Rebuild Space with Leaderboard / Instructions / Submit tabs and verified CLI commands

README.md

@@ -2,9 +2,9 @@
 title: TeamBench Leaderboard
 emoji: 📊
 colorFrom: blue
-colorTo: green
+colorTo: indigo
 sdk: gradio
-sdk_version: "5.12.0"
+sdk_version: 5.12.0
 python_version: "3.12"
 app_file: app.py
 pinned: true
@@ -13,14 +13,7 @@ license: mit
 
 # TeamBench Leaderboard
 
-Multi-agent benchmark evaluating LLM teamwork with OS-enforced role separation.
+Multi-agent LLM coordination on 90 stratified tasks under OS-enforced Planner / Executor / Verifier role separation. Three tabs: **Leaderboard** (current results), **Instructions** (install + run + grade), **Submit** (upload your results JSON).
 
-## Running locally
-
-```bash
-cd leaderboard
-pip install -r requirements.txt
-python app.py
-```
-
-The app loads data from `data/cross_model_stats.json`.
+- Code: <https://github.com/ybkim95/TeamBench>
+- Dataset: <https://huggingface.co/datasets/ybkim95/teambench>

app.py

@@ -1,803 +1,226 @@
-"""
-TeamBench Leaderboard — HuggingFace Spaces Gradio Application
+"""TeamBench Leaderboard — Hugging Face Spaces app
 
-Evaluates LLM teamwork via OS-enforced role separation (Planner/Executor/Verifier).
-Primary metric: TNI (Teamwork Necessity Index).
+Three tabs:
+  - Leaderboard  (paper Table 4 / LB90, 13 models, 5 conditions)
+  - Instructions (install, run an evaluation, compute TNI)
+  - Submit       (model name + results JSON upload, written to submissions/)
 """
-
 from __future__ import annotations
-
 import json
 import os
-import zipfile
-import tempfile
+import re
+import time
 from pathlib import Path
 
 import gradio as gr
 import pandas as pd
 
-# ---------------------------------------------------------------------------
-# Data loading
-# ---------------------------------------------------------------------------
-
-DATA_DIR = Path(__file__).parent / "data"
-SUBMISSIONS_DIR = Path(__file__).parent / "submissions"
+ROOT = Path(__file__).parent
+SUBMISSIONS_DIR = ROOT / "submissions"
 SUBMISSIONS_DIR.mkdir(exist_ok=True)
 
-_data: dict = {}
-
-
-def _load() -> dict:
-    global _data
-    if _data:
-        return _data
-    path = DATA_DIR / "leaderboard_data.json"
-    with open(path) as f:
-        _data = json.load(f)
-    return _data
-
-
-# ---------------------------------------------------------------------------
-# Display name maps
-# ---------------------------------------------------------------------------
-
-MODEL_DISPLAY = {
-    "gemini-3-flash-preview": "Gemini 3 Flash",
-    "gemini-3.1-flash-lite-preview": "Gemini 3.1 Flash Lite",
-    "gpt-5-mini": "GPT-5 Mini",
-    "gpt-5-nano": "GPT-5 Nano",
-}
-
-# ---------------------------------------------------------------------------
-# CSS
-# ---------------------------------------------------------------------------
-
-CUSTOM_CSS = """
-:root {
-    --bench-blue: #2563eb; --bench-green: #16a34a; --bench-red: #dc2626;
-    --bench-amber: #d97706; --bench-purple: #7c3aed; --bench-gray: #6b7280;
-    --badge-radius: 4px;
-}
-
-.tb-header {
-    background: linear-gradient(135deg, #1e3a8a 0%, #1d4ed8 50%, #2563eb 100%);
-    border-radius: 12px; padding: 28px 32px 24px; margin-bottom: 8px;
-}
-.tb-header h1 {
-    margin: 0 0 6px; font-size: 2rem; font-weight: 700;
-    letter-spacing: -0.5px; color: #ffffff !important;
-}
-.tb-header p {
-    margin: 0; font-size: 1rem; color: #ffffff !important; opacity: 0.92;
-}
-
-.stat-grid {
-    display: grid; grid-template-columns: repeat(auto-fit, minmax(150px, 1fr));
-    gap: 12px; margin-bottom: 16px;
-}
-.stat-card {
-    background: var(--background-fill-primary, #fff);
-    border: 1px solid var(--border-color-primary, #e5e7eb);
-    border-radius: 10px; padding: 16px 18px; text-align: center;
-}
-.stat-card .stat-value { font-size: 1.5rem; font-weight: 700; color: var(--bench-blue); line-height: 1; }
-.stat-card .stat-label { font-size: 0.78rem; color: var(--bench-gray); margin-top: 4px; }
-
-.badge-HIGH-TNI   { background:#dbeafe; color:#1e40af; padding:2px 8px; border-radius:var(--badge-radius); font-size:0.78rem; font-weight:600; }
-.badge-TEAM-HELPS { background:#dcfce7; color:#166534; padding:2px 8px; border-radius:var(--badge-radius); font-size:0.78rem; font-weight:600; }
-.badge-NEUTRAL    { background:#f3f4f6; color:#374151; padding:2px 8px; border-radius:var(--badge-radius); font-size:0.78rem; font-weight:600; }
-.badge-TEAM-HURTS { background:#fee2e2; color:#991b1b; padding:2px 8px; border-radius:var(--badge-radius); font-size:0.78rem; font-weight:600; }
-
-.section-title {
-    font-size: 1.05rem; font-weight: 600; margin: 20px 0 10px;
-    padding-bottom: 6px; border-bottom: 2px solid var(--border-color-primary, #e5e7eb);
-}
-
-.about-card {
-    background: var(--background-fill-primary, #fff);
-    border: 1px solid var(--border-color-primary, #e5e7eb);
-    border-radius: 10px; padding: 20px 24px; margin-bottom: 12px;
-}
-.about-card h3 { margin: 0 0 10px; font-size: 1rem; font-weight: 700; }
-.about-card p, .about-card li { font-size: 0.9rem; line-height: 1.65; }
-.about-card ul { padding-left: 20px; margin: 8px 0; }
-
-.role-planner  { background:#ede9fe; color:#5b21b6; padding:3px 10px; border-radius:20px; font-size:0.82rem; font-weight:600; }
-.role-executor { background:#fef3c7; color:#92400e; padding:3px 10px; border-radius:20px; font-size:0.82rem; font-weight:600; }
-.role-verifier { background:#d1fae5; color:#065f46; padding:3px 10px; border-radius:20px; font-size:0.82rem; font-weight:600; }
-
-.submit-info {
-    background: #eff6ff; border: 1px solid #bfdbfe; border-radius: 10px;
-    padding: 18px 22px; margin-bottom: 16px; font-size: 0.9rem; line-height: 1.7;
-}
-.submit-info code {
-    background: #dbeafe; padding: 2px 6px; border-radius: 4px; font-size: 0.82rem;
-}
-.submit-step {
-    background: var(--background-fill-primary, #fff);
-    border: 1px solid var(--border-color-primary, #e5e7eb);
-    border-radius: 10px; padding: 16px 20px; margin-bottom: 12px;
-}
-.submit-step h4 { margin: 0 0 8px; font-size: 0.95rem; }
-
-.gr-dataframe table { width: 100%; font-size: 0.875rem; }
-.gr-dataframe th { font-weight: 600; }
-"""
-
-# ---------------------------------------------------------------------------
-# HTML builders
-# ---------------------------------------------------------------------------
-
-def create_header() -> str:
-    return """
-<div class="tb-header">
-  <h1>TeamBench Leaderboard</h1>
-  <p>Multi-agent benchmark evaluating LLM teamwork with OS-enforced role separation</p>
-</div>
-"""
-
-
-def create_stat_cards() -> str:
-    d = _load()
-    agg = d["aggregate"]
-    n = d["total_tasks"]
-    n_hard = d["total_hard"]
-    cats = len(d["categories"])
-    helps = agg["team_helps_count"]
-    high_tni = agg["high_tni_count"]
-    avg_tni = agg["avg_tni"]
-    return f"""
-<div class="stat-grid">
-  <div class="stat-card">
-    <div class="stat-value">{n}</div>
-    <div class="stat-label">Total Tasks</div>
-  </div>
-  <div class="stat-card">
-    <div class="stat-value">{n_hard}</div>
-    <div class="stat-label">Hard / Expert Tasks</div>
-  </div>
-  <div class="stat-card">
-    <div class="stat-value">{cats}</div>
-    <div class="stat-label">Categories</div>
-  </div>
-  <div class="stat-card">
-    <div class="stat-value">{helps}</div>
-    <div class="stat-label">Team-Helps Tasks</div>
-  </div>
-  <div class="stat-card">
-    <div class="stat-value">{high_tni}</div>
-    <div class="stat-label">High-TNI (team &ge; oracle)</div>
-  </div>
-  <div class="stat-card">
-    <div class="stat-value">{avg_tni:.2f}</div>
-    <div class="stat-label">Avg TNI</div>
-  </div>
-</div>
-"""
-
-
 # ---------------------------------------------------------------------------
-# Tab: Performance — Cross-Model (28 tasks, 4 models)
+# LB90 leaderboard from paper Table tab:lb90-leaderboard
+# Pass-rate (%) per condition. Sort: max(Solo, Full) desc.
 # ---------------------------------------------------------------------------
+LEADERBOARD = [
+    {"model": "Claude Opus 4.7",       "provider": "Anthropic", "Solo": 35.6, "Restricted": 33.3, "No Plan": 35.6, "No Eval": 33.3, "Full": 37.8},
+    {"model": "GPT-5.4 Mini",          "provider": "OpenAI",    "Solo": 33.3, "Restricted": 23.3, "No Plan": 25.6, "No Eval": 24.4, "Full": 28.9},
+    {"model": "Claude Haiku 4.5",      "provider": "Anthropic", "Solo": 12.2, "Restricted": 31.1, "No Plan": 18.9, "No Eval":  1.1, "Full": 28.9},
+    {"model": "Gemini-3.1 Pro",        "provider": "Google",    "Solo": 27.8, "Restricted": 22.2, "No Plan": 16.7, "No Eval": 25.6, "Full": 28.9},
+    {"model": "Claude Sonnet 4.6",     "provider": "Anthropic", "Solo":  7.8, "Restricted": 27.8, "No Plan": 10.0, "No Eval":  6.7, "Full": 27.8},
+    {"model": "GPT-5.4",               "provider": "OpenAI",    "Solo": 12.2, "Restricted": 35.6, "No Plan": 23.3, "No Eval": 34.4, "Full": 27.8},
+    {"model": "Gemma 4 31B",           "provider": "Google",    "Solo": 27.8, "Restricted": 25.6, "No Plan": 24.4, "No Eval": 20.0, "Full": 22.2},
+    {"model": "Gemini-3 Flash",        "provider": "Google",    "Solo": 13.3, "Restricted": 18.9, "No Plan": 14.4, "No Eval": 27.8, "Full": 25.6},
+    {"model": "Gemini-3.1 Flash Lite", "provider": "Google",    "Solo":  5.6, "Restricted": 21.1, "No Plan":  8.9, "No Eval": 17.8, "Full": 17.8},
+    {"model": "gpt-oss-20b",           "provider": "OpenAI",    "Solo": 17.8, "Restricted": 17.8, "No Plan": 12.2, "No Eval":  7.8, "Full":  2.2},
+    {"model": "Qwen 3 14B",            "provider": "Alibaba",   "Solo":  5.6, "Restricted":  2.2, "No Plan":  2.2, "No Eval":  1.1, "Full":  2.2},
+    {"model": "Qwen 3 32B",            "provider": "Alibaba",   "Solo":  5.6, "Restricted":  3.3, "No Plan":  0.0, "No Eval":  5.6, "Full":  1.1},
+    {"model": "Qwen 3 8B",             "provider": "Alibaba",   "Solo":  2.2, "Restricted":  5.6, "No Plan":  1.1, "No Eval":  3.3, "Full":  0.0},
+]
 
-def build_crossmodel_leaderboard() -> pd.DataFrame:
-    d = _load()
-    cm = d["cross_model"]
-    models_ordered = sorted(cm["per_model"].items(),
-                            key=lambda kv: kv[1]["avg_uplift"], reverse=True)
-    rows = []
-    for i, (mid, stats) in enumerate(models_ordered):
-        u = stats["avg_uplift"]
-        rows.append({
-            "Rank": f"#{i+1}",
-            "Model": MODEL_DISPLAY.get(mid, mid),
-            "Team Score": f"{stats['avg_team']:.1%}",
-            "Oracle Score": f"{stats['avg_oracle']:.1%}",
-            "Team Uplift": f"+{u:.1%}" if u >= 0 else f"{u:.1%}",
-            "Avg TNI": f"{stats['avg_tni']:.3f}",
-            "Team > Oracle": f"{stats['team_helps_pct']:.0%}",
-        })
-    return pd.DataFrame(rows)
-
-
-def build_crossmodel_task_df(category: str = "All",
-                             classification: str = "All") -> pd.DataFrame:
-    d = _load()
-    cm = d["cross_model"]
-    models = cm["models"]
-
-    rows = []
-    for task in cm["per_task"]:
-        cat = task["category"]
-        if category != "All" and cat != category:
-            continue
-
-        classes = [task["models"][m]["classification"] for m in models if m in task["models"]]
-        counts: dict[str, int] = {}
-        for c in classes:
-            counts[c] = counts.get(c, 0) + 1
-        agg_class = max(counts, key=counts.get) if counts else "NEUTRAL"
-        if classification != "All" and agg_class != classification:
-            continue
-
-        row: dict = {"Task": task["task_id"], "Category": cat}
-        for mid in models:
-            short = MODEL_DISPLAY.get(mid, mid)
-            mdata = task["models"].get(mid, {})
-            team = mdata.get("team", 0) or 0
-            uplift = mdata.get("team_uplift", 0) or 0
-            row[f"{short}"] = f"{team:.0%}"
-            row[f"{short} Uplift"] = f"+{uplift:.0%}" if uplift >= 0 else f"{uplift:.0%}"
-        row["Class"] = agg_class
-        rows.append(row)
-
-    if not rows:
-        return pd.DataFrame()
-    return pd.DataFrame(rows).sort_values("Task").reset_index(drop=True)
 
-
-# ---------------------------------------------------------------------------
-# Tab: Overview — All 155 tasks (reference model)
-# ---------------------------------------------------------------------------
-
-def build_overview_df(track: str = "Full", category: str = "All",
-                      classification: str = "All") -> pd.DataFrame:
-    d = _load()
-    rows = []
-    for t in d["per_task"]:
-        if track == "Hard" and t["difficulty_track"] != "hard":
-            continue
-        if category != "All" and t["category"] != category:
-            continue
-        if classification != "All" and t["classification"] != classification:
-            continue
-        uplift = t["team_uplift"]
-        rows.append({
-            "Task": t["task_id"],
-            "Category": t["category"],
-            "Difficulty": t["difficulty"].capitalize(),
-            "Oracle": f"{t['oracle']:.0%}",
-            "Restricted": f"{t['restricted']:.0%}",
-            "Team": f"{t['team']:.0%}",
-            "Uplift": f"+{uplift:.0%}" if uplift >= 0 else f"{uplift:.0%}",
-            "TNI": f"{t['tni']:.2f}" if t["tni"] is not None else "—",
-            "Class": t["classification"],
-        })
-    if not rows:
-        return pd.DataFrame()
-    return pd.DataFrame(rows).sort_values("Task").reset_index(drop=True)
-
-
-def build_category_summary_df(track: str = "Full") -> pd.DataFrame:
-    d = _load()
-    cat_data: dict[str, list] = {}
-    for t in d["per_task"]:
-        if track == "Hard" and t["difficulty_track"] != "hard":
-            continue
-        cat = t["category"]
-        if cat not in cat_data:
-            cat_data[cat] = []
-        cat_data[cat].append(t)
-
-    rows = []
-    for cat in sorted(cat_data):
-        tasks = cat_data[cat]
-        n = len(tasks)
-        avg_team = sum(t["team"] for t in tasks) / n
-        avg_oracle = sum(t["oracle"] for t in tasks) / n
-        avg_uplift = sum(t["team_uplift"] for t in tasks) / n
-        helps = sum(1 for t in tasks if t["team_uplift"] > 0.05)
-        high = sum(1 for t in tasks if t["classification"] == "HIGH-TNI")
-        rows.append({
-            "Category": cat,
-            "Tasks": n,
-            "Avg Team": f"{avg_team:.0%}",
-            "Avg Oracle": f"{avg_oracle:.0%}",
-            "Avg Uplift": f"+{avg_uplift:.0%}" if avg_uplift >= 0 else f"{avg_uplift:.0%}",
-            "Team Helps": f"{helps}/{n}",
-            "High-TNI": high,
-        })
-    return pd.DataFrame(rows).sort_values("Category").reset_index(drop=True)
+def build_leaderboard_df():
+    df = pd.DataFrame(LEADERBOARD)
+    df.insert(0, "#", range(1, len(df) + 1))
+    return df
 
 
 # ---------------------------------------------------------------------------
-# Submission handler — workspace artifact zip (server-graded)
+# Submission validator
 # ---------------------------------------------------------------------------
-
-MAX_UPLOAD_MB = 100
+SAFE_NAME = re.compile(r"^[A-Za-z0-9_.\-]+$")
 
 
-def _get_valid_task_ids() -> set[str]:
-    return {t["task_id"] for t in _load()["per_task"]}
-
-
-def validate_and_accept_submission(
-    file, model_name: str, team_name: str, framework: str,
-    contact: str, description: str, seed: str,
-) -> str:
-    """Validate a workspace artifact zip and queue it for server-side grading."""
-
-    # --- basic field validation ---
-    if not file:
-        return "Please upload a workspace artifact zip file."
-    if not model_name or not model_name.strip():
-        return "Please provide a model name."
-    if not team_name or not team_name.strip():
-        return "Please provide a team or organization name."
-
-    file_path = file.name if hasattr(file, "name") else str(file)
-
-    # --- file size check ---
-    file_size_mb = os.path.getsize(file_path) / (1024 * 1024)
-    if file_size_mb > MAX_UPLOAD_MB:
-        return f"File too large ({file_size_mb:.1f} MB). Maximum is {MAX_UPLOAD_MB} MB."
-
-    # --- zip validation ---
-    if not zipfile.is_zipfile(file_path):
-        return "Uploaded file is not a valid zip archive."
-
-    valid_tasks = _get_valid_task_ids()
-    found_tasks: dict[str, list[str]] = {}  # task_id -> list of files
-    has_meta = False
+def validate_submission(file_obj, model: str, team: str, contact: str, framework: str, description: str):
+    if file_obj is None:
+        return "**Error.** Please attach a JSON results file produced by `harness.run_all` or `harness.ablation`."
+    if not model or not SAFE_NAME.match(model):
+        return "**Error.** Model name is required and may only contain letters, digits, `.`, `_`, `-`."
+    if not team:
+        return "**Error.** Team / organization is required."
 
+    src = Path(file_obj.name if hasattr(file_obj, "name") else file_obj)
     try:
-        with zipfile.ZipFile(file_path, "r") as zf:
-            names = zf.namelist()
-
-            # Security: reject paths with .. or absolute paths
-            for name in names:
-                if ".." in name or name.startswith("/"):
-                    return f"Rejected: unsafe path in zip: `{name}`"
-
-            # Check for meta.json
-            for name in names:
-                if name.endswith("meta.json") and name.count("/") <= 1:
-                    has_meta = True
-
-            # Find workspace directories. Accept multiple layouts:
-            #   tasks/{TASK_ID}/workspace/...
-            #   tasks/{TASK_ID}/{condition}/workspace/...
-            #   {TASK_ID}/workspace/...
-            #   {TASK_ID}/{condition}/workspace/...
-            valid_upper = {t.upper(): t for t in valid_tasks}
-
-            for name in names:
-                parts = name.split("/")
-                if "workspace" not in parts:
-                    continue
-                ws_idx = parts.index("workspace")
-                # Try task_id at ws_idx-1 or ws_idx-2
-                for offset in [1, 2]:
-                    candidate_idx = ws_idx - offset
-                    if candidate_idx < 0:
-                        continue
-                    candidate = parts[candidate_idx]
-                    matched = (
-                        candidate if candidate in valid_tasks
-                        else valid_upper.get(candidate.upper())
-                    )
-                    if matched:
-                        if matched not in found_tasks:
-                            found_tasks[matched] = []
-                        found_tasks[matched].append(name)
-                        break
-
-    except zipfile.BadZipFile:
-        return "Corrupted zip file. Please re-create and upload again."
-
-    if not found_tasks:
-        return (
-            "No valid task workspaces found in the zip.\n\n"
-            "Expected structure:\n"
-            "```\n"
-            "submission.zip/\n"
-            "  meta.json\n"
-            "  tasks/\n"
-            "    S1_hidden_spec/\n"
-            "      workspace/\n"
-            "        <files your agent produced>\n"
-            "    D1_schema_drift/\n"
-            "      workspace/\n"
-            "        ...\n"
-            "```"
-        )
-
-    # --- save submission ---
-    safe_model = "".join(c if c.isalnum() or c in "-_" else "_" for c in model_name.strip())
-    safe_team = "".join(c if c.isalnum() or c in "-_" else "_" for c in team_name.strip())
-    submission_id = f"{safe_team}__{safe_model}"
-
-    # Save metadata
+        with open(src) as f:
+            payload = json.load(f)
+    except Exception as e:
+        return f"**Error.** Could not parse JSON: `{e}`"
+
+    if not isinstance(payload, (list, dict)):
+        return "**Error.** Results JSON must be a list of run records or a `{tasks: [...], conditions: {...}}` object."
+
+    ts = time.strftime("%Y%m%dT%H%M%SZ", time.gmtime())
+    out_dir = SUBMISSIONS_DIR / f"{model}-{ts}"
+    out_dir.mkdir(parents=True, exist_ok=True)
+    (out_dir / "results.json").write_text(json.dumps(payload, indent=2))
     meta = {
-        "submission_id": submission_id,
-        "model": model_name.strip(),
-        "team": team_name.strip(),
-        "framework": framework.strip() if framework else "",
-        "contact": contact.strip() if contact else "",
-        "description": description.strip() if description else "",
-        "seed": int(seed) if seed and seed.isdigit() else 0,
-        "tasks_submitted": sorted(found_tasks.keys()),
-        "task_count": len(found_tasks),
-        "file_count": sum(len(v) for v in found_tasks.values()),
-        "status": "pending_grading",
+        "submitted_at_utc": ts,
+        "model": model,
+        "team": team,
+        "framework": framework or None,
+        "contact": contact or None,
+        "description": description or None,
+        "filename": src.name,
+        "n_records": len(payload) if isinstance(payload, list) else None,
     }
-
-    meta_path = SUBMISSIONS_DIR / f"{submission_id}.json"
-    with open(meta_path, "w") as f:
-        json.dump(meta, f, indent=2)
-
-    # Save the zip itself
-    import shutil
-    zip_path = SUBMISSIONS_DIR / f"{submission_id}.zip"
-    shutil.copy2(file_path, zip_path)
-
-    # --- build response ---
-    missing = valid_tasks - set(found_tasks.keys())
-    conditions = ["oracle", "restricted", "full", "team_no_plan", "team_no_verify"]
-
-    response_parts = [
-        f"### Submission Received",
-        f"",
-        f"**Model**: {model_name.strip()}",
-        f"**Team**: {team_name.strip()}",
-        f"**Tasks**: {len(found_tasks)} / {len(valid_tasks)}",
-        f"**Files**: {sum(len(v) for v in found_tasks.values())} workspace artifacts",
-        f"**Seed**: {meta['seed']}",
-    ]
-
-    if framework and framework.strip():
-        response_parts.append(f"**Framework**: {framework.strip()}")
-
-    response_parts.append("")
-
-    if len(found_tasks) < len(valid_tasks):
-        response_parts.append(
-            f"*{len(missing)} tasks not included (partial submission accepted).*"
-        )
-
-    response_parts += [
-        "",
-        "**Status: Queued for server-side grading.**",
-        "",
-        "Your workspace artifacts will be graded against our hidden test suites. "
-        "Scores are computed server-side — not self-reported — to ensure benchmark integrity.",
-        "",
-        "Results will appear on the leaderboard once grading completes (typically within 24-48 hours). "
-        "You will be contacted at the provided email if there are any issues.",
-    ]
-
-    return "\n".join(response_parts)
+    (out_dir / "meta.json").write_text(json.dumps(meta, indent=2))
+    return (
+        f"**Submission accepted** at `submissions/{out_dir.name}/`. We will re-run the "
+        f"deterministic graders server-side and add the model to the leaderboard once "
+        f"verified (typical turnaround: 24-48h). For status, open an issue at "
+        f"https://github.com/ybkim95/TeamBench/issues."
+    )
 
 
 # ---------------------------------------------------------------------------
-# About HTML
+# UI
 # ---------------------------------------------------------------------------
-
-def create_about_html() -> str:
-    return """
-<div class="about-card">
-  <h3>What is TeamBench?</h3>
-  <p>TeamBench is a multi-agent benchmark that evaluates whether structured LLM teamwork
-  improves task performance beyond what a single all-capable (oracle) agent can achieve.
-  155 tasks span 19 categories including Security, Data Engineering, Distributed Systems,
-  Testing, and Adversarial Specification &mdash; designed to require coordination, verification,
-  and planning across roles.</p>
-  <p>All roles execute in isolated Docker containers with OS-enforced permission boundaries,
-  preventing role confusion and ensuring genuine separation of concerns.</p>
-</div>
-
-<div class="about-card">
-  <h3>The Three Roles</h3>
-  <ul>
-    <li><span class="role-planner">Planner</span> &mdash; reads code and documentation, identifies what needs to change,
-    produces a structured plan. Has read-only filesystem access plus static analysis tools.</li>
-    <li><span class="role-executor">Executor</span> &mdash; implements the plan. Has read-write access to the workspace
-    but cannot run tests or arbitrary shell commands.</li>
-    <li><span class="role-verifier">Verifier</span> &mdash; runs tests and validates the implementation. Has pytest,
-    hypothesis, and mutation testing but cannot modify source files.</li>
-  </ul>
-</div>
-
-<div class="about-card">
-  <h3>Ablation Conditions</h3>
-  <ul>
-    <li><strong>Oracle</strong> &mdash; single agent with full unrestricted access (upper bound).</li>
-    <li><strong>Restricted</strong> &mdash; single agent with only executor-level permissions (lower bound).</li>
-    <li><strong>Full Team</strong> &mdash; all three roles: Planner &rarr; Executor &rarr; Verifier.</li>
-    <li><strong>Team (No Plan)</strong> &mdash; Executor + Verifier only; measures planning value.</li>
-    <li><strong>Team (No Verify)</strong> &mdash; Planner + Executor only; measures verification value.</li>
-  </ul>
-</div>
-
-<div class="about-card">
-  <h3>TNI &mdash; Teamwork Necessity Index</h3>
-  <p style="font-family:monospace; background:#f3f4f6; padding:8px 12px; border-radius:6px; display:inline-block;">
-    TNI = (team &minus; restricted) / (oracle &minus; restricted)
-  </p>
-  <p>TNI = 1.0 means the team matches the oracle. TNI &gt; 1.0 means the team <em>exceeds</em>
-  the oracle &mdash; true super-additive teamwork.</p>
-  <ul>
-    <li><span class="badge-HIGH-TNI">HIGH-TNI</span> TNI &ge; 1.0: team closes or exceeds the oracle gap.</li>
-    <li><span class="badge-TEAM-HELPS">TEAM-HELPS</span> positive uplift, TNI &lt; 1.0.</li>
-    <li><span class="badge-NEUTRAL">NEUTRAL</span> no significant change.</li>
-    <li><span class="badge-TEAM-HURTS">TEAM-HURTS</span> coordination degrades performance.</li>
-  </ul>
-</div>
-
-<div class="about-card">
-  <h3>Difficulty Tracks</h3>
-  <ul>
-    <li><strong>Full</strong> &mdash; all 155 tasks (easy + medium + hard + expert).</li>
-    <li><strong>Hard</strong> &mdash; 122 tasks rated hard or expert only.</li>
-  </ul>
-</div>
-
-<div class="about-card">
-  <h3>Citation</h3>
-  <p>If you use TeamBench in your research, please cite:</p>
-  <pre style="background:#f3f4f6; padding:12px 16px; border-radius:8px; font-size:0.82rem; line-height:1.6; overflow-x:auto;">@misc{kim2026teambench,
-  title   = {TeamBench: Evaluating Structured LLM Teamwork
-             via OS-Enforced Role Separation},
-  author  = {Kim, Yubin},
-  year    = {2026},
-  url     = {https://github.com/ybkim95/TeamBench}
-}</pre>
-</div>
-
-<div class="about-card">
-  <h3>Links</h3>
-  <ul>
-    <li>Repository: <a href="https://github.com/ybkim95/TeamBench" target="_blank">github.com/ybkim95/TeamBench</a></li>
-    <li>Leaderboard: <a href="https://huggingface.co/spaces/ybkim95/teambench-leaderboard" target="_blank">huggingface.co/spaces/ybkim95/teambench-leaderboard</a></li>
-  </ul>
-</div>
+CSS = """
+.tb-hero { padding: 1.25rem 1.5rem; border-radius: 12px; background: linear-gradient(135deg,#1e3a8a 0%,#1d4ed8 50%,#2563eb 100%); color: #fff; margin-bottom: 1rem; }
+.tb-hero h1 { color: #fff !important; margin: 0 0 0.35rem; font-weight: 700; }
+.tb-hero p  { color: rgba(255,255,255,0.95) !important; margin: 0; }
 """
 
 
-# ---------------------------------------------------------------------------
-# Main app
-# ---------------------------------------------------------------------------
-
-def build_app() -> gr.Blocks:
-    _load()
-
-    categories_all = ["All"] + _load()["categories"]
-    classifications = ["All", "HIGH-TNI", "TEAM-HELPS", "NEUTRAL", "TEAM-HURTS"]
-    cm_categories = ["All"] + sorted({
-        t["category"] for t in _load()["cross_model"]["per_task"]
-    })
-
-    _theme = gr.themes.Default(
-        primary_hue="blue",
-        secondary_hue="emerald",
-        font=[gr.themes.GoogleFont("Inter"), "ui-sans-serif", "sans-serif"],
-    )
-
-    with gr.Blocks(title="TeamBench Leaderboard", theme=_theme, css=CUSTOM_CSS) as demo:
-        gr.HTML(create_header())
+def build_app():
+    with gr.Blocks(css=CSS, title="TeamBench Leaderboard") as demo:
+        gr.HTML(
+            """
+            <div class="tb-hero">
+              <h1>TeamBench Leaderboard</h1>
+              <p>Multi-agent LLM coordination on 90 stratified tasks under OS-enforced Planner / Executor / Verifier role separation. Five conditions per task, deterministic graders, MIT-licensed.</p>
+            </div>
+            """
+        )
 
         with gr.Tabs():
-
-            # ==============================================================
-            # TAB 1: Performance (Cross-Model) — default tab
-            # ==============================================================
-            with gr.Tab("Performance"):
-                gr.HTML(create_stat_cards())
-
+            with gr.Tab("Leaderboard"):
                 gr.Markdown(
-                    "### Cross-Model Leaderboard\n"
-                    "28 tasks evaluated across 4 models. "
-                    "Ranked by average team uplift over oracle."
-                )
-                gr.Dataframe(
-                    value=build_crossmodel_leaderboard(),
-                    interactive=False, wrap=False,
-                    column_widths=["70px", "180px", "110px", "110px", "110px", "90px", "120px"],
+                    "Pass rate (%) per condition on **TeamBench-90**. Models are sorted by `max(Solo, Full)`. "
+                    "Source: paper main results (TeamBench-Verified covers 57 of these 90 tasks)."
                 )
-
-                gr.Markdown("### Per-Task Breakdown (28 tasks)")
-                with gr.Row():
-                    cm_cat_dd = gr.Dropdown(
-                        choices=cm_categories, value="All",
-                        label="Category", scale=2,
-                    )
-                    cm_cls_dd = gr.Dropdown(
-                        choices=classifications, value="All",
-                        label="Classification", scale=2,
-                    )
-
-                cm_task_table = gr.Dataframe(
-                    value=build_crossmodel_task_df(),
-                    interactive=False, wrap=False,
-                )
-
-                def _update_cm_tasks(cat, cls):
-                    return build_crossmodel_task_df(cat, cls)
-
-                for w in [cm_cat_dd, cm_cls_dd]:
-                    w.change(fn=_update_cm_tasks,
-                             inputs=[cm_cat_dd, cm_cls_dd],
-                             outputs=cm_task_table)
-
-            # ==============================================================
-            # TAB 2: About
-            # ==============================================================
-            with gr.Tab("About"):
-                gr.HTML(create_about_html())
-
-            # ==============================================================
-            # TAB 3: Overview (all 155 tasks)
-            # ==============================================================
-            with gr.Tab("Overview"):
+                gr.Dataframe(value=build_leaderboard_df(), interactive=False, wrap=False)
                 gr.Markdown(
-                    "### All 155 Tasks\n"
-                    "Evaluated across 5 ablation conditions at seed 0. "
-                    "Use **Track** to switch between Full (all) and Hard (hard + expert only)."
-                )
-
-                with gr.Row():
-                    track_dd = gr.Dropdown(
-                        choices=["Full", "Hard"],
-                        value="Full", label="Track", scale=1,
-                    )
-                    cat_dd = gr.Dropdown(
-                        choices=categories_all,
-                        value="All", label="Category", scale=2,
-                    )
-                    class_dd = gr.Dropdown(
-                        choices=classifications,
-                        value="All", label="Classification", scale=2,
-                    )
-
-                task_table = gr.Dataframe(
-                    value=build_overview_df(), interactive=False, wrap=False,
+                    "**Five conditions** &nbsp; "
+                    "`Solo` (one agent, full access) · "
+                    "`Restricted` (one agent, executor tools only) · "
+                    "`No Plan` (Executor + Verifier; Verifier holds the spec) · "
+                    "`No Eval` (Planner + Executor) · "
+                    "`Full` (Planner + Executor + Verifier)."
                 )
 
-                def _update_overview(track, cat, cls):
-                    return build_overview_df(track, cat, cls)
-
-                for w in [track_dd, cat_dd, class_dd]:
-                    w.change(fn=_update_overview,
-                             inputs=[track_dd, cat_dd, class_dd],
-                             outputs=task_table)
-
-                gr.Markdown("### Per-Category Summary")
-                cat_summary = gr.Dataframe(
-                    value=build_category_summary_df(), interactive=False, wrap=False,
+            with gr.Tab("Instructions"):
+                gr.Markdown(
+                    """
+### 1. Install
+
+```bash
+git clone https://github.com/ybkim95/TeamBench.git
+cd TeamBench
+pip install -e ".[all]"
+docker compose build
+```
+
+Set the providers you want to evaluate:
+
+```bash
+export ANTHROPIC_API_KEY=...
+export OPENAI_API_KEY=...
+export GEMINI_API_KEY=...
+```
+
+### 2. Run a single task across all five conditions
+
+```bash
+python -m harness.ablation \\
+    --model gemini-3-flash-preview \\
+    --tasks DIST1_queue_race \\
+    --seeds 0 \\
+    --conditions oracle restricted full team_no_plan team_no_verify \\
+    --output shared/runs/example
+```
+
+Each `score.json` under `shared/runs/example/<task>/<condition>/` contains `passed: true|false` and a partial score in `[0, 1]` from the deterministic shell-script grader. No API keys handy? `--model mock` runs an in-process stub that exercises the grader pipeline without provider calls.
+
+### 3. Run the full TeamBench-90 sweep
+
+```bash
+python -m harness.run_all \\
+    --tasks $(jq -r '.[]' leaderboard/data/leaderboard_90_tasks.json) \\
+    --seeds 0 \\
+    --runs_dir shared/runs/lb90_<model> \\
+    --output shared/ablation_results/lb90_<model>_seed0.json
+```
+
+### 4. Compute TNI and the paper tables
+
+```bash
+python -m harness.compute_tni \\
+    --ablation shared/ablation_results/lb90_<model>_seed0.json \\
+    --output shared/ablation_results/tni_<model>.json
+
+python -m harness.paper_tables \\
+    --ablation shared/ablation_results/lb90_<model>_seed0.json \\
+    --output-dir shared/paper/
+```
+
+### 5. Submit
+
+Upload `shared/ablation_results/lb90_<model>_seed0.json` from the Submit tab. We re-run the deterministic graders server-side and add your model to the leaderboard once verified.
+                    """
                 )
 
-                def _update_cat_summary(track):
-                    return build_category_summary_df(track)
-
-                track_dd.change(fn=_update_cat_summary,
-                                inputs=[track_dd], outputs=cat_summary)
-
-            # ==============================================================
-            # TAB 4: Submit
-            # ==============================================================
             with gr.Tab("Submit"):
-                gr.HTML("""
-<div class="submit-info">
-  <h3 style="margin:0 0 12px;">Submit Your Results</h3>
-  <p>TeamBench uses <strong>server-side grading</strong> to ensure benchmark integrity.
-  You submit workspace artifacts (the files your agent produced) &mdash;
-  not self-reported scores. We grade them against hidden test suites.</p>
-</div>
-""")
-
-                gr.HTML("""
-<div class="submit-step">
-  <h4>Step 1: Install and Run</h4>
-  <pre style="background:#f3f4f6; padding:10px 14px; border-radius:6px; font-size:0.82rem; overflow-x:auto; margin:8px 0;">pip install teambench
-teambench run --model your-model --seed 0 --output submission/</pre>
-  <p style="font-size:0.85rem; color:#6b7280; margin:4px 0 0;">
-    This runs your agent on all tasks and saves the workspace artifacts (files your agent wrote).
-    You must run all 5 conditions: <code>oracle</code>, <code>restricted</code>, <code>full</code>,
-    <code>team_no_plan</code>, <code>team_no_verify</code>.
-  </p>
-</div>
-
-<div class="submit-step">
-  <h4>Step 2: Package</h4>
-  <pre style="background:#f3f4f6; padding:10px 14px; border-radius:6px; font-size:0.82rem; overflow-x:auto; margin:8px 0;">teambench package submission/ --output submission.zip</pre>
-  <p style="font-size:0.85rem; color:#6b7280; margin:4px 0 0;">
-    Creates a zip with the required structure. Partial submissions (subset of 155 tasks) are accepted.
-  </p>
-</div>
-
-<div class="submit-step">
-  <h4>Step 3: Upload</h4>
-  <p style="font-size:0.85rem; color:#6b7280;">
-    Fill in the fields below and upload your <code>submission.zip</code>.
-    Results appear on the leaderboard after server-side grading (24-48 hours).
-  </p>
-</div>
-""")
-
-                gr.HTML("""
-<details style="margin-bottom:16px;">
-  <summary style="cursor:pointer; font-weight:600; font-size:0.9rem; padding:8px 0;">
-    Expected zip structure
-  </summary>
-  <pre style="background:#f3f4f6; padding:12px 16px; border-radius:6px; font-size:0.82rem; margin-top:8px;">submission.zip/
-  meta.json                        # auto-generated by teambench package
-  tasks/
-    S1_hidden_spec/
-      oracle/workspace/            # workspace after oracle condition
-      restricted/workspace/        # workspace after restricted condition
-      full/workspace/              # workspace after full team condition
-      team_no_plan/workspace/      # workspace after no-plan condition
-      team_no_verify/workspace/    # workspace after no-verify condition
-    D1_schema_drift/
-      oracle/workspace/
-      ...
-    ...</pre>
-</details>
-""")
-
-                with gr.Row():
-                    model_input = gr.Textbox(
-                        label="Model Name *",
-                        placeholder="e.g. gpt-5, claude-opus-4, gemini-3-pro",
-                        scale=2,
-                    )
-                    team_input = gr.Textbox(
-                        label="Team / Organization *",
-                        placeholder="e.g. OpenAI, Google DeepMind, your-lab-name",
-                        scale=2,
-                    )
+                gr.Markdown(
+                    "Submit a results JSON produced by `python -m harness.run_all` or `python -m harness.ablation`. "
+                    "We re-run the deterministic graders server-side to verify."
+                )
                 with gr.Row():
-                    framework_input = gr.Textbox(
-                        label="Framework (optional)",
-                        placeholder="e.g. langgraph, crewai, google-adk, custom",
-                        scale=2,
-                    )
-                    contact_input = gr.Textbox(
-                        label="Contact Email (optional)",
-                        placeholder="your@email.com",
-                        scale=2,
-                    )
+                    model_in = gr.Textbox(label="Model name *", placeholder="e.g. claude-sonnet-4-6")
+                    team_in = gr.Textbox(label="Team / organization *", placeholder="e.g. Anthropic, Google DeepMind, your-lab")
                 with gr.Row():
-                    description_input = gr.Textbox(
-                        label="Description (optional)",
-                        placeholder="Brief description of your agent setup, tools used, etc.",
-                        lines=2,
-                        scale=3,
-                    )
-                    seed_input = gr.Textbox(
-                        label="Seed",
-                        value="0",
-                        placeholder="0",
-                        scale=1,
-                    )
-
-                file_input = gr.File(
-                    label="Upload submission.zip (workspace artifacts, max 100 MB)",
-                    file_types=[".zip"],
-                )
-
-                submit_btn = gr.Button("Submit for Grading", variant="primary")
-                result_output = gr.Markdown("")
-
-                submit_btn.click(
-                    fn=validate_and_accept_submission,
-                    inputs=[file_input, model_input, team_input,
-                            framework_input, contact_input,
-                            description_input, seed_input],
-                    outputs=result_output,
+                    framework_in = gr.Textbox(label="Framework (optional)", placeholder="e.g. langgraph, crewai, custom")
+                    contact_in = gr.Textbox(label="Contact email (optional)", placeholder="you@example.com")
+                description_in = gr.Textbox(label="Description (optional)", placeholder="Tools, prompts, framework version, anything reviewers should know.", lines=2)
+                file_in = gr.File(label="Results JSON (the file harness.run_all wrote)", file_types=[".json"])
+                btn = gr.Button("Submit for grading", variant="primary")
+                out = gr.Markdown("")
+                btn.click(
+                    fn=validate_submission,
+                    inputs=[file_in, model_in, team_in, contact_in, framework_in, description_in],
+                    outputs=out,
                 )
 
+        gr.Markdown(
+            "<div style='text-align:center;color:#9ca3af;font-size:0.78rem;margin-top:1.25rem;'>"
+            "TeamBench &middot; <a href='https://github.com/ybkim95/TeamBench'>GitHub</a> &middot; "
+            "<a href='https://huggingface.co/datasets/ybkim95/teambench'>HuggingFace</a> &middot; "
+            "<a href='https://teambench.github.io'>teambench.github.io</a>"
+            "</div>"
+        )
     return demo
 
 
-# ---------------------------------------------------------------------------
-# Entry point
-# ---------------------------------------------------------------------------
-
 if __name__ == "__main__":
-    app = build_app()
-    app.launch(
-        server_name="0.0.0.0",
-        server_port=int(os.environ.get("PORT", 7860)),
-        share=False,
-    )
+    build_app().launch(server_name="0.0.0.0", server_port=int(os.environ.get("PORT", 7860)))

requirements.txt

@@ -1,2 +1,2 @@
-gradio>=5.0.0
-pandas>=2.0.0
+gradio==5.12.0
+pandas>=2.0