One 'output not in trace' measure on the leaderboard; gaps table grouped; row page two-column layout

home.js

@@ -44,7 +44,6 @@
     const h = r.scan.holes, m = h.models, out = [];
     if (m.mismatch) out.push(["bad", "served " + shortModel(m.served[0].model)]);
     else if (m.foreignTrials) out.push(["warn", "fallback " + m.served.filter(s => !s.configured).map(s => shortModel(s.model).replace(/^claude-/, "")).join(", ") + " " + pct0(m.foreignStepShare)]);
-    if (h.text.uncounted) out.push(["bad", "output ≥" + pct0(h.text.uncounted) + " unbilled"]);
     const off = h.lbCost && Math.abs(h.costSum - h.lbCost) / h.lbCost > 0.02;
     if (off) out.push(["warn", "trial costs " + money(h.costSum) + "/" + money(h.lbCost)]);
     else if (h.uncosted / r.scan.trials > 0.05) out.push(["warn", h.uncosted + " uncosted"]);
@@ -57,17 +56,15 @@
   const leadsOnly = $("leads-only");
   const SORTS = {
     rank: r => r.rank, score: r => -r.lb.accuracy, leads: r => -r.scan.leads, hacks: r => -(r.lb.rewardHacks || 0),
-    trace: r => visibleShare(r.scan.holes) ?? 0, cost: r => -(r.lb.cost || 0),
+    trace: r => -(r.scan.holes.text.unseen ?? 0), cost: r => -(r.lb.cost || 0),
   };
   function board() {
     const cols = [
       { name: "89 tasks × 5 attempts, easiest first", cls: "l" },
       { name: "Score", sort: "score", help: "leaderboard accuracy ± 95% CI half-width" },
-      { name: "Rewarded", help: "rewarded trials of the row's listed trials" },
       { name: "Leads", sort: "leads", help: "rewarded trials with an unexcused high or critical finding" },
       { name: "LB hacks", sort: "hacks", help: "leaderboard's own reward-hack rate" },
-      { name: "In trace", sort: "trace", help: "share of output tokens with text in the trace" },
-      { name: "Reasoning", help: "how much reasoning the trajectories hold" },
+      { name: "Output not in trace", sort: "trace", cls: "l", help: "share of reported output tokens with no text in the trace, and why: the reasoning's form" },
       { name: "Cost", sort: "cost", help: "leaderboard total" },
       { name: "Notes", cls: "l" },
     ];
@@ -78,18 +75,19 @@
       const h = r.scan.holes, tr = el("tr");
       tr.append(rowHead(r));
       const s = el("td", "l"); s.append(strip(r, { leadsOnly: leadsOnly.checked })); tr.append(s);
-      const sc = el("td"); sc.append(num(fmt1(r.lb.accuracy), "n big"), el("span", "pm", "±" + fmt1(r.lb.ci95))); tr.append(sc);
-      tr.append(el("td", "n", r.scan.rewarded + "/" + r.scan.trials));
+      const sc = el("td"); sc.append(num(fmt1(r.lb.accuracy), "n big"), el("span", "pm", "±" + fmt1(r.lb.ci95)));
+      sc.title = r.scan.rewarded + " of " + r.scan.trials + " listed trials rewarded"; tr.append(sc);
       const ld = el("td");
       if (r.scan.leads) ld.append(link(rowPath(r) + "highlights/index.html", String(r.scan.leads), "n big hot"));
       else ld.append(el("span", "z", "0"));
       tr.append(ld);
       tr.append(el("td", "n" + (r.lb.rewardHacks >= 1 ? " hot" : ""), fmt1(r.lb.rewardHacks) + "%"));
-      const vis = visibleShare(h), tc = el("td", "vis");
-      const bar = el("span", "bar"); const fill = el("i"); fill.style.width = Math.round(100 * (vis ?? 0)) + "%"; bar.append(fill);
-      tc.append(bar, el("span", "n", pct0(vis)));
+      const miss = h.text.unseen ?? 0, tc = el("td", "vis l");
+      const bar = el("span", "bar bar--miss"); const fill = el("i"); fill.style.width = Math.round(100 * miss) + "%"; bar.append(fill);
+      const rz = el("span", "rz rz--" + h.text.form, RZ[h.text.form]); rz.title = RZ_HELP[h.text.form];
+      tc.append(bar, el("span", "n pc", pct0(miss)), rz);
+      if (h.text.uncounted) tc.append(el("span", "bad unb", " · ≥" + pct0(h.text.uncounted) + " unbilled"));
       tr.append(tc);
-      const rz = el("td", "rz rz--" + h.text.form, RZ[h.text.form]); rz.title = RZ_HELP[h.text.form]; tr.append(rz);
       tr.append(el("td", "n", money(r.lb.cost)));
       const nt = el("td", "l notes");
       flags(r).forEach(([lvl, text]) => nt.append(el("span", "fl fl--" + lvl, text)));
@@ -106,31 +104,38 @@
   }
   leadsOnly.addEventListener("change", board);
 
-  /* 2. Gaps & accounting. */
+  /* 2. Gaps & accounting: grouped columns, each analysis once. */
   function gaps() {
-    const cols = [
-      { name: "Models served", cls: "l", help: "by share of steps; red: not the configured model" },
-      { name: "Reasoning", help: "form of the reasoning in the trajectories" },
-      { name: "Text / tok", help: "recorded characters per output token (≈3.2 = all output in the trace)" },
-      { name: "In trace", help: "share of output tokens with text in the trace" },
-      { name: "Reas. tok", help: "reasoning tokens as a share of output, where reported" },
-      { name: "Low-text trials", help: "trials whose text covers under 80% of their output" },
-      { name: "Web results missing", help: "trials with web activity whose results aren't in the trace / trials with web activity" },
-      { name: "Web prov.", help: "web input or reference provenance unresolved" },
-      { name: "Unpaired results", help: "tool results not paired to their calls" },
-      { name: "Calls missing", help: "token totals exceed the recorded model calls" },
-      { name: "Ends on call", help: "trace ends on a tool call with no result" },
-      { name: "Compacted", help: "history compacted into a summary" },
-      { name: "No traj.", help: "listed trial without a trajectory" },
-      { name: "Time­stamps", help: "missing or out of order" },
-      { name: "Trial cost / LB", help: "sum of per-trial recorded cost against the leaderboard total" },
-      { name: "Uncosted", help: "trials without a recorded cost" },
-      { name: "Output unbilled", help: "recorded reasoning not counted in output tokens, as a share of reported output" },
-      { name: "Undecided", help: "trials with at least one behavioural check atif-scan couldn't decide" },
+    const groups = [
+      ["Models", [["Served, by share of steps", "l", "red: not the configured model; the count opens the trials"]]],
+      ["Output vs recorded text", [
+        ["Chars / output token", "", "recorded agent text (reasoning, messages, tool arguments) per reported output token; ≈3.2 when the trace holds all output"],
+        ["Reasoning tokens", "", "reasoning tokens as a share of output, where the harness reports them separately"],
+        ["Trials < 80% text", "", "trials whose recorded text covers under 80% of their output tokens"],
+        ["Unbilled", "", "recorded reasoning on top of output tokens the visible text already fills: missing from the reported output and cost"]]],
+      ["Trace gaps · trials", [
+        ["Web results", "", "trials with web activity whose results aren't in the trace / trials with web activity"],
+        ["Web prov.", "", "web input or reference provenance unresolved"], ["Unpaired results", "", "tool results not paired to their calls"],
+        ["Calls missing", "", "token totals exceed the recorded model calls"], ["Ends on call", "", "trace ends on a tool call with no result"],
+        ["Compacted", "", "history compacted into a summary"], ["No traj.", "", "listed trial without a trajectory"],
+        ["Time­stamps", "", "missing or out of order"]]],
+      ["Cost", [["Trials / LB", "", "sum of per-trial recorded cost against the leaderboard total"], ["Uncosted", "", "trials without a recorded cost"]]],
+      ["Evidence", [["Undecided", "", "trials with a behavioural check atif-scan couldn't decide"]]],
     ];
+    const g = el("tr", "groups"), hr = el("tr");
+    const g0 = el("th", "rowh", ""); g0.rowSpan = 1; g.append(g0);
+    hr.append(el("th", "rowh", "Row"));
+    groups.forEach(([name, cols], i) => {
+      const th = el("th", "grp" + (i ? " gstart" : ""), name); th.colSpan = cols.length; g.append(th);
+      cols.forEach(([cname, cls, help], j) => {
+        const c = el("th", (cls || "") + (j === 0 && i ? " gstart" : ""), cname); c.title = help; hr.append(c);
+      });
+    });
+    const thead = el("thead"); thead.append(g, hr);
     const t = $("gaps-table");
-    t.replaceChildren(headRow(cols, "Row"));
+    t.replaceChildren(thead);
     const tb = el("tbody");
+    const start = td => { td.classList.add("gstart"); return td; };
     D.rows.forEach(r => {
       const h = r.scan.holes, n = r.scan.trials, m = h.models, tr = el("tr");
       tr.append(rowHead(r));
@@ -138,24 +143,26 @@
       m.served.filter(x => x.share >= 0.005).slice(0, 3).forEach(x =>
         md.append(el("span", x.configured ? "mdl" : "mdl off", shortModel(x.model) + (m.served.length > 1 ? " " + pct0(x.share) : ""))));
       if (m.foreignTrials) md.append(link(holeHref(r, "foreign"), m.foreignTrials + " trials", "mlink"));
-      tr.append(md);
-      const rz = el("td", "rz rz--" + h.text.form, RZ[h.text.form]); rz.title = RZ_HELP[h.text.form]; tr.append(rz);
-      tr.append(el("td", "n", String(h.text.allCpt)));
-      const vis = visibleShare(h), vc = el("td", "n", pct0(vis)); heat(vc, 1 - (vis ?? 1), 1, "cool"); tr.append(vc);
+      tr.append(start(md));
+
+      tr.append(start(el("td", "n", String(h.text.allCpt))));
       tr.append(el("td", "n", h.text.reasoningShare !== null ? pct0(h.text.reasoningShare) : "–"));
-      tr.append(countCell(r, "unseen", h.unseen, n, { scale: 1, hue: "cool" }));
-      tr.append(h.webTrials ? countCell(r, "webUnrecorded", h.webUnrecorded, n, { of: h.webTrials, scale: 1 }) : el("td", "c z", "no web"));
+      tr.append(countCell(r, "unseen", h.unseen, n, { scale: 1 }));
+      const ub = el("td", "c");
+      if (h.text.uncounted) ub.append(link(holeHref(r, "uncounted"), "≥" + pct0(h.text.uncounted), "bad"));
+      else ub.append(el("span", "z", "·"));
+      tr.append(ub);
+
+      tr.append(start(h.webTrials ? countCell(r, "webUnrecorded", h.webUnrecorded, n, { of: h.webTrials, scale: 1 }) : el("td", "c z", "no web")));
       ["webUnresolved", "unpaired", "callsMissing", "endsOnCall", "compacted", "noTrajectory", "timestamps"]
         .forEach(k => tr.append(countCell(r, k, h[k], n)));
+
       const off = h.lbCost ? h.costSum - h.lbCost : 0;
-      const cost = el("td", "n" + (Math.abs(off) / (h.lbCost || 1) > 0.02 ? " bad" : ""),
-        Math.abs(off) >= 0.5 ? money(h.costSum) + " / " + money(h.lbCost) : money(h.costSum) + " ="); tr.append(cost);
+      tr.append(start(el("td", "n" + (Math.abs(off) / (h.lbCost || 1) > 0.02 ? " bad" : ""),
+        Math.abs(off) >= 0.5 ? money(h.costSum) + " / " + money(h.lbCost) : money(h.costSum) + " =")));
       tr.append(countCell(r, "uncosted", h.uncosted, n, { scale: 0.3 }));
-      const ub = el("td", "c");
-      if (h.text.uncounted) { const a = link(holeHref(r, "uncounted"), "≥" + pct0(h.text.uncounted), "bad"); ub.append(a); }
-      else ub.append(el("span", "z", "·"));
-      tr.append(ub);
-      const und = el("td", "n", String(h.evidenceIncomplete)); heat(und, h.evidenceIncomplete / n, 1.5, "cool"); tr.append(und);
+      const und = el("td", "n", String(h.evidenceIncomplete)); heat(und, h.evidenceIncomplete / n, 1.5, "cool");
+      tr.append(start(und));
       tb.append(tr);
     });
     t.append(tb);

index.html

@@ -48,12 +48,19 @@
       <p class="key"><i class="sq sq--1"></i>rewarded <i class="sq sq--0"></i>failed <i class="sq sq--t"></i>timed out
         <i class="sq sq--e"></i>errored · corner: highest finding <i class="sq sq--1 sq--c"></i>critical
         <i class="sq sq--1 sq--h"></i>high <i class="sq sq--1 sq--m"></i>medium · <b>Leads</b>: rewarded trials whose
-        highest unexcused finding is high or critical · <b>In trace</b>: share of output tokens with text in the trace.</p>
+        highest unexcused finding is high or critical.
+        <b>Output not in trace</b>: share of reported output tokens with no text in the trace, and why —
+        <span class="rz rz--hidden">hidden</span> encrypted, tokens reported · <span class="rz rz--unrecorded">unrecorded</span> no text
+        or count · <span class="rz rz--dropped">dropped</span> lost converting the session · <span class="rz rz--summary">summary</span> ·
+        <span class="rz rz--full">full text</span>.</p>
     </section>
 
     <section id="gaps" class="block">
       <header class="block-head"><h2>Gaps &amp; accounting</h2>
-        <span class="muted small">trials per row unless marked · shading is the share · click a count for its trials</span></header>
+        <span class="muted small">trials per row unless marked · shading is the share · click a figure for its trials</span></header>
+      <p class="explain-line">The leaderboard's <b>output not in trace</b> comes from this group: recorded text per reported
+        output token, ≈3.2 when the trace holds everything. Where reasoning tokens are reported, their share is shown; trials under 80%
+        open from their count. <b>Unbilled</b> is the reverse gap: recorded reasoning missing from the output tokens, and so the cost.</p>
       <div class="scroll"><table id="gaps-table" class="grid grid--gaps"></table></div>
     </section>
 

row.js

@@ -69,7 +69,7 @@
   stats.append(
     stat("Score", fmt1(r.lb.accuracy) + "%", "± " + fmt1(r.lb.ci95) + " · " + r.scan.rewarded + " of " + n + " trials rewarded"),
     stat("Review leads", String(r.scan.leads), "leaderboard reward hacks " + fmt1(r.lb.rewardHacks) + "%", r.scan.leads ? "stat--hot" : ""),
-    stat("Trace visibility", pct0(visibleShare(h)), "of output tokens have text in the trace"),
+    stat("Output not in trace", pct0(h.text.unseen), "reasoning " + T.RZ[h.text.form] + (h.text.uncounted ? " · ≥" + pct0(h.text.uncounted) + " unbilled" : "")),
     stat("Cost", money(r.lb.cost), "leaderboard · trials sum to " + money(h.costSum)));
   const actions = el("div", "actions");
   actions.append(link(rowPath(r) + "highlights/index.html", "Highlights", "btn btn--primary"),
@@ -150,8 +150,12 @@
   if (h.uncosted) cost.append(trials("uncosted", "Trials without a recorded cost"));
   if (!h.text.uncounted && !h.uncosted && Math.abs(gap) < 0.5) cost.append(el("p", "muted small", "Every trial costed; the trials add up to the leaderboard's total."));
 
+  // Two columns that fill independently: findings on the left; what ran, the trace and the cost on the right.
   const grid = el("div", "panels");
-  grid.append(found, ran, trace, cost);
+  const left = el("div", "pcol"), right = el("div", "pcol");
+  left.append(found);
+  right.append(ran, trace, cost);
+  grid.append(left, right);
   main.replaceChildren(head, stats, actions, stripWrap, grid);
 
   // #k=<gap> (from the home page's tables) opens that list of trials; #found etc. scroll to a panel.

style.css

@@ -72,7 +72,8 @@ a.sq:hover { outline: 2px solid var(--ink); }
 .btn:hover { background: var(--inset); color: var(--ink); }
 .strip-wrap { padding: 14px 18px; border: 1px solid var(--line-soft); border-radius: var(--radius); background: var(--card); overflow-x: auto; }
 .strip-wrap p { margin: 0 0 8px; }
-.panels { display: grid; grid-template-columns: repeat(auto-fit, minmax(440px, 1fr)); gap: 22px; margin-top: 22px; align-items: start; }
+.panels { display: grid; grid-template-columns: minmax(0, 1fr) minmax(0, 1fr); gap: 22px; margin-top: 22px; align-items: start; }
+.pcol { display: flex; flex-direction: column; gap: 22px; min-width: 0; }
 .panel { padding: 18px 20px; border: 1px solid var(--line-soft); border-radius: var(--radius); background: var(--card); scroll-margin-top: 70px; }
 .panel h2 { font-size: 24px; margin: 0 0 4px; }
 .checks, .models, .trial-list { list-style: none; padding: 0; margin: 0; }
@@ -192,3 +193,18 @@ details.trials summary .label { font-weight: 650; }
 .grid--fam td { min-width: 96px; }
 .grid th, .grid td { padding: 5px 6px; }
 .grid .notes { max-width: 220px; }
+
+/* Grouped headers. */
+.grid thead tr.groups th { position: static; border-bottom: 0; padding-bottom: 0; font-size: 10px; color: var(--ink); text-align: left; }
+.grid .gstart { border-left: 1.5px solid var(--line-soft); }
+.grid thead tr.groups th.gstart { border-left: 1.5px solid var(--line); }
+.explain-line { margin: 0 0 10px; font-size: 13.5px; color: var(--muted); max-width: 1100px; }
+.explain-line b { color: var(--ink); }
+.vis .pc { display: inline-block; min-width: 36px; margin-right: 6px; }
+.grid td .rz { display: block; font-size: 10.5px; }
+.vis .rz { display: inline; font-size: 12px; }
+.grid td.vis .rz { display: inline; font-size: 12px; }
+.bar--miss { background: color-mix(in srgb, var(--ink) 10%, transparent); }
+.bar--miss i { background: var(--hl-h); }
+.unb { font-size: 12px; }
+.vis .bar { width: 40px; }