First multi-prompt result: Qwen 6-bit TTS = 49 (vs 82 LFU) + per-prompt schema

TTS grade for qwen3.6-35b-a3b-6bit-mlx: 49/100 Critical (same model that
scored 82 on LFU). File doesn't parse + bounded-concurrency is fake
(1 worker + inner semaphore = real concurrency 1). Per-task signal:
strong on data-structures, weak on async-pipeline work.

Schema: prompt_id + PILLARS_BY_PROMPT so each entry uses its own 5 pillars.
TODO_submission_tool.md sketches the grade-as-a-tool idea for later.

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

generate_dashboard.py

@@ -11,14 +11,40 @@ DATA = os.path.join(HERE, "data", "benchmark_history.json")
 OUT_DASH = os.path.join(HERE, "dashboard.html")
 PAGES_DIR = os.path.join(HERE, "pages")
 
-PILLARS = ["complexity", "concurrency", "isolation", "memory_edge_cases", "test_integrity"]
-PILLAR_LABELS = {
-    "complexity": "Complexity (O(1))",
-    "concurrency": "Concurrency / Races",
-    "isolation": "Tx Isolation",
-    "memory_edge_cases": "Memory & Edges",
-    "test_integrity": "Test Integrity",
+# Per-prompt pillar definitions. Each prompt is graded on 5 pillars (0-20 each = 100),
+# but the pillar NAMES differ by prompt type. The generator reads the entry's own pillars.
+PILLARS_BY_PROMPT = {
+    "lfu": ["complexity", "concurrency", "isolation", "memory_edge_cases", "test_integrity"],
+    "tts": ["complexity", "concurrency", "error_handling", "resource_safety", "test_integrity"],
+    "mcp": ["schema_io", "transport", "error_handling", "state_safety", "test_integrity"],
+    "rust": ["ownership", "concurrency", "error_handling", "cancellation", "test_integrity"],
+    "data": ["query_safety", "pooling", "transactions", "pagination", "test_integrity"],
+    "automation": ["idempotency", "retry_backoff", "checkpointing", "signal_handling", "test_integrity"],
 }
+PILLARS = PILLARS_BY_PROMPT["lfu"]  # default for any code that still references the global
+PILLAR_LABELS = {
+    # lfu
+    "complexity": "Complexity (O(1))", "concurrency": "Concurrency / Races",
+    "isolation": "Tx Isolation", "memory_edge_cases": "Memory & Edges",
+    "test_integrity": "Test Integrity",
+    # tts / shared
+    "error_handling": "Error Handling", "resource_safety": "Resource & State Safety",
+    # mcp
+    "schema_io": "Schema / I/O", "transport": "Transport",
+    "state_safety": "State Safety",
+    # rust
+    "ownership": "Ownership / Types", "cancellation": "Cancellation / Shutdown",
+    # data
+    "query_safety": "Query Safety", "pooling": "Pooling", "transactions": "Transactions", "pagination": "Pagination",
+    # automation
+    "idempotency": "Idempotency", "retry_backoff": "Retry / Backoff",
+    "checkpointing": "Checkpointing", "signal_handling": "Signal Handling",
+}
+
+def pillars_for(m):
+    """Return the 5 pillar keys for a model entry based on its prompt_id."""
+    pid = m.get("prompt_id", "lfu")
+    return PILLARS_BY_PROMPT.get(pid, PILLARS)
 
 # ---- cyberpunk palette ----
 NEON_CYAN = "#00ffc8"
@@ -144,6 +170,13 @@ tr:hover .score-bar > i{box-shadow:0 0 12px currentColor}
 .fmt-mlx{color:var(--cyan);background:rgba(0,255,200,0.08)}
 .fmt-gguf{color:var(--mag);background:rgba(255,43,214,0.08)}
 .fmt-cloud{color:var(--blue);background:rgba(91,139,255,0.08)}
+.prompt-chip{display:inline-block;font-size:.62rem;letter-spacing:.06em;padding:2px 6px;border-radius:3px;border:1px solid currentColor;font-family:'Fira Code',monospace;text-transform:uppercase}
+.p-lfu{color:var(--cyan);background:rgba(0,255,200,0.06)}
+.p-tts{color:var(--lime);background:rgba(182,255,58,0.06)}
+.p-mcp{color:var(--blue);background:rgba(91,139,255,0.06)}
+.p-rust{color:var(--amber);background:rgba(255,176,0,0.06)}
+.p-data{color:var(--mag);background:rgba(255,43,214,0.06)}
+.p-automation{color:#c084fc;background:rgba(192,132,252,0.06)}
 /* format/quant showdown cards */
 .fcard{background:var(--panel2);border:1px solid rgba(255,255,255,0.06);border-radius:6px;padding:12px 14px;margin-bottom:10px}
 .fcard-h{display:flex;justify-content:space-between;align-items:baseline;gap:10px;flex-wrap:wrap;margin-bottom:9px}
@@ -204,6 +237,7 @@ FOOT = """</div>
 </body></html>"""
 
 def render_dashboard(data):
+    # leaderboard ranked by score (prompt badge distinguishes same-model entries across prompts)
     models = sorted(data["models"], key=lambda m: -m["total_score"])
     n = len(models)
     avg = sum(m["total_score"] for m in models) / n if n else 0
@@ -232,9 +266,11 @@ def render_dashboard(data):
         else:
             fmt_chip = f'<span class="fmt-chip">{esc(m.get("format") or "—")}</span>'
         bar_color = col
+        pid = m.get("prompt_id", "lfu")
+        pchip = f'<span class="prompt-chip p-{pid}">{pid}</span>' if pid != "lfu" else '<span class="prompt-chip p-lfu">lfu</span>'
         rows.append(f"""<tr>
   <td class="rank {'top' if i<=3 else ''}">#{i}</td>
-  <td><div class="model-name">{esc(m['model_name'])}</div>{caveat}</td>
+  <td><div class="model-name">{esc(m['model_name'])} {pchip}</div>{caveat}</td>
   <td><span class="quant-cell mono">{esc(m['quant'])}</span></td>
   <td>{fmt_chip}</td>
   <td class="mono">{speed_str(m)} <span style="color:var(--dim);font-size:.72rem">t/s</span></td>
@@ -257,9 +293,11 @@ def render_dashboard(data):
     bar_speed = json.dumps([m["tok_sec"] for m in local_sorted])
     bar_score = json.dumps([m["total_score"] for m in local_sorted])
 
+    # radar compares top models on the LFU exam (apples-to-apples, same 5 axes)
+    radar_lfu = [m for m in models if m.get("prompt_id", "lfu") == "lfu"][:3]
     radar_labels = json.dumps([PILLAR_LABELS[p] for p in PILLARS])
     radar_sets = []
-    for idx, m in enumerate(radar_models):
+    for idx, m in enumerate(radar_lfu):
         col = SERIES[idx % len(SERIES)]
         radar_sets.append({
             "label": m["model_name"][:24],
@@ -374,7 +412,9 @@ def render_dashboard(data):
     n_crit = sum(1 for m in local if m["verdict"] == "Critical Bugs")
     n_slots = sum(1 for m in local if scans[m["id"]]["slots"])
     n_mono = sum(1 for m in local if scans[m["id"]]["monotonic"])
-    pillar_avg = {p: round(sum(m["breakdown"][p] for m in local)/len(local), 1) for p in PILLARS}
+    # findings stats are LFU-exam-scoped (the common comparison set)
+    lfu_models = [m for m in local if m.get("prompt_id", "lfu") == "lfu"]
+    pillar_avg = {p: round(sum(m["breakdown"][p] for m in lfu_models)/len(lfu_models), 1) for p in PILLARS}
     weakest = min(PILLARS, key=lambda p: pillar_avg[p])
 
     def _fcard(num, label, sub):
@@ -519,14 +559,15 @@ new Chart(document.getElementById('radar'),{{
 
 def render_detail(m, data):
     col, chip = verdict_meta(m["verdict"])
+    m_pillars = pillars_for(m)  # each entry uses its own prompt's 5 pillars
     # derive "what went right" from high pillars, "wrong" from low + critical_bugs
     bd = m["breakdown"]
-    ranked = sorted(PILLARS, key=lambda p: -bd[p])
+    ranked = sorted(m_pillars, key=lambda p: -bd[p])
     rights = [f"{PILLAR_LABELS[p]} ({bd[p]}/20)" for p in ranked if bd[p] >= 16]
     wrongs_pillars = [f"{PILLAR_LABELS[p]} ({bd[p]}/20)" for p in ranked if bd[p] <= 13]
 
     pillar_bars = ""
-    for p in PILLARS:
+    for p in m_pillars:
         v = bd[p]
         c = NEON_LIME if v >= 17 else (NEON_AMBER if v >= 13 else NEON_RED)
         pillar_bars += f"""