Upload app.py with huggingface_hub

app.py

@@ -874,21 +874,24 @@ VALIDATION = f"""
     </div>
   </div>
 
-  <div style="margin:18px 0 8px">
-    <div style="color:#f4f4f5;font-size:0.95rem;font-weight:600;margin-bottom:4px">Demo cohort · 5 held-out sites, 10 subjects</div>
-    <div style="color:#8b95a7;font-size:0.82rem;line-height:1.6">Two subjects per site were randomly selected from the ABIDE I held-out pool. The model that trained on <em>all other sites</em> was used — it has never seen any data from these institutions.</div>
+  <div style="border-left:3px solid #22c55e;background:#0f1f14;border-radius:0 8px 8px 0;padding:14px 18px;margin:24px 0 14px">
+    <div style="color:#22c55e;font-size:0.7rem;text-transform:uppercase;letter-spacing:1.5px;font-weight:600;margin-bottom:5px">01 · Demo cohort</div>
+    <div style="color:#f4f4f5;font-size:1rem;font-weight:600;margin-bottom:6px">5 held-out sites · 10 subjects · never seen during training</div>
+    <div style="color:#8b95a7;font-size:0.84rem;line-height:1.65">Two subjects per site were randomly drawn from the ABIDE I pool. The fold model was trained on <em>all other 19 sites</em> — these institutions were completely blind to it. Green = definitive correct, orange = correctly flagged inconclusive.</div>
   </div>
-  <img src="data:image/png;base64,{VAL_B64}" style="width:100%;border-radius:6px;margin-bottom:22px;border:1px solid #252a35"/>
+  <img src="data:image/png;base64,{VAL_B64}" style="width:100%;border-radius:6px;margin-bottom:32px;border:1px solid #252a35"/>
 
-  <div style="margin:0 0 8px">
-    <div style="color:#f4f4f5;font-size:0.95rem;font-weight:600;margin-bottom:4px">Full 20-fold LOSO AUC · all 1,102 subjects</div>
-    <div style="color:#8b95a7;font-size:0.82rem;line-height:1.6">Each bar is one site blind to training. AUC 0.7298 is the mean across all 20 folds — every subject is test-only exactly once. Prior art uses same-site splits; this is a strictly harder evaluation.</div>
+  <div style="border-left:3px solid #ef4444;background:#1a0f0f;border-radius:0 8px 8px 0;padding:14px 18px;margin:0 0 14px">
+    <div style="color:#ef4444;font-size:0.7rem;text-transform:uppercase;letter-spacing:1.5px;font-weight:600;margin-bottom:5px">02 · Full 20-fold LOSO</div>
+    <div style="color:#f4f4f5;font-size:1rem;font-weight:600;margin-bottom:6px">1,102 subjects · every subject is test-only exactly once</div>
+    <div style="color:#8b95a7;font-size:0.84rem;line-height:1.65">Each bar is one site's model evaluated on held-out subjects from that site. Mean AUC <span style="color:#ef4444;font-weight:600">0.7298</span> across all 20 folds. All published baselines use same-site train/test — cross-site is a fundamentally harder bar.</div>
   </div>
-  <img src="data:image/png;base64,{AUC_B64}" style="width:100%;border-radius:6px;margin-bottom:22px;border:1px solid #252a35"/>
+  <img src="data:image/png;base64,{AUC_B64}" style="width:100%;border-radius:6px;margin-bottom:32px;border:1px solid #252a35"/>
 
-  <div style="margin:0 0 8px">
-    <div style="color:#f4f4f5;font-size:0.95rem;font-weight:600;margin-bottom:4px">Subject-level predictions · demo cohort</div>
-    <div style="color:#8b95a7;font-size:0.82rem;line-height:1.6">Predictions with 0.4 &lt; p &lt; 0.6 are flagged as inconclusive rather than forced to a label — surfacing borderline cases for clinical review instead of producing a confident wrong answer.</div>
+  <div style="border-left:3px solid #f59e0b;background:#181208;border-radius:0 8px 8px 0;padding:14px 18px;margin:0 0 14px">
+    <div style="color:#f59e0b;font-size:0.7rem;text-transform:uppercase;letter-spacing:1.5px;font-weight:600;margin-bottom:5px">03 · Subject-level predictions</div>
+    <div style="color:#f4f4f5;font-size:1rem;font-weight:600;margin-bottom:6px">Inconclusive threshold · clinical safety over forced errors</div>
+    <div style="color:#8b95a7;font-size:0.84rem;line-height:1.65">When 0.4 &lt; p &lt; 0.6 the model flags the case for human review rather than committing to a label. On this cohort: <span style="color:#f59e0b;font-weight:600">2 correctly deferred</span>, zero confident misclassifications.</div>
   </div>
   <div style="background:#161922;border:1px solid #252a35;border-radius:8px;overflow:hidden">
     <table style="width:100%;border-collapse:collapse;font-size:0.86rem">
@@ -908,9 +911,10 @@ VALIDATION = f"""
     <span style="color:#cbd5e1">Zero confident misclassifications across 5 unseen sites.</span>
   </div>
 
-  <div style="margin:22px 0 8px">
-    <div style="color:#f4f4f5;font-size:0.95rem;font-weight:600;margin-bottom:4px">Diagnostic breakdown &amp; published baselines</div>
-    <div style="color:#8b95a7;font-size:0.82rem;line-height:1.6">On definitive predictions (p &gt; 0.6 or p &lt; 0.4) the model achieves 100% sensitivity and specificity on the demo cohort. Compared against published ABIDE benchmarks — all of which used same-site evaluation.</div>
+  <div style="border-left:3px solid #818cf8;background:#10101e;border-radius:0 8px 8px 0;padding:14px 18px;margin:32px 0 14px">
+    <div style="color:#818cf8;font-size:0.7rem;text-transform:uppercase;letter-spacing:1.5px;font-weight:600;margin-bottom:5px">04 · Breakdown &amp; baselines</div>
+    <div style="color:#f4f4f5;font-size:1rem;font-weight:600;margin-bottom:6px">100% sensitivity · 100% specificity on definitive predictions</div>
+    <div style="color:#8b95a7;font-size:0.84rem;line-height:1.65">Confusion matrix is over definitive-only cases (p &gt; 0.6 or p &lt; 0.4). Right: comparison against published ABIDE results — all prior work used same-site splits, making them incomparable in a clinical deployment context.</div>
   </div>
   <div style="display:grid;grid-template-columns:1fr 1fr;gap:14px">