: auto; border: 1px solid var(--border-soft);
+    border-radius: var(--radius); background: var(--panel); }
+  table { border-collapse: separate; border-spacing: 0; width: 100%; }
+  thead th {
+    position: sticky; top: 0; z-index: 5; background: var(--bg-elev);
+    text-align: right; padding: 7px 9px; font-size: 11px; font-weight: 600; color: var(--text-dim);
+    border-bottom: 1px solid var(--border); white-space: nowrap; cursor: help; user-select: none;
+  }
+  thead th.lbl, thead th.sortable { cursor: pointer; }
+  thead th.txt { text-align: left; }
+  thead th .arrow { color: var(--accent); font-size: 9px; margin-left: 3px; }
+  tbody td { padding: 4px 9px; text-align: right; font-variant-numeric: tabular-nums;
+    border-bottom: 1px solid var(--border-soft); white-space: nowrap; height: var(--row-h); }
+  tbody td.txt { text-align: left; font-family: var(--mono); font-size: 12px; }
+  tbody tr:hover td { background: rgba(127,150,170,.07); }
+  tbody tr.neon td.txt .pipe { font-weight: 600; }
+
+  .badge { display: inline-block; padding: 1px 6px; border-radius: 999px; font-size: 10px; font-weight: 650;
+    letter-spacing: .3px; vertical-align: middle; }
+  .badge.neon { background: rgba(90,209,194,.16); color: var(--neon); border: 1px solid rgba(90,209,194,.3); }
+  .badge.pareto { background: rgba(120,170,255,.14); color: #8fb6ff; }
+  .badge.dq { background: rgba(255,120,90,.13); color: #ff9b7d; border: 1px solid rgba(255,120,90,.25); }
+  html[data-theme="light"] .badge.pareto { color: #2f5fc0; }
+  .star { color: #8fb6ff; font-size: 11px; }
+  .pipe small { color: var(--text-faint); }
+
+  .verdict { display: inline-block; padding: 2px 8px; border-radius: 6px; font-size: 11px; font-weight: 600; }
+  .v-clear { background: rgba(63,182,168,.18); color: var(--neon); }
+  .v-inconclusive { background: rgba(190,170,90,.16); color: #d9c178; }
+  .v-dq { background: rgba(255,120,90,.13); color: #ff9b7d; }
+  .v-none { background: var(--chip-bg); color: var(--text-faint); }
+
+  /* Picks grid + Metrics grid share matrix styles */
+  .matrix-wrap { flex: 1 1 auto; min-height: 0; overflow: auto; border: 1px solid var(--border-soft);
+    border-radius: var(--radius); background: var(--panel); }
+  .matrix { width: 100%; border-collapse: separate; border-spacing: 0; }
+  .matrix th, .matrix td { padding: 8px 10px; border-bottom: 1px solid var(--border-soft); border-right: 1px solid var(--border-soft); }
+  .matrix thead th { position: sticky; top: 0; background: var(--bg-elev); z-index: 5; cursor: help; font-size: 11.5px; }
+  .matrix tbody th { position: sticky; left: 0; background: var(--bg-elev); z-index: 4; text-align: left; cursor: help; }
+  .matrix td { text-align: center; font-variant-numeric: tabular-nums; }
+  .matrix .cell-sub { display: block; font-size: 10px; color: var(--text-faint); margin-top: 2px; }
+
+  /* Models tab */
+  .model-page { flex: 1 1 auto; min-height: 0; display: grid; grid-template-columns: 280px 1fr; gap: 14px; }
+  .card { background: var(--panel); border: 1px solid var(--border-soft); border-radius: var(--radius); padding: 16px; }
+  .card.scroll { overflow: auto; }
+  .kv { display: flex; justify-content: space-between; gap: 12px; padding: 6px 0; border-bottom: 1px solid var(--border-soft); font-size: 12.5px; }
+  .kv:last-child { border-bottom: 0; }
+  .kv .k { color: var(--text-faint); }
+  .kv .v { color: var(--text); text-align: right; font-family: var(--mono); }
+  .stat-row { display: flex; gap: 10px; flex-wrap: wrap; margin-bottom: 12px; }
+  .stat { background: var(--bg-elev2); border: 1px solid var(--border-soft); border-radius: 8px; padding: 9px 13px; min-width: 92px; }
+  .stat .n { font-size: 19px; font-weight: 650; font-variant-numeric: tabular-nums; }
+  .stat .l { font-size: 10.5px; color: var(--text-faint); text-transform: uppercase; letter-spacing: .5px; cursor: help; }
+
+  /* Guide tab */
+  .guide-grid { flex: 1 1 auto; min-height: 0; display: grid; grid-template-columns: 1fr 1fr; gap: 14px; }
+  .guide-col { overflow: auto; }
+  .guide-col h3 { margin: 0 0 8px; font-size: 13px; color: var(--accent); }
+  .gloss { width: 100%; border-collapse: collapse; }
+  .gloss td { padding: 5px 8px; border-bottom: 1px solid var(--border-soft); vertical-align: top; font-size: 12px; }
+  .gloss td.term { color: var(--text); font-weight: 600; white-space: nowrap; width: 78px; font-family: var(--mono); }
+  .gloss td.def { color: var(--text-dim); }
+  .prof { background: var(--bg-elev2); border: 1px solid var(--border-soft); border-radius: 8px; padding: 10px 12px; margin-bottom: 8px; }
+  .prof h4 { margin: 0 0 4px; font-size: 12.5px; }
+  .prof .gates { display: flex; gap: 6px; flex-wrap: wrap; }
+  .gate { background: var(--chip-bg); border-radius: 6px; padding: 2px 8px; font-size: 11px; color: var(--text-dim); font-family: var(--mono); }
+
+  footer.bottom { flex: 0 0 auto; display: flex; align-items: center; gap: 16px; padding: 4px 16px;
+    height: 26px; background: var(--bg-elev); border-top: 1px solid var(--border); font-size: 10.5px; color: var(--text-faint); }
+  footer .lg { display: inline-flex; align-items: center; gap: 5px; }
+  footer .sw { width: 11px; height: 11px; border-radius: 3px; display: inline-block; }
+
+  /* tooltip */
+  #tip { position: fixed; z-index: 200; max-width: 320px; background: #0a0e12; color: #e7edf3;
+    border: 1px solid var(--border); border-radius: 8px; padding: 9px 11px; font-size: 12px; line-height: 1.45;
+    box-shadow: var(--shadow); pointer-events: none; opacity: 0; transition: opacity .1s; }
+  html[data-theme="light"] #tip { background: #1f2933; color: #f3f5f8; border-color: #1f2933; }
+  #tip b { color: var(--neon); }
+
+  .empty { color: var(--text-faint); padding: 30px; text-align: center; font-size: 13px; }
+  ::-webkit-scrollbar { width: 10px; height: 10px; }
+  ::-webkit-scrollbar-thumb { background: var(--border); border-radius: 6px; }
+  ::-webkit-scrollbar-track { background: transparent; }
+</style>
+</head>
+<body>
+<div class="app">
+  <header class="top">
+    <div class="brand"><span class="dot"></span><h1>Neon Leaderboard</h1><span class="meta" id="headerMeta"></span></div>
+    <nav class="tabs" id="tabs">
+      <button data-tab="compare" class="active">Compare</button>
+      <button data-tab="picks">Picks</button>
+      <button data-tab="models">Models</button>
+      <button data-tab="metrics">Metrics</button>
+      <button data-tab="guide">Guide</button>
+    </nav>
+    <span class="spacer"></span>
+    <button class="icon-btn" id="themeBtn" title="Toggle light / dark">◐</button>
+  </header>
+
+  <!-- COMPARE -->
+  <div class="controls" id="ctl-compare">
+    <div class="ctl"><label>View</label>
+      <div class="seg" id="cmpModeSeg">
+        <button data-mode="byBench" class="active">By benchmark</button>
+        <button data-mode="byModel">By model</button>
+      </div>
+    </div>
+    <div class="ctl" id="benchPickWrap"><label>Benchmark</label><div class="combo" id="benchCombo"></div></div>
+    <div class="ctl" id="modelPickWrap" style="display:none"><label>Model</label><div class="combo" id="modelCombo"></div></div>
+    <div class="ctl"><label>Filter model</label><div class="combo" id="filterCombo"></div></div>
+    <div class="ctl"><label>Sort</label><select class="field" id="sortSel"></select>
+      <button class="icon-btn" id="sortDir" title="Sort direction" style="width:30px;height:30px">▾</button>
+    </div>
+    <span class="spacer"></span>
+    <label class="toggle"><input type="checkbox" id="neonOnly" /> Neon only</label>
+    <label class="toggle"><input type="checkbox" id="showDQ" checked /> Show DQ</label>
+    <label class="toggle"><input type="checkbox" id="heat" checked /> Color</label>
+  </div>
+
+  <!-- PICKS -->
+  <div class="controls" id="ctl-picks" style="display:none">
+    <div class="ctl"><label>Router recommendations</label></div>
+    <span class="meta" style="color:var(--text-faint);font-size:11.5px">Pick = model · retriever · reranker per client profile. Hover a profile column for its hard gates.</span>
+  </div>
+
+  <!-- MODELS -->
+  <div class="controls" id="ctl-models" style="display:none">
+    <div class="ctl"><label>Model</label><div class="combo" id="modelPageCombo"></div></div>
+    <span class="meta" id="modelPageBadge"></span>
+  </div>
+
+  <!-- METRICS -->
+  <div class="controls" id="ctl-metrics" style="display:none">
+    <div class="ctl"><label>Show</label>
+      <div class="seg" id="metricModeSeg">
+        <button data-mode="scores" class="active">Mean scores</button>
+        <button data-mode="winloss">Wins / losses</button>
+      </div>
+    </div>
+    <div class="ctl"><label>Sort by</label><select class="field" id="metricSortSel"></select></div>
+    <span class="spacer"></span>
+    <label class="toggle"><input type="checkbox" id="metricNeonOnly" /> Neon only</label>
+    <label class="toggle"><input type="checkbox" id="metricHeat" checked /> Color</label>
+  </div>
+
+  <!-- GUIDE -->
+  <div class="controls" id="ctl-guide" style="display:none">
+    <div class="ctl"><label>Reference</label></div>
+    <span class="meta" style="color:var(--text-faint);font-size:11.5px">Every column, profile gate, and provenance source behind the numbers.</span>
+  </div>
+
+  <main>
+    <section class="view active" id="view-compare"><div class="table-wrap"><table id="cmpTable"></table></div></section>
+    <section class="view" id="view-picks"><div class="matrix-wrap"><table class="matrix" id="picksTable"></table></div></section>
+    <section class="view" id="view-models"><div class="model-page" id="modelPage"></div></section>
+    <section class="view" id="view-metrics"><div class="matrix-wrap"><table class="matrix" id="metricsTable"></table></div></section>
+    <section class="view" id="view-guide"><div class="guide-grid" id="guideGrid"></div></section>
+  </main>
+
+  <footer class="bottom">
+    <span class="lg"><span class="sw" id="lgGood"></span> higher rank</span>
+    <span class="lg"><span class="sw" id="lgBad"></span> lower rank</span>
+    <span class="lg"><span class="badge neon" style="padding:0 6px">NEON</span> Neon.ai model</span>
+    <span class="lg"><span class="star">★</span> Pareto-optimal</span>
+    <span class="lg"><span class="badge dq" style="padding:0 6px">DQ</span> disqualified by a profile gate</span>
+    <span class="spacer"></span>
+    <span id="provFoot"></span>
+  </footer>
+</div>
+<div id="tip"></div>
+<script src="data.js"></script>
+<script src="app.js"></script>
+</body>
+</html>

scripts/build_data.py

@@ -0,0 +1,493 @@
+#!/usr/bin/env python3
+"""Build the static dashboard data bundle (``data.js``) for the Neon leaderboard.
+
+This reads the leaderboard's own source of truth -- ``run.json`` plus the
+profile YAMLs -- and recomputes picks, the Pareto front, disqualifications,
+verdicts, and per-metric win/loss tallies using logic faithful to
+``router/pareto.py`` and ``router/reports/leaderboard.py`` in the neon-router
+repo. The output is a single ``data.js`` file that assigns
+``window.LEADERBOARD_DATA`` so the dashboard loads with no server / no CORS.
+
+Usage:
+    python scripts/build_data.py \
+        --run-json /path/to/reports/dashboard/raw/run.json \
+        --profiles-dir /path/to/configs/profiles \
+        --out ../data.js
+
+Defaults point at a sibling ``neon-router`` checkout.
+"""
+from __future__ import annotations
+
+import argparse
+import json
+import math
+import os
+from dataclasses import dataclass, field
+from datetime import datetime, timezone
+from pathlib import Path
+from typing import Dict, List, Optional, Sequence, Tuple
+
+try:
+    import yaml
+except ImportError:  # pragma: no cover
+    raise SystemExit("PyYAML is required: pip install pyyaml")
+
+
+# ---------------------------------------------------------------------------
+# Profile model (mirror of router.pareto.ProfilePreference, subset we need)
+# ---------------------------------------------------------------------------
+@dataclass
+class Profile:
+    name: str
+    latency_budget_ms: Optional[float] = None
+    generation_latency_budget_ms: Optional[float] = None
+    cost_ceiling_per_query_usd: Optional[float] = None
+    accuracy_floor_mrr: Optional[float] = None
+    accuracy_floor_generation_acc: Optional[float] = None
+    corpus_profile: Optional[str] = None
+    query_mix: Dict[str, float] = field(default_factory=dict)
+    deployment_constraint: Optional[str] = None
+    quality_weight: float = 0.7
+    latency_penalty_per_ms: float = 0.0001
+    generation_latency_penalty_per_ms: float = 0.0
+    cost_penalty_per_token: float = 0.0
+    metric: str = "mrr"
+
+
+# Fields the dashboard reads off each aggregate row.
+NUMERIC_FIELDS = [
+    "mrr", "hit_at_1", "hit_at_5", "ndcg_at_10",
+    "p50_retrieval_latency_ms", "p95_retrieval_latency_ms",
+    "avg_retrieval_latency_ms", "index_build_seconds",
+    "est_cost_per_query_usd", "generation_acc",
+    "p50_generation_latency_ms", "avg_generation_latency_ms",
+    "generation_tokens_per_sec", "est_generation_cost_per_query_usd",
+    "paraphrase_drop", "shift_drop", "n_queries",
+]
+
+DQ_LATENCY, DQ_COST, DQ_ACCURACY = "latency", "cost", "accuracy"
+
+# Models served/fine-tuned by Neon.ai (vs. retrieval-only baselines and any
+# external references). Edit this set to control the "Neon" badge in the UI.
+# Match on the candidate's `model` name (the first ``::`` segment).
+NEON_MODELS = {
+    "devstral-123b",
+    "ministral-14b-reasoning",
+    "magistral-small",
+    "devstral-small-24b",
+    "mistral-small-3.2-24b",
+    "granite-4-30b",
+}
+
+
+def _get(agg: dict, key: str):
+    return agg.get(key)
+
+
+def _acc_axis(agg: dict, metric: str) -> float:
+    return float(agg.get(metric) or 0.0)
+
+
+def _cost_per_query(agg: dict) -> float:
+    return float(agg.get("est_cost_per_query_usd") or 0.0)
+
+
+# ---------------------------------------------------------------------------
+# Faithful port of router.pareto
+# ---------------------------------------------------------------------------
+def apply_hard_constraints(aggs: List[dict], p: Profile) -> Tuple[List[dict], List[Tuple[dict, str]]]:
+    survivors: List[dict] = []
+    dq: List[Tuple[dict, str]] = []
+    for agg in aggs:
+        reason: Optional[str] = None
+        avg_lat = agg.get("avg_retrieval_latency_ms")
+        avg_gen_lat = agg.get("avg_generation_latency_ms")
+        if p.latency_budget_ms is not None and avg_lat is not None and avg_lat > p.latency_budget_ms:
+            reason = DQ_LATENCY
+        elif (
+            p.generation_latency_budget_ms is not None
+            and avg_gen_lat is not None
+            and avg_gen_lat > p.generation_latency_budget_ms
+        ):
+            reason = DQ_LATENCY
+        elif (
+            p.cost_ceiling_per_query_usd is not None
+            and p.cost_ceiling_per_query_usd > 0.0
+            and _cost_per_query(agg) > p.cost_ceiling_per_query_usd
+        ):
+            reason = DQ_COST
+        elif (
+            p.accuracy_floor_generation_acc is not None
+            and agg.get("generation_acc") is not None
+            and agg.get("generation_acc") < p.accuracy_floor_generation_acc
+        ):
+            reason = DQ_ACCURACY
+        elif (
+            p.accuracy_floor_mrr is not None
+            and p.metric == "mrr"
+            and _acc_axis(agg, p.metric) < p.accuracy_floor_mrr
+        ):
+            reason = DQ_ACCURACY
+        if reason is None:
+            survivors.append(agg)
+        else:
+            dq.append((agg, reason))
+    return survivors, dq
+
+
+def pareto_front(aggs: List[dict], metric: str = "mrr") -> List[dict]:
+    front: List[dict] = []
+    for a in aggs:
+        dominated = False
+        for b in aggs:
+            if a is b:
+                continue
+            a_lat = a.get("avg_retrieval_latency_ms") or 0.0
+            b_lat = b.get("avg_retrieval_latency_ms") or 0.0
+            a_build = a.get("index_build_seconds") or 0.0
+            b_build = b.get("index_build_seconds") or 0.0
+            ge = (
+                _acc_axis(b, metric) >= _acc_axis(a, metric)
+                and b_lat <= a_lat
+                and b_build <= a_build
+            )
+            gt = (
+                _acc_axis(b, metric) > _acc_axis(a, metric)
+                or b_lat < a_lat
+                or b_build < a_build
+            )
+            if ge and gt:
+                dominated = True
+                break
+        if not dominated:
+            front.append(a)
+    return front
+
+
+def select_winner(aggs: List[dict], p: Profile) -> Optional[dict]:
+    survivors, _ = apply_hard_constraints(aggs, p)
+    front = pareto_front(survivors, metric=p.metric)
+    if not front:
+        return None
+
+    def utility(agg: dict) -> float:
+        accuracy = _acc_axis(agg, p.metric)
+        lat = agg.get("avg_retrieval_latency_ms") or 0.0
+        gen_lat = agg.get("avg_generation_latency_ms") or 0.0
+        return (
+            p.quality_weight * accuracy
+            - p.latency_penalty_per_ms * lat
+            - p.generation_latency_penalty_per_ms * gen_lat
+        )
+
+    return sorted(front, key=utility, reverse=True)[0]
+
+
+def _wilson_ci(n_succ: float, n_total: int, z: float = 1.96) -> Tuple[float, float]:
+    if n_total == 0:
+        return (0.0, 0.0)
+    pp = n_succ / n_total
+    denom = 1.0 + z * z / n_total
+    centre = (pp + z * z / (2.0 * n_total)) / denom
+    margin = z * math.sqrt(pp * (1.0 - pp) / n_total + z * z / (4.0 * n_total * n_total)) / denom
+    return (max(0.0, centre - margin), min(1.0, centre + margin))
+
+
+def value_add_verdict(aggs: List[dict], p: Profile) -> str:
+    survivors, _ = apply_hard_constraints(aggs, p)
+    if len(survivors) < 2:
+        return "no signal"
+
+    def key(a: dict) -> float:
+        if p.metric == "generation_acc" and a.get("generation_acc") is not None:
+            return a["generation_acc"]
+        return a.get("mrr") or 0.0
+
+    ranked = sorted(survivors, key=key, reverse=True)
+    winner, runner = ranked[0], ranked[1]
+    n = int(winner.get("n_queries") or 0)
+    w_lo, _ = _wilson_ci(key(winner) * n, n)
+    _, r_hi = _wilson_ci(key(runner) * n, n)
+    return "winner clear" if w_lo > r_hi else "inconclusive"
+
+
+# ---------------------------------------------------------------------------
+# Loaders
+# ---------------------------------------------------------------------------
+def load_profiles(profiles_dir: Path) -> List[Profile]:
+    valid = set(Profile.__dataclass_fields__)
+    profiles: List[Profile] = []
+    for path in sorted(profiles_dir.glob("*.yaml")):
+        raw = yaml.safe_load(path.read_text(encoding="utf-8")) or {}
+        clean = {k: v for k, v in raw.items() if k in valid}
+        profiles.append(Profile(**clean))
+    return profiles
+
+
+def split_candidate(cid: str) -> Tuple[str, str, str]:
+    parts = cid.split("::")
+    return (parts[0], parts[1], parts[2]) if len(parts) == 3 else (cid, "", "")
+
+
+# Mirror of leaderboard.py _GLOSSARY (kept in sync with the source report).
+GLOSSARY = {
+    "model": "The generation model in the pipeline. 'retrieval-only' means no LLM generation step was run for this row.",
+    "retriever": "Retrieval strategy used to fetch candidate documents (e.g. hybrid = BM25 + dense embeddings; bm25-only = lexical; splade-only = learned sparse).",
+    "rerank": "Cross-encoder reranker applied after retrieval ('none' = no reranking; minilm / mxbai are reranker models).",
+    "MRR": "Mean Reciprocal Rank \u2014 average of 1/rank for the first relevant doc retrieved per query. Range 0\u20131; higher is better.",
+    "H@1": "Hit@1 \u2014 fraction of queries where the top-1 retrieved doc is relevant. Higher is better.",
+    "H@5": "Hit@5 \u2014 fraction of queries where any of the top-5 retrieved docs is relevant. Higher is better.",
+    "nDCG": "Normalized Discounted Cumulative Gain @10 \u2014 position-weighted relevance of the top-10 results. Higher is better.",
+    "P50": "Median (50th percentile) per-query retrieval latency in milliseconds. Lower is better.",
+    "P95": "95th percentile retrieval latency \u2014 the tail (worst-case) cost per query, in milliseconds. Lower is better.",
+    "build": "Index build time in seconds for this benchmark corpus. Lower is better.",
+    "gen": "Generation score (task-dependent): for multiple-choice benchmarks, the fraction of questions answered correctly given retrieved context; for the language-following benchmark, the fraction of responses written in the user's language. Higher is better.",
+    "gen ms": "Median per-query generation latency (time-to-first-token + generation) in milliseconds. Lower is better.",
+    "tok/s": "LLM output tokens generated per second on this benchmark. Higher is better.",
+    "gen $": "Estimated generation cost per query in USD (token usage \u00d7 model price). Lower is better.",
+    "$/q": "Total estimated cost per query in USD (retrieval + generation). Lower is better.",
+    "par": "Paraphrase robustness \u2014 MRR on paraphrased queries minus MRR on originals. Near zero is expected; large negative means fragile exact-match retrieval.",
+    "shift": "Held-out (domain-shifted) slice MRR \u2014 measures out-of-distribution generalization on an unseen query subset.",
+    "n": "Number of queries evaluated for this row.",
+    "verdict": "Whether the winning candidate is statistically distinct from the runner-up. 'winner clear' = winner's Wilson CI lower bound exceeds runner-up's upper bound; 'inconclusive' = confidence intervals overlap (within noise); 'no signal' = too few valid candidates; 'all disqualified' = every candidate failed a hard gate.",
+    "Pareto": "Pareto front \u2014 a pipeline not dominated by any other on (accuracy, latency, build-cost) simultaneously. Marked with \u2605.",
+    "DQ": "Disqualified \u2014 violated one of the profile's hard gates: DQ:latency (p50 over budget), DQ:cost (cost per query over ceiling), or DQ:accuracy (MRR or generation score below the floor).",
+    "pick": "The router's single recommended pipeline (model / retriever / reranker) for this client profile on this benchmark.",
+    "profile": "A client requirements file (not a model) declaring latency / cost / accuracy hard gates plus utility weights used to rank surviving candidates.",
+}
+
+# Per-metric definitions used by the Compare and Metrics tabs.
+# key -> (label, field, higher_is_better, fmt, tooltip-key)
+METRICS = [
+    ("mrr", "MRR", "mrr", True, "f3", "MRR"),
+    ("h1", "H@1", "hit_at_1", True, "f2", "H@1"),
+    ("h5", "H@5", "hit_at_5", True, "f2", "H@5"),
+    ("ndcg", "nDCG", "ndcg_at_10", True, "f3", "nDCG"),
+    ("p50", "P50 ms", "p50_retrieval_latency_ms", False, "f0", "P50"),
+    ("p95", "P95 ms", "p95_retrieval_latency_ms", False, "f0", "P95"),
+    ("build", "Build s", "index_build_seconds", False, "f1", "build"),
+    ("gen", "gen", "generation_acc", True, "f3", "gen"),
+    ("gen_ms", "gen ms", "p50_generation_latency_ms", False, "f0", "gen ms"),
+    ("tok_s", "tok/s", "generation_tokens_per_sec", True, "f0", "tok/s"),
+    ("gen_cost", "gen $", "est_generation_cost_per_query_usd", False, "f5", "gen $"),
+    ("cost", "$/q", "est_cost_per_query_usd", False, "f5", "$/q"),
+    ("par", "par", "paraphrase_drop", True, "f3", "par"),
+    ("shift", "shift", "shift_drop", True, "f3", "shift"),
+]
+
+
+def build(run_json: Path, profiles_dir: Path) -> dict:
+    raw = json.loads(run_json.read_text(encoding="utf-8"))
+    profiles = load_profiles(profiles_dir)
+    benchmarks: List[str] = list(raw.get("benchmarks", []))
+    aggregates: List[dict] = raw.get("aggregates", [])
+    candidates_meta: List[dict] = raw.get("candidates", [])
+
+    # Candidate metadata map keyed by candidate_id.
+    cand_map: Dict[str, dict] = {}
+    for c in candidates_meta:
+        llm = c.get("llm", {}) or {}
+        retr = c.get("retriever", {}) or {}
+        rer = c.get("reranker", {}) or {}
+        cid = f"{llm.get('name')}::{retr.get('name')}::{rer.get('name')}"
+        cand_map[cid] = {
+            "id": cid,
+            "model": llm.get("name"),
+            "model_adapter": llm.get("adapter"),
+            "model_repo": llm.get("model"),
+            "retriever": retr.get("name"),
+            "retriever_adapter": retr.get("adapter"),
+            "reranker": rer.get("name"),
+            "reranker_model": rer.get("model"),
+            "is_neon": llm.get("name") in NEON_MODELS,
+            "is_retrieval_only": llm.get("name") == "retrieval-only",
+        }
+
+    # Trim aggregates to the fields the dashboard needs + identity.
+    def trim(agg: dict) -> dict:
+        model, retriever, reranker = split_candidate(agg["candidate_id"])
+        out = {
+            "candidate_id": agg["candidate_id"],
+            "benchmark": agg["benchmark"],
+            "model": model,
+            "retriever": retriever,
+            "reranker": reranker,
+            "is_neon": model in NEON_MODELS,
+            "error": agg.get("error"),
+        }
+        for f in NUMERIC_FIELDS:
+            out[f] = agg.get(f)
+        return out
+
+    trimmed = [trim(a) for a in aggregates]
+
+    by_bench: Dict[str, List[dict]] = {}
+    for a in trimmed:
+        by_bench.setdefault(a["benchmark"], []).append(a)
+
+    # ----- Picks (per benchmark x profile) -----
+    picks: Dict[str, Dict[str, dict]] = {}
+    for bench in benchmarks:
+        aggs = by_bench.get(bench, [])
+        picks[bench] = {}
+        for p in profiles:
+            winner = select_winner(aggs, p)
+            if winner is None:
+                survivors, _ = apply_hard_constraints(aggs, p)
+                status = "all disqualified" if not survivors else "no winner"
+                picks[bench][p.name] = {"pick": None, "verdict": status}
+                continue
+            verdict = value_add_verdict(aggs, p)
+            metric_val = (
+                winner.get("generation_acc")
+                if p.metric == "generation_acc"
+                else winner.get("hit_at_1")
+            )
+            picks[bench][p.name] = {
+                "pick": winner["candidate_id"],
+                "model": winner["model"],
+                "retriever": winner["retriever"],
+                "reranker": winner["reranker"],
+                "mrr": winner.get("mrr"),
+                "metric": p.metric,
+                "metric_val": metric_val,
+                "p50": winner.get("p50_retrieval_latency_ms"),
+                "verdict": verdict,
+            }
+
+    # ----- Per-benchmark ordering, Pareto, DQ -----
+    bench_tables: Dict[str, dict] = {}
+    for bench in benchmarks:
+        aggs = sorted(by_bench.get(bench, []), key=lambda a: a.get("mrr") or 0.0, reverse=True)
+        front_ids = {a["candidate_id"] for a in pareto_front(aggs)}
+        dq_by_id: Dict[str, str] = {}
+        for p in profiles:
+            _, dqs = apply_hard_constraints(aggs, p)
+            for agg, reason in dqs:
+                dq_by_id.setdefault(agg["candidate_id"], reason)
+        bench_tables[bench] = {
+            "order": [a["candidate_id"] for a in aggs],
+            "pareto": sorted(front_ids),
+            "dq": dq_by_id,
+        }
+
+    # ----- Per-metric win / loss tally across benchmarks -----
+    EPS = 1e-9
+    metric_stats: Dict[str, dict] = {}
+    for key, label, fieldname, higher, _fmt, _tip in METRICS:
+        per_cand: Dict[str, dict] = {}
+        for bench in benchmarks:
+            vals = [
+                (a["candidate_id"], a.get(fieldname))
+                for a in by_bench.get(bench, [])
+                if a.get(fieldname) is not None
+            ]
+            if not vals:
+                continue
+            best = (max if higher else min)(v for _, v in vals)
+            worst = (min if higher else max)(v for _, v in vals)
+            for cid, v in vals:
+                s = per_cand.setdefault(cid, {"wins": 0, "losses": 0, "sum": 0.0, "count": 0})
+                s["sum"] += v
+                s["count"] += 1
+                if abs(v - best) <= EPS:
+                    s["wins"] += 1
+                if abs(v - worst) <= EPS and abs(best - worst) > EPS:
+                    s["losses"] += 1
+        for cid, s in per_cand.items():
+            s["mean"] = s["sum"] / s["count"] if s["count"] else None
+            del s["sum"]
+        metric_stats[key] = per_cand
+
+    profiles_out = []
+    for p in profiles:
+        gates = []
+        if p.latency_budget_ms:
+            gates.append(f"p50 \u2264 {p.latency_budget_ms:g} ms")
+        if p.generation_latency_budget_ms:
+            gates.append(f"gen \u2264 {p.generation_latency_budget_ms:g} ms")
+        if p.cost_ceiling_per_query_usd:
+            gates.append(f"$/q \u2264 {p.cost_ceiling_per_query_usd:g}")
+        if p.accuracy_floor_mrr:
+            gates.append(f"MRR \u2265 {p.accuracy_floor_mrr:g}")
+        if p.accuracy_floor_generation_acc:
+            gates.append(f"gen \u2265 {p.accuracy_floor_generation_acc:g}")
+        profiles_out.append({
+            "name": p.name,
+            "metric": p.metric,
+            "gates": gates,
+            "corpus_profile": p.corpus_profile,
+            "query_mix": p.query_mix,
+            "deployment_constraint": p.deployment_constraint,
+            "quality_weight": p.quality_weight,
+        })
+
+    header = raw.get("provenance", {}) or {}
+    date_str = ""
+    merged_at = header.get("merged_at")
+    if merged_at:
+        date_str = merged_at[:10]
+
+    return {
+        "generated_at": datetime.now(timezone.utc).isoformat(),
+        "date": date_str,
+        "benchmarks": benchmarks,
+        "candidates": cand_map,
+        "aggregates": trimmed,
+        "profiles": profiles_out,
+        "picks": picks,
+        "bench_tables": bench_tables,
+        "metric_stats": metric_stats,
+        "metrics_def": [
+            {"key": k, "label": l, "field": f, "higher": h, "fmt": fmt, "tip": t}
+            for (k, l, f, h, fmt, t) in METRICS
+        ],
+        "glossary": GLOSSARY,
+        "provenance": header,
+        "n_candidates": len({a["candidate_id"] for a in trimmed}),
+        "n_queries": (trimmed[0].get("n_queries") if trimmed else None),
+    }
+
+
+def main() -> None:
+    here = Path(__file__).resolve().parent
+    default_repo = here.parent.parent / "neon-router"
+    ap = argparse.ArgumentParser()
+    ap.add_argument(
+        "--run-json",
+        type=Path,
+        default=default_repo / "reports" / "dashboard" / "raw" / "run.json",
+    )
+    ap.add_argument(
+        "--profiles-dir",
+        type=Path,
+        default=default_repo / "configs" / "profiles",
+    )
+    ap.add_argument("--out", type=Path, default=here.parent / "data.js")
+    args = ap.parse_args()
+
+    if not args.run_json.exists():
+        raise SystemExit(f"run.json not found: {args.run_json}")
+    if not args.profiles_dir.exists():
+        raise SystemExit(f"profiles dir not found: {args.profiles_dir}")
+
+    data = build(args.run_json, args.profiles_dir)
+    payload = json.dumps(data, ensure_ascii=False, separators=(",", ":"))
+    args.out.write_text(
+        "// Auto-generated by scripts/build_data.py - do not edit by hand.\n"
+        "window.LEADERBOARD_DATA = " + payload + ";\n",
+        encoding="utf-8",
+    )
+    # Also emit a plain JSON sidecar for non-browser consumers.
+    (args.out.parent / "data.json").write_text(
+        json.dumps(data, ensure_ascii=False, indent=2), encoding="utf-8"
+    )
+    print(f"Wrote {args.out} ({args.out.stat().st_size/1024:.1f} KB)")
+    print(f"  benchmarks={len(data['benchmarks'])} "
+          f"candidates={data['n_candidates']} "
+          f"aggregates={len(data['aggregates'])} "
+          f"profiles={len(data['profiles'])}")
+
+
+if __name__ == "__main__":
+    main()