diff --git a/training/predict.py b/training/predict.py
index 5389141..8702f9d 100644
--- a/training/predict.py
+++ b/training/predict.py
@@ -1149,6 +1149,17 @@ def _ensure_sub_index():
                 _build_sub_index()
 
 
+def _predicted_tastes_at(profiles, i):
+    """The PREDICTED taste heads (score >= 0.5) for reference-set row i, from the profile matrix.
+    Lets neighbor / substitute cards show a taste read even when nothing is *documented* — the
+    taste columns are the first len(_CLASSIFIERS) of the profile vector. Marked predicted in the UI."""
+    if profiles is None:
+        return []
+    taste_heads = sorted(_CLASSIFIERS)
+    row = profiles[i]
+    return [t for j, t in enumerate(taste_heads) if float(row[j]) >= 0.5]
+
+
 def structural_neighbors(smiles: str, k: int = 8, min_similarity: float = 0.0) -> dict:
     """STRUCTURAL neighbors: the k labeled molecules most structurally similar to the query
     (Tanimoto over Morgan fingerprints), each with its known tastes. Structural look-alikes —
@@ -1158,7 +1169,7 @@ def structural_neighbors(smiles: str, k: int = 8, min_similarity: float = 0.0) -
     if mol is None:
         return {"error": f"unparseable SMILES: {smiles}"}
     _ensure_sub_index()
-    fps, smis, tastes, _aromas, _profiles, _dims = _SUB_INDEX
+    fps, smis, tastes, _aromas, profiles, _dims = _SUB_INDEX
     if not fps:
         return {"neighbors": [], "note": "no reference set loaded (taste_master.parquet absent)"}
     q = _MORGAN.GetFingerprint(mol)
@@ -1179,6 +1190,7 @@ def structural_neighbors(smiles: str, k: int = 8, min_similarity: float = 0.0) -
             continue
         neighbors.append({"smiles": smis[i], "similarity": round(float(sims[i]), 3),
                           "known_tastes": tastes[i],
+                          "predicted_tastes": _predicted_tastes_at(profiles, i),
                           # confident aromas precomputed once in the index — reused so the
                           # endpoint never re-runs the 24 aroma heads per neighbor (8x ~1.3s saved)
                           "aromas": _aromas[i] if i < len(_aromas) else []})
@@ -1220,7 +1232,8 @@ def substitutes(smiles: str, k: int = 8) -> dict:
         if ni is None or Chem.MolToInchiKey(ni).split("-")[0] == self_skel:
             continue
         subs.append({"smiles": smis[i], "profile_match": round(float(sims[i]), 3),
-                     "known_tastes": tastes[i], "aromas": aromas[i] if i < len(aromas) else []})
+                     "known_tastes": tastes[i], "predicted_tastes": _predicted_tastes_at(profiles, i),
+                     "aromas": aromas[i] if i < len(aromas) else []})
         if len(subs) >= k:
             break
     return {"query": Chem.MolToSmiles(mol), "substitutes": subs,
diff --git a/training/workbench.html b/training/workbench.html
index b1d2b44..862a0fa 100644
--- a/training/workbench.html
+++ b/training/workbench.html
@@ -177,6 +177,7 @@
   .kt{display:flex;flex-wrap:wrap;gap:4px;margin-top:3px}
   .rx-from{font-size:10px;color:var(--muted);line-height:1.3;margin-top:3px}
   .kt .tchip{font-size:10px;font-weight:600;padding:1px 8px;border-radius:20px;color:#10140F;text-transform:capitalize}
+  .kt .tchip.pred{background:transparent!important;border:1px solid;font-weight:500}
   .kt .muted{color:var(--muted);font-size:11px}
   .browse{margin:22px 0 4px;border:1px solid #2C6E74;border-radius:12px;padding:15px 18px;background:linear-gradient(180deg,#15211F,#181D24)}
   .browse-head{font-size:12px;font-weight:650;text-transform:uppercase;letter-spacing:.04em;color:var(--brand-2);margin:0 0 12px;display:flex;align-items:center;justify-content:space-between;gap:8px}
@@ -1416,7 +1417,7 @@ <h4>Software &amp; type</h4>
           <div class="nb-name">${x.name||'—'}${x.gras?'<span class="nb-gras" title="Listed in a food-use reference (FDA Substances-Added-to-Food / EU flavourings) — a listing flag, not a GRAS or safety clearance">Food-listed</span>':''}</div>
           ${x.iupac?`<div class="nb-iupac">${x.iupac}</div>`:''}
           <div class="s">${x.smiles}</div>
-          <div class="kt">${x.known_tastes.length?x.known_tastes.map(t=>`<span class="tchip" style="background:${TASTE_COLORS[t]||'var(--muted)'}">${t}</span>`).join(''):'<span class="muted">no known taste</span>'}</div>
+          <div class="kt">${(()=>{const doc=(x.known_tastes||[]).map(t=>`<span class="tchip" style="background:${TASTE_COLORS[t]||'var(--muted)'}">${t}</span>`);const pred=(x.predicted_tastes||[]).filter(t=>!(x.known_tastes||[]).includes(t)).map(t=>`<span class="tchip pred" style="color:${TASTE_COLORS[t]||'var(--muted)'};border-color:${TASTE_COLORS[t]||'var(--muted)'}" title="predicted taste (model)">${t} •</span>`);return [...doc,...pred].join('')||'<span class="muted">no taste read</span>';})()}</div>
           ${(x.aroma&&x.aroma.length)?`<div class="nb-aroma" title="aroma (found = documented, else predicted •)">${x.aroma.map(a=>`<span class="atag" style="color:${aromaColor(a.odor)};border-color:${aromaColor(a.odor)}44">${a.odor}${a.source==='predicted'?' •':''}</span>`).join('')}</div>`:''}
         </div>
         <div class="sim">${Math.round((x[scoreKey]||0)*100)}% match</div>