add balance/mean-reversion signal and Cloudflare visitor logging

Balance signal (15% weight) favors under-represented chairs over last 50
games. Visitor middleware captures real IPs from CF headers, batched into
ClickHouse with 90-day TTL.

app/db.py

@@ -61,6 +61,15 @@ def run_migrations():
         )
         client.insert("_migrations", [["swap_ac_chairs"]], column_names=["name"])
         log.info("Migration swap_ac_chairs applied")
+    # Ensure visitors table exists (for existing deployments)
+    client.command(
+        "CREATE TABLE IF NOT EXISTS visitors ("
+        "  ip String, country String, path String, method String,"
+        "  user_agent String, referer String, accept_lang String,"
+        "  created_at DateTime DEFAULT now()"
+        ") ENGINE = MergeTree() ORDER BY (created_at, ip)"
+        " TTL created_at + INTERVAL 90 DAY"
+    )
     _migrations_applied = True
 
 
@@ -143,6 +152,31 @@ def upsert_user(user: dict):
     )
 
 
+@_with_lock
+def insert_visitors(batch: list[dict]):
+    """Bulk insert visitor records."""
+    if not batch:
+        return
+    client = get_client()
+    rows = [
+        [
+            v.get("ip", ""),
+            v.get("country", ""),
+            v.get("path", ""),
+            v.get("method", ""),
+            v.get("user_agent", ""),
+            v.get("referer", ""),
+            v.get("accept_lang", ""),
+        ]
+        for v in batch
+    ]
+    client.insert(
+        "visitors",
+        rows,
+        column_names=["ip", "country", "path", "method", "user_agent", "referer", "accept_lang"],
+    )
+
+
 @_with_lock
 def get_recent_games(n: int = 50) -> list[dict]:
     """Get last N completed games."""
@@ -1000,8 +1034,17 @@ def _bayesian_prediction(winners, markov1, markov2):
     else:
         streak = {c: 1 / 3 for c in CHAIR_LABELS}
 
-    weights = {"base_rate": 0.20, "markov_1": 0.30, "markov_2": 0.25, "recent_20": 0.15, "streak": 0.10}
+    # Signal 6: Balance / Mean Reversion — 15%
-    signals = {"base_rate": base, "markov_1": m1, "markov_2": m2, "recent_20": rec, "streak": streak}
+    # Look at last 50 games, invert frequencies to favor under-represented chairs
+    window = min(50, len(winners))
+    recent_50 = winners[-window:]
+    freq = {c: recent_50.count(c) / window for c in CHAIR_LABELS}
+    balance = {c: max(0.01, 2 / 3 - freq[c]) for c in CHAIR_LABELS}
+    bal_total = sum(balance.values())
+    balance = {c: balance[c] / bal_total for c in CHAIR_LABELS}
+
+    weights = {"base_rate": 0.15, "markov_1": 0.25, "markov_2": 0.25, "recent_20": 0.10, "streak": 0.10, "balance": 0.15}
+    signals = {"base_rate": base, "markov_1": m1, "markov_2": m2, "recent_20": rec, "streak": streak, "balance": balance}
 
     combined = {c: 0 for c in CHAIR_LABELS}
     for sig_name, weight in weights.items():
@@ -1122,7 +1165,7 @@ def _backtest_theories(winners):
     if len(winners) <= warmup:
         return {"error": "Not enough data for backtesting"}
 
-    theories = ["base_rate", "markov_1", "markov_2", "recent_20", "streak", "combined"]
+    theories = ["base_rate", "markov_1", "markov_2", "recent_20", "streak", "balance", "combined"]
     full_hits = {t: 0 for t in theories}
     semi_hits = {t: 0 for t in theories}
     total_tested = 0
@@ -1176,10 +1219,19 @@ def _backtest_theories(winners):
             streak_probs = {c: 1 / 3 for c in CHAIR_LABELS}
         streak_ranked = sorted(CHAIR_LABELS, key=lambda c: streak_probs[c], reverse=True)
 
+        # Balance / Mean Reversion
+        bal_window = min(50, len(history))
+        bal_recent = history[-bal_window:]
+        bal_freq = {c: bal_recent.count(c) / bal_window for c in CHAIR_LABELS}
+        bal_probs = {c: max(0.01, 2 / 3 - bal_freq[c]) for c in CHAIR_LABELS}
+        bal_t = sum(bal_probs.values())
+        bal_probs = {c: bal_probs[c] / bal_t for c in CHAIR_LABELS}
+        bal_ranked = sorted(CHAIR_LABELS, key=lambda c: bal_probs[c], reverse=True)
+
         # Combined Bayesian
         combined = {c: 0 for c in CHAIR_LABELS}
-        weights = {"base_rate": 0.20, "markov_1": 0.30, "markov_2": 0.25, "recent_20": 0.15, "streak": 0.10}
+        weights = {"base_rate": 0.15, "markov_1": 0.25, "markov_2": 0.25, "recent_20": 0.10, "streak": 0.10, "balance": 0.15}
-        signals = {"base_rate": base, "markov_1": m1_probs, "markov_2": m2_probs, "recent_20": rec, "streak": streak_probs}
+        signals = {"base_rate": base, "markov_1": m1_probs, "markov_2": m2_probs, "recent_20": rec, "streak": streak_probs, "balance": bal_probs}
         for sig_name, weight in weights.items():
             for c in CHAIR_LABELS:
                 combined[c] += weight * signals[sig_name].get(c, 1 / 3)
@@ -1187,7 +1239,8 @@ def _backtest_theories(winners):
 
         ranked = {
             "base_rate": base_ranked, "markov_1": m1_ranked, "markov_2": m2_ranked,
-            "recent_20": rec_ranked, "streak": streak_ranked, "combined": combined_ranked,
+            "recent_20": rec_ranked, "streak": streak_ranked, "balance": bal_ranked,
+            "combined": combined_ranked,
         }
         for t in theories:
             pick = ranked[t][0]

app/server.py

@@ -30,7 +30,65 @@ class WebServer:
     def __init__(self):
         self.app = web.Application()
         self.clients: set[web.WebSocketResponse] = set()
+        self._visitor_buffer: list[dict] = []
+        self._visitor_lock = asyncio.Lock()
         self._setup_routes()
+        self.app.middlewares.append(self._make_visitor_middleware())
+
+    def _make_visitor_middleware(self):
+        server = self
+
+        @web.middleware
+        async def visitor_middleware(request: web.Request, handler):
+            response = await handler(request)
+            path = request.path
+            # Skip static files and WebSocket upgrades
+            if path.startswith("/static/") or request.headers.get("Upgrade", "").lower() == "websocket":
+                return response
+            ip = (
+                request.headers.get("CF-Connecting-IP")
+                or (request.headers.get("X-Forwarded-For", "").split(",")[0].strip())
+                or request.remote
+                or ""
+            )
+            visitor = {
+                "ip": ip,
+                "country": request.headers.get("CF-IPCountry", ""),
+                "path": path,
+                "method": request.method,
+                "user_agent": request.headers.get("User-Agent", ""),
+                "referer": request.headers.get("Referer", ""),
+                "accept_lang": request.headers.get("Accept-Language", ""),
+            }
+            batch = None
+            async with server._visitor_lock:
+                server._visitor_buffer.append(visitor)
+                if len(server._visitor_buffer) >= 20:
+                    batch = server._visitor_buffer[:]
+                    server._visitor_buffer.clear()
+            if batch:
+                try:
+                    await _run_sync(db.insert_visitors, batch)
+                except Exception as e:
+                    log.warning("Visitor insert failed: %s", e)
+            return response
+
+        return visitor_middleware
+
+    async def _flush_visitors(self):
+        """Periodically flush visitor buffer so low-traffic visits aren't lost."""
+        while True:
+            await asyncio.sleep(30)
+            batch = None
+            async with self._visitor_lock:
+                if self._visitor_buffer:
+                    batch = self._visitor_buffer[:]
+                    self._visitor_buffer.clear()
+            if batch:
+                try:
+                    await _run_sync(db.insert_visitors, batch)
+                except Exception as e:
+                    log.warning("Visitor flush failed: %s", e)
 
     def _setup_routes(self):
         self.app.router.add_get("/", self._handle_index)
@@ -219,6 +277,7 @@ class WebServer:
         site = web.TCPSite(runner, "0.0.0.0", config.WEB_PORT)
         await site.start()
         log.info("Web server listening on http://0.0.0.0:%s", config.WEB_PORT)
+        flush_task = asyncio.create_task(self._flush_visitors())
         # Keep running until cancelled
         try:
             while True:
@@ -226,4 +285,5 @@ class WebServer:
         except asyncio.CancelledError:
             pass
         finally:
+            flush_task.cancel()
             await runner.cleanup()

clickhouse/init.sql

@@ -40,3 +40,16 @@ CREATE TABLE IF NOT EXISTS users (
     updated_at    DateTime DEFAULT now()
 ) ENGINE = ReplacingMergeTree(updated_at)
 ORDER BY user_id;
+
+CREATE TABLE IF NOT EXISTS visitors (
+    ip            String,
+    country       String,
+    path          String,
+    method        String,
+    user_agent    String,
+    referer       String,
+    accept_lang   String,
+    created_at    DateTime DEFAULT now()
+) ENGINE = MergeTree()
+ORDER BY (created_at, ip)
+TTL created_at + INTERVAL 90 DAY;

static/predictions.html

@@ -456,7 +456,7 @@ function renderPrediction(data) {
 
   // Signal table
   const tbody = $('signal-table').querySelector('tbody');
-  const sigNames = {'base_rate':'Base Rate','markov_1':'Markov-1','markov_2':'Markov-2','recent_20':'Recent 20','streak':'Streak'};
+  const sigNames = {'base_rate':'Base Rate','markov_1':'Markov-1','markov_2':'Markov-2','recent_20':'Recent 20','streak':'Streak','balance':'Balance'};
   tbody.innerHTML = Object.entries(data.signals).map(([key, sig]) =>
     `<tr><td style="text-align:left">${sigNames[key]||key}</td><td>${(sig.weight*100).toFixed(0)}%</td>` +
     CHAIRS.map(c => `<td style="color:${CHAIR_COLORS[c]}">${pct(sig.probs[c])}</td>`).join('') + '</tr>'
@@ -914,7 +914,7 @@ function renderRunsTest(runs) {
 
 function renderBacktest(bt) {
   if (bt.error) { $('backtest-cards').innerHTML = `<div style="color:var(--text2)">${bt.error}</div>`; return; }
-  const names = {base_rate:'Base Rate',markov_1:'Markov-1',markov_2:'Markov-2',recent_20:'Recent 20',streak:'Streak',combined:'Combined'};
+  const names = {base_rate:'Base Rate',markov_1:'Markov-1',markov_2:'Markov-2',recent_20:'Recent 20',streak:'Streak',balance:'Balance',combined:'Combined'};
   $('backtest-cards').innerHTML = Object.entries(bt.accuracy).map(([key, acc]) => {
     const fh = bt.full_hits?.[key] ?? '?';
     const sh = bt.semi_hits?.[key] ?? '?';
@@ -925,7 +925,7 @@ function renderBacktest(bt) {
   }).join('');
   if (bt.rolling_accuracy) {
     const ctx = $('backtest-chart').getContext('2d');
-    const colors = {base_rate:'#8b8fa3',markov_1:'#3b82f6',markov_2:'#ec4899',recent_20:'#f59e0b',streak:'#10b981',combined:'#6c5ce7'};
+    const colors = {base_rate:'#8b8fa3',markov_1:'#3b82f6',markov_2:'#ec4899',recent_20:'#f59e0b',streak:'#10b981',balance:'#f472b6',combined:'#6c5ce7'};
     const datasets = Object.entries(bt.rolling_accuracy).map(([key, data]) => ({
       label: names[key]||key, data, borderColor: colors[key]||'#fff', backgroundColor: 'transparent',
       borderWidth: key === 'combined' ? 3 : 1.5, pointRadius: 0, tension: 0.3,