Report the scheme-only floor beneath the sentence results

Sphragis merges seven treebanks that do not overlap in authors, so the
annotation scheme partly predicts the author. A collapsed note under the
sentence table gives what classifiers score when they see annotation markup
and no Greek at all: 34.02 macro-F1 on one sentence and 65.98 on ten, against
a 1.50 majority baseline. A model that does not clearly beat those has
demonstrated source identification rather than authorship attribution.

Figures come from scripts/audit_leakage.py in the dataset repository and are
regenerated with the data.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

src/app.css

@@ -119,6 +119,20 @@ tbody tr.selected td.classifier-column { overflow: visible; white-space: normal;
 .tooltip { position: fixed; z-index: 50; top: var(--tip-y); left: var(--tip-x); display: grid; gap: 5px; width: 250px; padding: 11px 12px; border: 1px solid var(--border-strong); border-radius: 6px; background: var(--text); color: var(--page); box-shadow: 0 12px 30px rgb(0 0 0 / 18%); font-size: 11px; font-weight: 400; line-height: 1.45; white-space: normal; transform: translateX(-50%); }
 .tooltip strong { font-size: 11px; }
 .table-caption { display: flex; align-items: center; gap: 6px; margin: 11px 2px 0; color: var(--text-subtle); font-size: 11px; }
+.scheme-note { margin: 7px 2px 0; color: var(--text-subtle); font-size: 11px; }
+.scheme-note > summary { display: flex; align-items: center; gap: 6px; cursor: pointer; list-style: none; }
+.scheme-note > summary::-webkit-details-marker { display: none; }
+.scheme-note > summary::after { content: '\203A'; margin-left: 2px; transition: transform 0.12s; }
+.scheme-note[open] > summary::after { transform: rotate(90deg); }
+.scheme-note > summary:hover { color: var(--text-muted); }
+.scheme-note > summary:focus-visible { outline: 2px solid var(--text); outline-offset: 2px; border-radius: 3px; }
+.scheme-note > p { margin: 8px 0 0; max-width: 78ch; line-height: 1.55; }
+.scheme-note .scheme-source { color: var(--text-subtle); }
+.scheme-note code { font-size: 10px; }
+.scheme-table { margin: 8px 0 0; border-collapse: collapse; font-variant-numeric: tabular-nums; }
+.scheme-table th, .scheme-table td { padding: 3px 12px 3px 0; text-align: right; font-weight: 400; }
+.scheme-table thead th { color: var(--text-subtle); font-size: 10px; text-transform: uppercase; letter-spacing: 0.05em; }
+.scheme-table tbody th { text-align: left; color: var(--text-muted); }
 .accuracy-chart { margin-top: 28px; padding: 22px; border: 1px solid var(--border); border-radius: 8px; background: var(--surface); box-shadow: var(--shadow-sm); }
 .chart-heading { display: flex; align-items: end; justify-content: space-between; gap: 24px; margin-bottom: 10px; }
 .chart-heading h3 { margin: 0 0 5px; font-size: 20px; font-weight: 600; letter-spacing: -0.025em; }

src/lib/components/LeaderboardTable.svelte

@@ -1,5 +1,5 @@
 <script lang="ts">
-  import { displayScore, metreTaskColumns, sentenceTaskColumns, type ResultRow, type TaskKey } from '$lib/data';
+  import { displayScore, metreTaskColumns, schemeBaselineNote, schemeOnlyBaselines, sentenceTaskColumns, type ResultRow, type TaskKey } from '$lib/data';
   import AuthorAccuracyChart from './AuthorAccuracyChart.svelte';
   import ConfusionMatrices from './ConfusionMatrices.svelte';
   import InfoTip from './InfoTip.svelte';
@@ -81,6 +81,21 @@
     </table>
   </div>
   <p class="table-caption"><InfoTip title="Scores" text="Values are held-out test macro-F1 percentages. Select a model row to inspect accuracy for each author." /> Scores are test macro-F1 (%). Select a row for per-author accuracy.</p>
+  {#if unit === 'sentence' && schemeOnlyBaselines.length}
+    <details class="scheme-note">
+      <summary><InfoTip title="Scheme-only floor" text={schemeBaselineNote} /> Scheme-only floor: {schemeOnlyBaselines.map((baseline) => `${displayScore(baseline.scores.sentence1)} / ${displayScore(baseline.scores.sentence10)}`).join(', ')} on 1 / 10 sentences</summary>
+      <p>{schemeBaselineNote} These classifiers see annotation markup and no Greek at all, so a model that does not clearly beat them has demonstrated source identification rather than authorship attribution.</p>
+      <table class="scheme-table">
+        <thead><tr><th>Probe</th>{#each columns as column}<th>{column.shortLabel}</th>{/each}</tr></thead>
+        <tbody>
+          {#each schemeOnlyBaselines as baseline}
+            <tr><th scope="row">{baseline.probe}</th>{#each columns as column}<td>{displayScore(baseline.scores[column.key])}</td>{/each}</tr>
+          {/each}
+        </tbody>
+      </table>
+      <p class="scheme-source">Regenerate with <code>scripts/audit_leakage.py</code>; full report in <code>metadata/leakage_report.json</code>. The <code>source_held_out_*</code> configurations evaluate each author on a treebank absent from its training data.</p>
+    </details>
+  {/if}
   {#if selectedRow}
     <AuthorAccuracyChart row={selectedRow} {columns} onselectauthor={(author) => selectedAuthor = author} />
     <ConfusionMatrices row={selectedRow} {columns} {selectedAuthor} />

src/lib/data.ts

@@ -802,6 +802,32 @@ export const metreResults: ResultRow[] = [
   }
 ];
 
+/**
+ * Scheme-only baselines from `scripts/audit_leakage.py`.
+ *
+ * Sphragis merges seven annotation projects that do not overlap in authors, so
+ * the annotation scheme is itself a partial answer key. These classifiers see
+ * annotation markup and no Greek at all: `inventory` is binary presence of each
+ * tag, feature and relation; `markup` adds non-lexical rates. A model that does
+ * not clearly beat them has demonstrated source identification rather than
+ * authorship attribution.
+ *
+ * Regenerate with:
+ *   python scripts/audit_leakage.py --data data --metadata metadata
+ */
+export interface SchemeBaseline {
+  probe: string;
+  scores: Partial<Record<TaskKey, number>>;
+}
+
+export const schemeOnlyBaselines: SchemeBaseline[] = [
+  { probe: 'Tag inventory only', scores: { sentence1: 17.95, sentence10: 16.65, sentence50: 28.92, sentence100: 27.27 } },
+  { probe: 'Non-lexical markup', scores: { sentence1: 34.02, sentence10: 65.98, sentence50: 48.91, sentence100: 41.67 } }
+];
+
+export const schemeBaselineNote =
+  'Sphragis merges seven treebanks that do not overlap in authors, so the annotation scheme partly predicts the author.';
+
 export function displayScore(score: number | undefined): string {
   return score === undefined ? '—' : score.toFixed(2);
 }