Generate one rank per evaluation cohort

README.md

@@ -181,13 +181,15 @@ SHA-256 so consumers can reject a partially updated artifact pair.
 Every aggregate is separated by task/config revision, model revision, runner
 revision, container digest, hardware, resolution, appliance, target-data access,
 smoke/full scope, and a digest of every protocol override. Context length,
-epochs, and sample limits remain visible in the generated table.
-The website's rank groups use a separate comparison-protocol digest covering
-effective context length and epochs, tuning method/trial budget/selection
-protocol, and the immutable runner, contrib, container, and hardware tuple. It
-intentionally excludes model-specific selected parameters and the study digest,
-so different models remain comparable and evaluation seeds from one frozen
-study can still aggregate. A corrected
+epochs, sample limits, tuning records, and runtime provenance remain visible in
+the generated artifacts.
+
+Public rank is generated once, in Python, within an evaluation cohort defined by
+task/config revision, profile, resolution, appliance, sample limit, target-data
+access, and smoke/full scope. Architecture choices such as context length and
+epoch count, and provenance changes such as a later container build, do not
+reset rank. The complete comparison-protocol digest still records those details
+for audit and exact reproduction. A corrected
 full run becomes `full-verified`, and a smoke run becomes `smoke-verified`, only
 after the required seeds 10, 20, and 42 pass source, container, and dataset
 provenance checks. Incomplete clean smoke matrices are labelled `smoke-partial`.

index.html

@@ -925,7 +925,7 @@ <h2>Built to become NILM's ImageNet</h2>
     <div class="lb-sub" id="live-leaderboard-summary">Loading immutable result bundles&hellip;</div>
     <div class="table-container">
         <table class="lb">
-            <thead><tr><th>Group rank</th><th>Model</th><th>Task / appliance</th><th>MAE&darr; (W)</th><th>F1&uarr;</th><th>Params</th><th>Time&darr;</th><th>Peak VRAM&darr;</th><th>Status</th></tr></thead>
+            <thead><tr><th>Rank</th><th>Model</th><th>Task / appliance</th><th>MAE&darr; (W)</th><th>F1&uarr;</th><th>Params</th><th>Time&darr;</th><th>Peak VRAM&darr;</th><th>Status</th></tr></thead>
             <tbody id="live-leaderboard-body">
                 <tr><td colspan="9" style="color:var(--text-muted)">Loading&hellip;</td></tr>
             </tbody>
@@ -1558,12 +1558,6 @@ <h2>Cite NILMBench2026</h2>
 function memoryInMiB(value) {
   return value == null ? '&mdash;' : `${(Number(value) / 1048576).toFixed(1)} MiB`;
 }
-function comparisonKey(entry) {
-  return [
-    entry.comparison_protocol_sha256,
-    entry.appliance,
-  ].join('|');
-}
 async function loadLiveLeaderboard() {
   const body = document.getElementById('live-leaderboard-body');
   const summary = document.getElementById('live-leaderboard-summary');
@@ -1578,9 +1572,10 @@ <h2>Cite NILMBench2026</h2>
       return;
     }
     const statusOrder = {'full-verified': 0, 'smoke-verified': 1, 'smoke-partial': 2, 'candidate': 3, 'smoke-unverified': 4};
-    entries.sort((a,b)=>comparisonKey(a).localeCompare(comparisonKey(b)) ||
-      (statusOrder[a.status]??9)-(statusOrder[b.status]??9) || a.mae_mean-b.mae_mean);
-    const ranks = new Map();
+    entries.sort((a,b)=>a.ranking_protocol_sha256.localeCompare(b.ranking_protocol_sha256) ||
+      (statusOrder[a.status]??9)-(statusOrder[b.status]??9) ||
+      (a.rank??Number.MAX_SAFE_INTEGER)-(b.rank??Number.MAX_SAFE_INTEGER) ||
+      a.mae_mean-b.mae_mean);
     body.innerHTML = entries.map((entry) => {
       const protocol = [
         entry.sequence_length ? `seq ${entry.sequence_length}` : null,
@@ -1589,10 +1584,7 @@ <h2>Cite NILMBench2026</h2>
       ].filter(Boolean).join(' / ');
       const task = `${entry.task} / ${entry.appliance} / ${entry.sample_period}s` +
         (protocol ? ` / ${protocol}` : '');
-      const key = comparisonKey(entry);
-      const isVerified = entry.status === 'full-verified' || entry.status === 'smoke-verified';
-      const rank = isVerified ? (ranks.get(key) ?? 0) + 1 : null;
-      if (rank != null) ranks.set(key, rank);
+      const rank = Number.isInteger(entry.rank) ? entry.rank : null;
       const cls = rank === 1 ? ' class="gold"' : '';
       return `<tr${cls}><td class="rank">${rank ?? '&mdash;'}</td><td>${escHTML(entry.model)}</td>` +
         `<td class="fam">${escHTML(task)}</td>` +