Update app.py

app.py

@@ -1,4 +1,4 @@
-import json, os, time, csv, io, re, html, hashlib, sqlite3, threading
+import json, os, time, csv, io, re, html, hashlib, sqlite3, threading, random
 from datetime import datetime
 from dataclasses import dataclass, field
 from typing import List, Dict
@@ -99,10 +99,10 @@ def load_tasks():
         raise FileNotFoundError("Dataset not found!")
 
 ALL_TASKS = load_tasks()
-print(f"✅ FINAL Bench v4.1: {len(ALL_TASKS)} tasks loaded")
+print(f"✅ FINAL Bench v4.2: {len(ALL_TASKS)} tasks loaded")
 
 # ════════════════════════════════════════════════════════════════
-# §3. Multi-Provider Model Registry (Eval + Judge SAME)
+# §3. Multi-Provider Model Registry
 # ════════════════════════════════════════════════════════════════
 
 PROVIDER_MODELS = {
@@ -119,13 +119,12 @@ PROVIDER_MODELS = {
         "claude-haiku-4-5-20251001":    "Claude Haiku 4.5",
     },
     "Google": {
-        "gemini-3-pro":         "Gemini 3 Pro",
-        "gemini-2.5-pro":       "Gemini 2.5 Pro",
-        "gemini-2.5-flash":     "Gemini 2.5 Flash",
+        "gemini-2.5-flash":             "Gemini 2.5 Flash",
+        "gemini-2.5-pro":               "Gemini 2.5 Pro",
+        "gemini-2.0-flash":             "Gemini 2.0 Flash",
     },
 }
 
-# Build unified model list — used for BOTH eval and judge dropdowns
 ALL_MODELS = {}
 for prov, models in PROVIDER_MODELS.items():
     for mid, label in models.items():
@@ -136,7 +135,6 @@ DEFAULT_EVAL  = "GPT-5.2 (flagship) [OpenAI]"
 DEFAULT_JUDGE = "GPT-5.2 (flagship) [OpenAI]"
 
 def _resolve_model(choice):
-    """Resolve dropdown choice → (model_id, provider)"""
     info = ALL_MODELS.get(choice, {})
     return info.get("id", "gpt-5.2"), info.get("provider", "OpenAI")
 
@@ -150,7 +148,7 @@ def _strip_think(text):
         text = re.sub(rf'<{tag}>.*?</{tag}>','',text,flags=re.DOTALL)
     return text.strip()
 
-# --- OpenAI ---
+# ── OpenAI ──
 def call_openai(prompt, system="", api_key="", model="gpt-5.2",
                 max_tokens=8192, temperature=0.6, reasoning_effort=None,
                 json_mode=False, json_schema=None):
@@ -171,19 +169,20 @@ def call_openai(prompt, system="", api_key="", model="gpt-5.2",
         try:
             r=requests.post("https://api.openai.com/v1/chat/completions",
                             headers=headers,data=json.dumps(payload),timeout=300)
-            if r.status_code==429: time.sleep(5*(attempt+1)); continue
-            r.raise_for_status(); c=r.json()["choices"][0]["message"]["content"]
+            r.raise_for_status()
+            c=r.json()["choices"][0]["message"]["content"]
             return _strip_think(c) if c else "[EMPTY]"
         except requests.exceptions.HTTPError:
+            if r.status_code==429: time.sleep(5*(attempt+1)); continue
             try: err=r.json().get("error",{}).get("message","")
             except: err=str(r.status_code)
             if attempt<2: time.sleep(3*(attempt+1)); continue
-            return f"[API_ERROR] {err}"
+            return f"[API_ERROR] OpenAI {r.status_code}: {err}"
         except Exception as e:
             if attempt<2: time.sleep(3*(attempt+1))
             else: return f"[API_ERROR] {e}"
 
-# --- Anthropic (★ data=json.dumps, 429+529 retry) ---
+# ── Anthropic Claude (★ 참고코드 동일 패턴) ──
 def call_anthropic(prompt, system="", api_key="", model="claude-opus-4-6",
                    max_tokens=8192, temperature=0.6):
     headers={
@@ -198,8 +197,6 @@ def call_anthropic(prompt, system="", api_key="", model="claude-opus-4-6",
         try:
             r=requests.post("https://api.anthropic.com/v1/messages",
                             headers=headers,data=json.dumps(payload),timeout=300)
-            if r.status_code==429: time.sleep(5*(attempt+1)); continue
-            if r.status_code==529: time.sleep(8*(attempt+1)); continue
             r.raise_for_status()
             resp=r.json()
             text_parts=[]
@@ -209,69 +206,92 @@ def call_anthropic(prompt, system="", api_key="", model="claude-opus-4-6",
             c="\n".join(text_parts)
             return _strip_think(c) if c else "[EMPTY]"
         except requests.exceptions.HTTPError:
+            if r.status_code==429: time.sleep(5*(attempt+1)); continue
+            if r.status_code==529: time.sleep(8*(attempt+1)); continue
             try: err=r.json().get("error",{}).get("message","")
             except: err=str(r.status_code)
-            if attempt<2: time.sleep(3*(attempt+1)); continue
-            return f"[API_ERROR] {err}"
+            return f"[API_ERROR] Claude {r.status_code}: {err}"
         except Exception as e:
             if attempt<2: time.sleep(3*(attempt+1))
             else: return f"[API_ERROR] {e}"
 
-# --- Google Gemini (★ x-goog-api-key header, data=json.dumps, thinking filter) ---
+# ── Google Gemini (★★★ 참고코드와 100% 동일 패턴 ★★★) ──
 GEMINI_API_BASE = "https://generativelanguage.googleapis.com/v1beta"
 
-def call_gemini(prompt, system="", api_key="", model="gemini-3-pro",
+def call_gemini(prompt, system="", api_key="", model="gemini-2.5-flash",
                 max_tokens=8192, temperature=1.0, json_mode=False):
-    url=f"{GEMINI_API_BASE}/models/{model}:generateContent"
-    headers={
-        "Content-Type":"application/json",
-        "x-goog-api-key":api_key,
+    """Google Gemini generateContent REST API
+    ★ x-goog-api-key 헤더 인증
+    ★ data=json.dumps(payload) 전송
+    ★ thinking part (thought:True) 스킵
+    """
+    url = f"{GEMINI_API_BASE}/models/{model}:generateContent"
+    headers = {
+        "Content-Type": "application/json",
+        "x-goog-api-key": api_key,
     }
-    contents=[{"role":"user","parts":[{"text":prompt}]}]
-    gen_config={"maxOutputTokens":max_tokens,"temperature":temperature}
-    payload={"contents":contents,"generationConfig":gen_config}
+    contents = [{"role": "user", "parts": [{"text": prompt}]}]
+    gen_config = {"maxOutputTokens": max_tokens, "temperature": temperature}
+    payload = {"contents": contents, "generationConfig": gen_config}
     if system:
-        payload["systemInstruction"]={"parts":[{"text":system}]}
+        payload["systemInstruction"] = {"parts": [{"text": system}]}
     if json_mode:
-        gen_config["responseMimeType"]="application/json"
+        gen_config["responseMimeType"] = "application/json"
     for attempt in range(3):
         try:
-            r=requests.post(url,headers=headers,data=json.dumps(payload),timeout=300)
-            if r.status_code==429: time.sleep(5*(attempt+1)); continue
-            if r.status_code==503: time.sleep(8*(attempt+1)); continue
+            r = requests.post(url, headers=headers, data=json.dumps(payload), timeout=300)
+            # ★ raise_for_status FIRST — 참고코드 동일 패턴
             r.raise_for_status()
-            data=r.json()
-            candidates=data.get("candidates",[])
+            data = r.json()
+            candidates = data.get("candidates", [])
             if not candidates:
-                block_reason=data.get("promptFeedback",{}).get("blockReason","UNKNOWN")
-                return f"[BLOCKED] Gemini blocked response: {block_reason}"
-            parts=candidates[0].get("content",{}).get("parts",[])
-            result=[]
+                block_reason = data.get("promptFeedback", {}).get("blockReason", "UNKNOWN")
+                print(f"  [Gemini] BLOCKED: {block_reason}")
+                return f"[API_ERROR] Gemini BLOCKED: {block_reason}"
+            parts = candidates[0].get("content", {}).get("parts", [])
+            result = []
             for p in parts:
                 if "text" in p:
-                    if p.get("thought",False): continue  # skip thinking parts
+                    if p.get("thought", False):
+                        continue  # ★ thinking part skip
                     result.append(p["text"])
-            c="\n".join(result) if result else ""
+            c = "\n".join(result) if result else ""
             return _strip_think(c) if c else "[EMPTY]"
         except requests.exceptions.HTTPError:
-            try: err=r.json().get("error",{}).get("message","")
-            except: err=str(r.status_code)
-            if attempt<2: time.sleep(3*(attempt+1)); continue
+            # ★ 참고코드 동일: 429/503만 retry, 나머지는 즉시 에러 반환
+            if r.status_code == 429:
+                time.sleep(5 * (attempt + 1) + random.uniform(0, 2))
+                continue
+            if r.status_code == 503:
+                time.sleep(8 * (attempt + 1) + random.uniform(0, 3))
+                continue
+            try:
+                err = r.json().get("error", {}).get("message", "")
+            except:
+                err = str(r.status_code)
+            print(f"  [Gemini] ERROR {r.status_code}: {err[:200]}")
             return f"[API_ERROR] Gemini {r.status_code}: {err}"
         except Exception as e:
-            if attempt<2: time.sleep(3*(attempt+1))
-            else: return f"[API_ERROR] {e}"
+            print(f"  [Gemini] Exception: {e}")
+            if attempt < 2:
+                time.sleep(3 * (attempt + 1))
+            else:
+                return f"[API_ERROR] Gemini: {e}"
 
-# --- Unified Dispatcher ---
+# ── Unified Dispatcher ──
 def call_model(prompt, system="", api_key="", model_id="gpt-5.2",
                provider="OpenAI", max_tokens=8192, temperature=0.6):
-    if provider=="OpenAI":    return call_openai(prompt,system,api_key,model_id,max_tokens,temperature)
-    elif provider=="Anthropic": return call_anthropic(prompt,system,api_key,model_id,max_tokens,temperature)
-    elif provider=="Google":  return call_gemini(prompt,system,api_key,model_id,max_tokens,temperature)
+    if provider == "OpenAI":
+        return call_openai(prompt, system, api_key, model_id, max_tokens, temperature)
+    elif provider == "Anthropic":
+        return call_anthropic(prompt, system, api_key, model_id, max_tokens, temperature)
+    elif provider == "Google":
+        # ★ Gemini는 temperature=1.0 권장 (thinking 모델)
+        return call_gemini(prompt, system, api_key, model_id, max_tokens, temperature=1.0)
     return f"[API_ERROR] Unknown provider: {provider}"
 
 # ════════════════════════════════════════════════════════════════
-# §5. Judge — Multi-Provider (OpenAI / Anthropic / Google)
+# §5. Judge — Multi-Provider
 # ════════════════════════════════════════════════════════════════
 
 JUDGE_SYSTEM = """You are a FINAL Bench judge for AGI-Level Verification.
@@ -292,9 +312,8 @@ G_PivotDetection: Found reversing premise? H_DecisionUnderUncertainty: Scenario
 
 STRICT: 1.0=AGI-worthy 0.75=expert 0.5=competent 0.25=gaps 0.0=failure
 
-IMPORTANT: Output ONLY valid JSON with NO extra text before or after:
-{"scores":{"process_quality":X,"metacognitive_accuracy":X,"error_recovery":X,"integration_depth":X,"final_correctness":X},"comment":"<50 words>"}
-Where each X is one of: 0.0, 0.25, 0.5, 0.75, 1.0"""
+IMPORTANT: Output ONLY valid JSON with NO extra text:
+{"scores":{"process_quality":X,"metacognitive_accuracy":X,"error_recovery":X,"integration_depth":X,"final_correctness":X},"comment":"<50 words>"}"""
 
 def _build_judge_schema():
     sp={k:{"type":"number","enum":[0.0,0.25,0.5,0.75,1.0]} for k in RUBRIC}
@@ -323,18 +342,14 @@ Score: process_quality, metacognitive_accuracy, error_recovery, integration_dept
 Apply {task.ticos_type} bonus criteria.
 Output ONLY JSON: {{"scores":{{...}},"comment":"..."}}"""
 
-
 def _parse_judge_json(text):
-    """Parse judge response → dict with scores, works for all providers"""
     if not text or text.startswith("[API_ERROR") or text=="[EMPTY]":
         return None
     cleaned = _strip_think(text)
     VALID = {0.0, 0.25, 0.5, 0.75, 1.0}
     keys = list(RUBRIC.keys())
-
-    # Method 1: Direct JSON parse
+    # Method 1: Direct JSON
     try:
-        # Strip markdown fences
         t = re.sub(r'^```(?:json)?\s*', '', cleaned.strip())
         t = re.sub(r'\s*```$', '', t.strip())
         data = json.loads(t)
@@ -344,10 +359,8 @@ def _parse_judge_json(text):
                 v = float(data["scores"].get(k, 0.5))
                 scores[k] = min(VALID, key=lambda x, v=v: abs(x - v))
             return {"scores": scores, "comment": data.get("comment", "ok")}
-    except:
-        pass
-
-    # Method 2: Find JSON object in text
+    except: pass
+    # Method 2: Search JSON
     try:
         m = re.search(r'\{[^{}]*"scores"\s*:\s*\{[^{}]*\}[^{}]*\}', cleaned, re.DOTALL)
         if m:
@@ -358,62 +371,44 @@ def _parse_judge_json(text):
                     v = float(data["scores"].get(k, 0.5))
                     scores[k] = min(VALID, key=lambda x, v=v: abs(x - v))
                 return {"scores": scores, "comment": data.get("comment", "parsed")}
-    except:
-        pass
-
-    # Method 3: Regex extraction
+    except: pass
+    # Method 3: Regex
     try:
         sc = {}
         for k in keys:
             m2 = re.search(rf'["\']?{re.escape(k)}["\']?\s*[:=]\s*([\d.]+)', cleaned, re.IGNORECASE)
             if m2:
                 v = float(m2.group(1))
-                if 0 <= v <= 1:
-                    sc[k] = min(VALID, key=lambda x, v=v: abs(x - v))
+                if 0 <= v <= 1: sc[k] = min(VALID, key=lambda x, v=v: abs(x - v))
         if len(sc) >= 3:
             for k in keys:
                 if k not in sc: sc[k] = 0.5
             return {"scores": sc, "comment": "regex_parsed"}
-    except:
-        pass
-
+    except: pass
     return None
 
-
 def call_judge(prompt, system, api_key, model_id, provider, temperature=0.1, max_tokens=2048):
-    """★ Universal Judge — routes to correct provider with JSON enforcement"""
-
     if provider == "OpenAI":
-        # OpenAI: use structured output (best quality)
         raw = call_openai(prompt, system=system, api_key=api_key, model=model_id,
-                          max_tokens=max_tokens, temperature=temperature,
-                          json_schema=JUDGE_SCHEMA)
+                          max_tokens=max_tokens, temperature=temperature, json_schema=JUDGE_SCHEMA)
         result = _parse_judge_json(raw)
-        if result:
-            return result
-        # Fallback: try without structured output
+        if result: return result
         raw2 = call_openai(prompt, system=system, api_key=api_key, model=model_id,
                            max_tokens=max_tokens, temperature=temperature, json_mode=True)
         return _parse_judge_json(raw2)
-
     elif provider == "Anthropic":
-        # Anthropic: prompt-based JSON enforcement
         raw = call_anthropic(prompt, system=system, api_key=api_key, model=model_id,
                              max_tokens=max_tokens, temperature=temperature)
         return _parse_judge_json(raw)
-
     elif provider == "Google":
-        # Google: JSON mode supported
+        # ★ Gemini judge도 temperature=1.0 고정 (thinking 모델 호환)
         raw = call_gemini(prompt, system=system, api_key=api_key, model=model_id,
-                          max_tokens=max_tokens, temperature=temperature, json_mode=True)
+                          max_tokens=max_tokens, temperature=1.0, json_mode=True)
         result = _parse_judge_json(raw)
-        if result:
-            return result
-        # Fallback without json_mode
+        if result: return result
         raw2 = call_gemini(prompt, system=system, api_key=api_key, model=model_id,
-                           max_tokens=max_tokens, temperature=temperature, json_mode=False)
+                           max_tokens=max_tokens, temperature=1.0, json_mode=False)
         return _parse_judge_json(raw2)
-
     return None
 
 # ════════════════════════════════════════════════════════════════
@@ -476,12 +471,23 @@ def _init_db():
     c=sqlite3.connect(DB_PATH)
     c.execute("CREATE TABLE IF NOT EXISTS eval_results(run_id TEXT,task_id TEXT,model_response TEXT,judge_response TEXT,weighted_score REAL,timestamp REAL,PRIMARY KEY(run_id,task_id))")
     c.commit(); c.close()
-def _make_run_id(m): return hashlib.md5(f"FINALv41_BL_{m}".encode()).hexdigest()[:12]
+def _make_run_id(m): return hashlib.md5(f"FINALv42_BL_{m}".encode()).hexdigest()[:12]
 def _save_result(rid,tid,resp,jresp,sc):
     c=sqlite3.connect(DB_PATH); c.execute("INSERT OR REPLACE INTO eval_results VALUES(?,?,?,?,?,?)",(rid,tid,resp,jresp,sc,time.time())); c.commit(); c.close()
 def _load_all(rid):
-    c=sqlite3.connect(DB_PATH); cur=c.execute("SELECT task_id,model_response,judge_response,weighted_score FROM eval_results WHERE run_id=?",(rid,)); rows=cur.fetchall(); c.close()
-    return {r[0]:{"response":r[1],"judge":r[2],"score":r[3]} for r in rows}
+    """★ 캐시 로드 시 실패 결과(score=0 + API_ERROR) 자동 제외 → 재시도 보장"""
+    c=sqlite3.connect(DB_PATH)
+    cur=c.execute("SELECT task_id,model_response,judge_response,weighted_score FROM eval_results WHERE run_id=?",(rid,))
+    rows=cur.fetchall(); c.close()
+    result = {}
+    for r in rows:
+        resp = r[1] or ""
+        score = r[3]
+        # ★ API 에러/빈 응답/0점은 캐시에서 제외 → 다음 실행 시 재시도
+        if score <= 0 and (resp.startswith("[API_ERROR") or resp.startswith("[BLOCKED") or resp == "[EMPTY]" or resp.startswith("[ERROR")):
+            continue
+        result[r[0]] = {"response": resp, "judge": r[2], "score": score}
+    return result
 def _clear_run(rid):
     c=sqlite3.connect(DB_PATH); c.execute("DELETE FROM eval_results WHERE run_id=?",(rid,)); c.commit(); c.close()
 _init_db()
@@ -546,9 +552,13 @@ def _build_progress_table(results, tasks):
         info=DOMAIN_INFO.get(t.domain,{"icon":"?","color":"#999"})
         gb=f'<span style="background:{"#c62828" if t.grade=="A" else "#1565c0" if t.grade=="B" else "#6a1b9a"};color:#fff;padding:1px 6px;border-radius:4px;font-size:0.8em">{t.grade}</span>'
         if t.task_id in results:
-            s=results[t.task_id]["score"]
+            d=results[t.task_id]; s=d["score"]; resp=d.get("response","")
             if s<0:
-                rows+=f'<tr style="background:#fff3e0"><td>{t.task_id}</td><td>{info["icon"]} {t.domain[:15]}</td><td>{gb}</td><td>{t.ticos_type.split("_")[0]}</td><td>{t.difficulty}</td><td style="color:#ff9800">❌</td><td>—</td></tr>'
+                rows+=f'<tr style="background:#fff3e0"><td>{t.task_id}</td><td>{info["icon"]} {t.domain[:15]}</td><td>{gb}</td><td>{t.ticos_type.split("_")[0]}</td><td>{t.difficulty}</td><td style="color:#ff9800">❌ JF</td><td>—</td></tr>'
+            elif s==0 and resp and (resp.startswith("[API_ERROR") or resp.startswith("[BLOCKED") or resp=="[EMPTY]"):
+                # ★ API 에러를 명확하게 표시
+                err_short=html.escape(resp[:60])
+                rows+=f'<tr style="background:#ffebee"><td>{t.task_id}</td><td>{info["icon"]} {t.domain[:15]}</td><td>{gb}</td><td>{t.ticos_type.split("_")[0]}</td><td>{t.difficulty}</td><td colspan="2" style="color:#c62828;font-size:0.75em">🚫 {err_short}</td></tr>'
             else:
                 c=_sc(s)
                 rows+=f'<tr><td>{t.task_id}</td><td>{info["icon"]} {t.domain[:15]}</td><td>{gb}</td><td>{t.ticos_type.split("_")[0]}</td><td>{t.difficulty}</td><td><div class="score-bar"><div class="score-fill" style="width:{min(s,100)}%;background:{c}"></div></div></td><td style="font-weight:700;color:{c}">{s:.1f}</td></tr>'
@@ -571,6 +581,8 @@ def _build_summary_card(results, tasks, eval_label, judge_label, hf_status):
         if gs: a=np.mean(gs); gh+=f'<span style="margin-right:14px">{g}×{GRADE_WEIGHT[g]}: <b style="color:{_sc(a)}">{a:.1f}</b></span>'
     done=sum(1 for t in tasks if t.task_id in results)
     jf=sum(1 for t in tasks if t.task_id in results and results[t.task_id]["score"]<0)
+    # API errors
+    api_errs=sum(1 for t in tasks if t.task_id in results and results[t.task_id]["score"]==0 and (results[t.task_id].get("response","") or "").startswith("["))
     # MA-ER Gap
     ma_vals,er_vals=[],[]
     for tid,d in results.items():
@@ -582,21 +594,21 @@ def _build_summary_card(results, tasks, eval_label, judge_label, hf_status):
             if "error_recovery" in sc: er_vals.append(float(sc["error_recovery"]))
         except: pass
     avg_ma=np.mean(ma_vals) if ma_vals else 0; avg_er=np.mean(er_vals) if er_vals else 0
-    gap=avg_ma-avg_er
-    gc="#f44336" if gap>0.2 else "#ff9800" if gap>0.1 else "#4caf50"
+    gap=avg_ma-avg_er; gc="#f44336" if gap>0.2 else "#ff9800" if gap>0.1 else "#4caf50"
     gl="Declaration-Action Gap" if gap>0.2 else "Moderate Gap" if gap>0.1 else "Balanced"
-    # Pass checks
     ad=[t.domain for t in tasks if t.grade=="A"]
     asc_vals=[dom_avgs[d] for d in set(ad) if d in dom_avgs]
     aa=np.mean(asc_vals) if asc_vals else 0
     checks=[("Score≥80",final>=80),("Axes≥60",all(v>=60 for v in axis.values())),(f"A-avg≥75({aa:.0f})",aa>=75)]
     ch="".join([f'<span style="margin-right:8px">{"✅" if ok else "❌"}{lb}</span>' for lb,ok in checks])
+    err_html=f'<div style="color:#ff5722;font-size:0.82em;margin-top:4px">⚠️ API Errors: {api_errs} tasks</div>' if api_errs else ""
     return f"""{CSS}<div class="summary-card">
 <div style="text-align:center">
 <div class="stage-badge" style="background:{stage['color']}">{stage['name']}</div>
 <h2 style="margin:6px 0;font-size:1.6em">🤖 Baseline FINAL: {final:.1f}</h2>
 <p style="color:#aaa;font-size:0.85em">{stage['label']} · Base {base:.1f} × HAR {har_p:.3f} · {done}/{len(tasks)}{f" · JF={jf}" if jf else ""}</p>
 <p style="color:#8af;font-size:0.82em;margin:4px 0">Eval: {eval_label} · Judge: {judge_label}</p>
+{err_html}
 </div><hr style="border-color:#333;margin:12px 0">
 <h4 style="color:#aaa;margin:6px 0">🎯 5-Axis Scores</h4>{ax_html}
 <hr style="border-color:#333;margin:10px 0">
@@ -637,18 +649,24 @@ def _build_detail_view(results, tasks):
 def _eval_single(task, run_id, eval_api_key, eval_model_id, eval_provider,
                  judge_api_key, judge_model_id, judge_provider, state):
     try:
-        # 1) Eval model call
         sys_p=(f"You are being evaluated on FINAL Bench.\nTask: {task.ticos_type}\n"
                f"State confidence (0-100%) for EVERY claim. If wrong, EXPLICITLY backtrack. "
                f"If unsure, say so honestly.")
+        print(f"  ▶ {task.task_id} → {eval_provider}/{eval_model_id}")
         model_response=call_model(task.prompt,system=sys_p,api_key=eval_api_key,
                                    model_id=eval_model_id,provider=eval_provider,max_tokens=12288)
-        if model_response.startswith("[API_ERROR") or model_response=="[EMPTY]":
+        if (model_response.startswith("[API_ERROR") or
+            model_response.startswith("[BLOCKED") or
+            model_response=="[EMPTY]"):
+            print(f"  ✗ {task.task_id}: {model_response[:100]}")
+            # ★ API 에러는 저장하되, _load_all에서 자동 제외됨
             _save_result(run_id,task.task_id,model_response,"{}",0)
-            with state["lock"]: state["done"]+=1; state["errors"].append(f"{task.task_id}: {model_response[:50]}")
+            with state["lock"]:
+                state["done"]+=1
+                state["errors"].append(f"{task.task_id}: {model_response[:80]}")
             return task.task_id,{"response":model_response,"judge":"{}","score":0}
 
-        # 2) Judge call — any provider
+        print(f"  ✓ {task.task_id} response len={len(model_response)}")
         jp = build_judge_prompt(task, model_response)
         jd = call_judge(jp, system=JUDGE_SYSTEM, api_key=judge_api_key,
                         model_id=judge_model_id, provider=judge_provider)
@@ -671,6 +689,7 @@ def _eval_single(task, run_id, eval_api_key, eval_model_id, eval_provider,
             if len(state["active"])>10: state["active"]=state["active"][-10:]
         return task.task_id,{"response":model_response,"judge":jj,"score":ws}
     except Exception as e:
+        print(f"  ✗ {task.task_id} EXCEPTION: {e}")
         with state["lock"]: state["done"]+=1; state["errors"].append(f"{task.task_id}: {str(e)[:60]}")
         _save_result(run_id,task.task_id,f"[ERROR] {e}","{}",0)
         return task.task_id,{"response":f"[ERROR] {e}","judge":"{}","score":0}
@@ -709,7 +728,11 @@ def _prog_html(state, pending):
     ac=state.get("active",[])
     if ac: o+='<div style="margin-top:8px">🔄 '+" ".join([f'<span style="background:#e3f2fd;padding:2px 6px;border-radius:4px;font-size:0.78em">{a}</span>' for a in ac[-8:]])+'</div>'
     er=state.get("errors",[])
-    if er: o+=f'<div style="color:#c62828;margin-top:6px;font-size:0.8em">{" · ".join(["⚠️"+html.escape(e[:30]) for e in er[-3:]])}</div>'
+    if er:
+        o+=f'<div style="color:#c62828;margin-top:6px;font-size:0.8em;max-height:120px;overflow-y:auto">'
+        for e in er[-6:]:
+            o+=f'<div>⚠️ {html.escape(e[:100])}</div>'
+        o+='</div>'
     return o+'</div>'
 
 def _bg_eval(eval_api_key, eval_model_id, eval_provider, eval_label,
@@ -720,9 +743,11 @@ def _bg_eval(eval_api_key, eval_model_id, eval_provider, eval_label,
         with _EVAL_STATE["lock"]:
             _EVAL_STATE["start_time"]=time.time()
             _EVAL_STATE["message"]=f"⚡ Eval: {eval_label} · Judge: {judge_label} · {len(tasks)} tasks"
+        # ★ _load_all은 이제 실패 결과를 자동 제외함
         results=dict(_load_all(run_id))
         cached=sum(1 for t in tasks if t.task_id in results)
         pending=[t for t in tasks if t.task_id not in results]
+        print(f"  📊 Cached (valid): {cached} / Pending: {len(pending)} / Total: {len(tasks)}")
         gt={}
         for t in pending: gt.setdefault(t.grade,[]).append(t)
         with _EVAL_STATE["lock"]:
@@ -766,6 +791,8 @@ def _bg_eval(eval_api_key, eval_model_id, eval_provider, eval_label,
             _EVAL_STATE["message"]=f"🏁 {stage['name']} — FINAL={final:.1f} · {elapsed}s"
             _EVAL_STATE["running"]=False; _EVAL_STATE["finished"]=True
     except Exception as e:
+        print(f"  ❌ Fatal: {e}")
+        import traceback; traceback.print_exc()
         with _EVAL_STATE["lock"]:
             _EVAL_STATE["message"]=f"❌ Fatal: {str(e)[:100]}"
             _EVAL_STATE["running"]=False; _EVAL_STATE["finished"]=True
@@ -834,7 +861,7 @@ def _poll():
 
 HEADER = """
 <div style="text-align:center;padding:16px 0">
-<h1 style="margin:0;font-size:1.8em">🏆 FINAL Bench v4.1 — Baseline Evaluation</h1>
+<h1 style="margin:0;font-size:1.8em">🏆 FINAL Bench v4.2 — Baseline Evaluation</h1>
 <h2 style="margin:4px 0;color:#555;font-size:1.05em">Frontier Intelligence Nexus for AGI-Level Verification</h2>
 <p style="color:#888;font-size:0.88em;max-width:720px;margin:8px auto">
 <b>100 Tasks · 15 Domains · 8 TICOS · 5-Axis · 5-Stage AGI Grade</b><br>
@@ -844,45 +871,38 @@ Both <b>Eval</b> and <b>Judge</b> support OpenAI / Anthropic / Google
 <div style="display:flex;justify-content:center;gap:6px;margin-top:8px;flex-wrap:wrap;font-size:0.82em">
 <span style="background:#e3f2fd;padding:2px 10px;border-radius:12px">OpenAI · GPT-5.2 / 5-Mini / 4.1 / o4-mini / 4o</span>
 <span style="background:#fce4ec;padding:2px 10px;border-radius:12px">Anthropic · Opus 4.6 / Sonnet 4.5 / Haiku 4.5</span>
-<span style="background:#e8f5e9;padding:2px 10px;border-radius:12px">Google · Gemini 3 Pro / 2.5 Pro / 2.5 Flash</span>
+<span style="background:#e8f5e9;padding:2px 10px;border-radius:12px">Google · Gemini 2.5 Flash / 2.5 Pro / 2.0 Flash</span>
 </div>
 <div style="background:rgba(233,69,96,0.1);border:1px solid #e94560;border-radius:10px;padding:10px;margin:12px auto;max-width:600px">
 <p style="color:#e94560;font-size:0.85em;margin:0">🔒 <b>MetaCog (Self-Correction Protocol): COMING SOON</b></p>
-<p style="color:#888;font-size:0.78em;margin:4px 0 0 0">
-3-Phase Protocol (Initial → Self-Review → Correction) — paper's core contribution.
-</p></div>
+</div>
 <div style="display:flex;justify-content:center;gap:8px;margin-top:8px;font-size:0.78em">
 <a href="https://huggingface.co/datasets/FINAL-Bench/Metacognitive" target="_blank" style="background:#333;color:#fff;padding:3px 10px;border-radius:10px;text-decoration:none">📊 Dataset</a>
 <a href="https://huggingface.co/spaces/FINAL-Bench/Leaderboard" target="_blank" style="background:#333;color:#fff;padding:3px 10px;border-radius:10px;text-decoration:none">🏆 Leaderboard</a>
 </div></div>"""
 
 def create_app():
-    with gr.Blocks(title="FINAL Bench v4.1",theme=gr.themes.Soft(),
+    with gr.Blocks(title="FINAL Bench v4.2",
                    css=".gradio-container{max-width:1100px !important} header{display:none!important}") as app:
         gr.HTML(HEADER)
 
-        # --- API Keys ---
         gr.Markdown("### 🔑 API Keys")
-        gr.HTML('<p style="color:#888;font-size:0.82em;margin:0 0 6px 0">Enter the API key matching each model\'s provider. If both models use the same provider, you can enter the same key.</p>')
+        gr.HTML('<p style="color:#888;font-size:0.82em;margin:0 0 6px 0">Enter the API key matching each model\'s provider. Same key OK if both use same provider.</p>')
         with gr.Row():
             eval_api_key=gr.Textbox(label="🤖 Eval Model API Key",type="password",
                 placeholder="sk-... / sk-ant-... / AIza...",
-                info="OpenAI / Anthropic / Google key for the evaluation target",scale=3)
+                info="OpenAI / Anthropic / Google key for eval",scale=3)
             judge_api_key=gr.Textbox(label="⚖️ Judge Model API Key",type="password",
                 placeholder="sk-... / sk-ant-... / AIza...",
-                info="OpenAI / Anthropic / Google key for the judge",scale=3)
+                info="OpenAI / Anthropic / Google key for judge",scale=3)
 
-        # --- Model Selection (SAME choices for both) ---
         gr.Markdown("### 🤖 Model Selection")
         with gr.Row():
             eval_m=gr.Dropdown(label="🤖 Evaluation Target",choices=MODEL_CHOICES,
-                value=DEFAULT_EVAL,
-                info="Model to be evaluated on FINAL Bench tasks",scale=3)
+                value=DEFAULT_EVAL,info="Model to evaluate",scale=3)
             judge_m=gr.Dropdown(label="⚖️ Judge Model",choices=MODEL_CHOICES,
-                value=DEFAULT_JUDGE,
-                info="Model that scores the evaluation responses",scale=3)
+                value=DEFAULT_JUDGE,info="Model that scores responses",scale=3)
 
-        # --- Settings ---
         gr.Markdown("### ⚙️ Settings")
         with gr.Row():
             gf=gr.Dropdown(["All","A","B","C"],value="All",label="Grade Filter",scale=1)
@@ -890,14 +910,12 @@ def create_app():
             mt=gr.Slider(1,100,value=100,step=1,label="Max Tasks",scale=1)
             nw=gr.Slider(1,10,value=5,step=1,label="Workers",scale=1)
 
-        # --- Buttons ---
         with gr.Row():
             s_btn=gr.Button("▶️ Start (Resume)",variant="primary",size="lg",scale=2)
             f_btn=gr.Button("🚀 Fresh Start",variant="secondary",size="lg",scale=2)
             x_btn=gr.Button("⏹️ Stop",variant="stop",size="lg",scale=1)
-        status=gr.Textbox(label="Status",interactive=False,max_lines=1)
+        status=gr.Textbox(label="Status",interactive=False,max_lines=2)
 
-        # --- Results ---
         with gr.Tabs():
             with gr.Tab("📊 Progress"): p_html=gr.HTML()
             with gr.Tab("📋 Results"): t_html=gr.HTML()
@@ -913,17 +931,10 @@ def create_app():
         f_btn.click(fn=lambda *a:_start_eval(*a,fresh=True),inputs=eval_ins,outputs=[status])
         x_btn.click(fn=_stop,outputs=[status])
 
-        gr.Markdown("""---
-<center><b>FINAL Bench v4.1</b> — Baseline (Non-AGI) · Multi-Provider Eval & Judge<br>
-100 Tasks · 5-Axis · 5-Stage · OpenAI / Anthropic / Google<br>
-🔒 MetaCog (Self-Correction Protocol): <b>COMING SOON</b><br>
-Apache 2.0 · <b>Ginigen AI</b> — Choi Sunyoung</center>""")
-    return app
-
 if __name__=="__main__":
     sg,sd={},{}
     for t in ALL_TASKS: sg[t.grade]=sg.get(t.grade,0)+1; sd[t.domain]=sd.get(t.domain,0)+1
-    print(f"\n{'='*60}\n  FINAL Bench v4.1 — Baseline (Non-AGI)\n  Eval & Judge: OpenAI / Anthropic / Google\n{'='*60}")
+    print(f"\n{'='*60}\n  FINAL Bench v4.2 — Baseline (Non-AGI)\n  Eval & Judge: OpenAI / Anthropic / Google\n{'='*60}")
     print(f"  {len(ALL_TASKS)} tasks | {len(sd)} domains")
     for g in ["A","B","C"]: print(f"  Grade {g} (×{GRADE_WEIGHT[g]}): {sg.get(g,0)}")
     print(f"  🔒 MetaCog: COMING SOON\n{'='*60}\n")