diff --git a/README.md b/README.md
index 58fd99c..f9d12e3 100644
--- a/README.md
+++ b/README.md
@@ -95,12 +95,16 @@ powershell -ExecutionPolicy Bypass -File tools\install_windows_torch_retrainer.p
 Для удалённого запуска с телефона или с бота используется Windows training agent. Бот на `tb.kusoft.xyz` хранит очередь заданий, а Windows-машина сама подключается к интернету, забирает задания, обучает модель и загружает артефакты обратно:
 
 ```powershell
-powershell -ExecutionPolicy Bypass -File tools\install_windows_training_agent.ps1 -ApiAuth "login:password" -StartNow
+powershell -ExecutionPolicy Bypass -File tools\install_windows_training_agent.ps1 -ApiAuth "<TRADEBOT_TRAINING_TOKEN>" -StartNow
 ```
 
-Установщик регистрирует Scheduled Task `TradeBot Windows Training Agent` при входе в Windows и удаляет старые локальные retrain-задачи, чтобы обучение запускалось через очередь, а не двумя независимыми механизмами.
+Установщик сохраняет worker-токен через Windows DPAPI, удаляет его старую plaintext-копию из пользовательского окружения и включает постоянный запуск агента. С правами администратора используется Scheduled Task с watchdog; без повышения прав — штатный ярлык в пользовательской папке Startup. Старые локальные retrain-задачи удаляются, чтобы обучение запускалось через очередь, а не двумя независимыми механизмами.
 
-По умолчанию Windows-расписание переобучает PyTorch `LSTM/GRU` каждые 6 часов с `--limit 3000` на 12 spot-парах из `SYMBOLS`. Параметры можно переопределить через env: `TORCH_RETRAIN_SYMBOLS`, `TORCH_RETRAIN_LIMIT`, `TORCH_RETRAIN_LOOKBACKS`, `TORCH_RETRAIN_ARCHITECTURES`, `TORCH_RETRAIN_HIDDEN_SIZES`, `TORCH_RETRAIN_LAYERS`, `TORCH_RETRAIN_DROPOUTS`, `TORCH_RETRAIN_HORIZON`, `TORCH_RETRAIN_HORIZONS`, `TORCH_RETRAIN_CONTEXT_SYMBOLS`, `TORCH_RETRAIN_FEATURES`, `TORCH_RETRAIN_SEED`, `TORCH_RETRAIN_EPOCHS`, `TORCH_RETRAIN_PATIENCE`, `TORCH_RETRAIN_INTERVAL`, `TORCH_RETRAIN_ENV`.
+По умолчанию Windows-agent обучает pooled multi-asset PyTorch `LSTM/GRU` на `6000` часовых свечах: общие recurrent-веса получают one-hot embedding символа, прогноз усредняется по seed `7/19/43`, модели сравниваются на validation-folds, а пороги калибруются отдельно для каждой пары. Search space использует lookback `32/64/128`, hidden `64/96`, dropout `0.20`, AdamW learning rate `0.0007` и weight decay `0.0005`; untouched holdout и quality gate не ослабляются. Параметры можно переопределить через env: `TORCH_RETRAIN_SYMBOLS`, `TORCH_RETRAIN_LIMIT`, `TORCH_RETRAIN_LOOKBACKS`, `TORCH_RETRAIN_ARCHITECTURES`, `TORCH_RETRAIN_HIDDEN_SIZES`, `TORCH_RETRAIN_LAYERS`, `TORCH_RETRAIN_DROPOUTS`, `TORCH_RETRAIN_HORIZON`, `TORCH_RETRAIN_HORIZONS`, `TORCH_RETRAIN_CONTEXT_SYMBOLS`, `TORCH_RETRAIN_FEATURES`, `TORCH_RETRAIN_SEED`, `TORCH_RETRAIN_ENSEMBLE_SEEDS`, `TORCH_RETRAIN_SELECTION_FOLDS`, `TORCH_RETRAIN_LEARNING_RATE`, `TORCH_RETRAIN_WEIGHT_DECAY`, `TORCH_RETRAIN_EPOCHS`, `TORCH_RETRAIN_PATIENCE`, `TORCH_RETRAIN_INTERVAL`, `TORCH_RETRAIN_ENV`.
+
+Loss и выбор гиперпараметров учитывают after-cost trading utility и ранговую связь прогноза с будущей доходностью, а не только MAE. В каждом walk-forward fold вероятность `P(up)` калибруется Platt-моделью исключительно на train-части; затем на этой же train-части выбираются глобальные и per-symbol пороги, которые применяются к test-части. Калибратор не имеет fallback на единичные сделки: если минимальная статистика не набрана, кандидат получает `calibration_insufficient` и не может пройти gate.
+
+Основной decision horizon — `12h`, дополнительные горизонты — `3/6/12/24`. Это согласует прогноз с round-trip cost: при текущих fee/slippage полный вход-выход стоит около `0.26%`, поэтому прежний `3h` target чаще описывал шум, который не покрывал издержки. Threshold search оценивается тем же execution replay со stop-loss, take-profit, ATR trailing и forecast-exit, который используется в walk-forward. `holdout_skill` остаётся только в финальном отчёте и никогда не участвует в фильтрации входов или подборе порогов.
 
 Если retrain запускается с `-DeployToPi`, после успешного guard он синхронизирует `runtime/lstm_forecaster.json`, `runtime/torch_retrain_guard.json` и `runtime/torch_threshold_calibration.json` на Raspberry Pi через SSH-ключ и перезапускает сервис `tradebot`. Отдельный запуск sync:
 
diff --git a/android/TradeBotMonitor/README.md b/android/TradeBotMonitor/README.md
index b2c5358..1d82a97 100644
--- a/android/TradeBotMonitor/README.md
+++ b/android/TradeBotMonitor/README.md
@@ -38,15 +38,11 @@ https://tb.kusoft.xyz
 
 Этот адрес установлен в приложении по умолчанию. Если в настройках ввести просто `tb.kusoft.xyz`, приложение само добавит `https://`.
 
-Если домен защищён авторизацией, в поле `API auth` можно указать:
-
-- `login:password` — приложение отправит HTTP Basic;
-- `Basic ...` — готовый Basic header;
-- `Bearer ...` или просто токен — приложение отправит Bearer.
+В поле `API-токен` указывается отдельный токен Android-клиента (`TRADEBOT_API_TOKEN` на сервере). Логин и пароль reverse proxy приложению не нужны. Токен отправляется как Bearer/X-TradeBot-Token и хранится зашифрованным ключом Android Keystore.
 
 ## Переобучение
 
-Телефон не обучает модель локально. Вкладка `Обучение` ставит задание в очередь на `tb.kusoft.xyz`, а Windows-agent на закреплённой машине `DESKTOP-TMFDL0H` сам выходит в интернет, забирает задание, обучает модель и отправляет артефакты обратно боту. Так телефон становится пультом запуска/расписания, а тяжёлый PyTorch retrain остаётся на нормальном компьютере даже если он находится в другой сети.
+Телефон не обучает модель локально. Вкладка `Обучение` ставит задание в очередь на `tb.kusoft.xyz`, а Windows-agent на закреплённой машине `SEVENHILL` (`G:\Repos\TradeBot`) сам выходит в интернет, забирает задание, обучает модель и отправляет артефакты обратно боту. Так телефон становится пультом запуска/расписания, а тяжёлый PyTorch retrain остаётся на нормальном компьютере даже если он находится в другой сети.
 
 ## Live-торговля
 
diff --git a/android/TradeBotMonitor/app/build.gradle.kts b/android/TradeBotMonitor/app/build.gradle.kts
index f681b5a..0c9f416 100644
--- a/android/TradeBotMonitor/app/build.gradle.kts
+++ b/android/TradeBotMonitor/app/build.gradle.kts
@@ -10,7 +10,7 @@ android {
         applicationId = "xyz.kusoft.tradebotmonitor"
         minSdk = 26
         targetSdk = 36
-        versionCode = 18
-        versionName = "0.3.0"
+        versionCode = 20
+        versionName = "0.4.1"
     }
 }
diff --git a/android/TradeBotMonitor/app/src/main/java/xyz/kusoft/tradebotmonitor/AppPrefs.kt b/android/TradeBotMonitor/app/src/main/java/xyz/kusoft/tradebotmonitor/AppPrefs.kt
index 1851deb..4b6a6b7 100644
--- a/android/TradeBotMonitor/app/src/main/java/xyz/kusoft/tradebotmonitor/AppPrefs.kt
+++ b/android/TradeBotMonitor/app/src/main/java/xyz/kusoft/tradebotmonitor/AppPrefs.kt
@@ -16,7 +16,14 @@ class AppPrefs(context: Context) {
         if (saved.isNullOrBlank() || saved == LEGACY_PI_API_BASE_URL) {
             prefs.edit().putString("api_base_url", DEFAULT_API_BASE_URL).apply()
         }
-        if (prefs.getString("training_computer_name", null).isNullOrBlank()) {
+        val trainingComputerName = prefs.getString("training_computer_name", null)?.trim()
+        val trainingComputerPath = prefs.getString("training_computer_path", null)?.trim()
+        if (
+            trainingComputerName.isNullOrBlank() ||
+            trainingComputerName == LEGACY_TRAINING_COMPUTER_NAME ||
+            trainingComputerPath.isNullOrBlank() ||
+            trainingComputerPath == LEGACY_TRAINING_COMPUTER_PATH
+        ) {
             pinDefaultTrainingComputer()
         }
     }
@@ -134,8 +141,10 @@ class AppPrefs(context: Context) {
     private companion object {
         const val DEFAULT_API_BASE_URL = "https://tb.kusoft.xyz"
         const val LEGACY_PI_API_BASE_URL = "http://192.168.0.185:8787"
-        const val DEFAULT_TRAINING_COMPUTER_NAME = "DESKTOP-TMFDL0H"
-        const val DEFAULT_TRAINING_COMPUTER_PATH = "C:\\Repos\\TradeBot"
+        const val DEFAULT_TRAINING_COMPUTER_NAME = "SEVENHILL"
+        const val DEFAULT_TRAINING_COMPUTER_PATH = "G:\\Repos\\TradeBot"
+        const val LEGACY_TRAINING_COMPUTER_NAME = "DESKTOP-TMFDL0H"
+        const val LEGACY_TRAINING_COMPUTER_PATH = "C:\\Repos\\TradeBot"
         const val TOKEN_KEY_ALIAS = "tradebot_api_auth_v1"
     }
 }
diff --git a/android/TradeBotMonitor/app/src/main/java/xyz/kusoft/tradebotmonitor/MainActivity.kt b/android/TradeBotMonitor/app/src/main/java/xyz/kusoft/tradebotmonitor/MainActivity.kt
index 4c4383b..c517a17 100644
--- a/android/TradeBotMonitor/app/src/main/java/xyz/kusoft/tradebotmonitor/MainActivity.kt
+++ b/android/TradeBotMonitor/app/src/main/java/xyz/kusoft/tradebotmonitor/MainActivity.kt
@@ -489,21 +489,19 @@ class MainActivity : Activity() {
             liveSignature = liveBlockSignature(snapshot),
         )
 
-        val (savedLogin, savedPassword) = authParts(prefs.commandToken)
         val apiInput = input("Адрес API бота", prefs.apiBaseUrl)
-        val loginInput = input("Логин API", savedLogin)
-        val passwordInput = input("Пароль API", savedPassword).apply {
+        val tokenInput = input("API-токен", prefs.commandToken).apply {
             inputType = InputType.TYPE_CLASS_TEXT or InputType.TYPE_TEXT_VARIATION_PASSWORD
         }
         box.addView(section("Подключение", LinearLayout(this).apply {
             orientation = LinearLayout.VERTICAL
             addView(apiInput)
-            addView(loginInput.top(dp(8)))
-            addView(passwordInput.top(dp(8)))
+            addView(tokenInput.top(dp(8)))
+            addView(text("Токен выдаётся сервером для Android-клиента и хранится в Android Keystore.", 12f, Typeface.NORMAL, palette.muted).top(dp(8)))
             addView(actionRow(
                 "Сохранить" to {
                     prefs.apiBaseUrl = apiInput.text.toString()
-                    prefs.commandToken = authToken(loginInput.text.toString(), passwordInput.text.toString())
+                    prefs.commandToken = tokenInput.text.toString()
                     toast("Подключение сохранено")
                     refreshData(silent = false)
                 },
@@ -1020,6 +1018,13 @@ class MainActivity : Activity() {
         LinearLayout(this).apply {
             val coordination = retrain.optJSONObject("coordination") ?: JSONObject()
             val activeJob = coordination.optJSONObject("active_job")
+            val worker = coordination.optJSONObject("worker")
+            val trainingComputerName = worker?.optStringClean("name")
+                ?.takeIf { it.isNotBlank() }
+                ?: prefs.trainingComputerName
+            val trainingComputerPath = worker?.optStringClean("path")
+                ?.takeIf { it.isNotBlank() }
+                ?: prefs.trainingComputerPath
             val agentRecentlySeen = coordination.optBoolean(
                 "agent_recently_seen",
                 coordination.optBoolean("agent_online", false),
@@ -1036,8 +1041,8 @@ class MainActivity : Activity() {
             val connectionColor = if (agentRecentlySeen || agentBusy) palette.green else palette.amber
             orientation = LinearLayout.VERTICAL
             addView(text("Компьютер обучения", 12f, Typeface.NORMAL, palette.muted))
-            addView(text(prefs.trainingComputerName, 18f, Typeface.BOLD, palette.green).top(dp(5)))
-            addView(text(prefs.trainingComputerPath, 12f, Typeface.NORMAL, palette.muted).top(dp(4)))
+            addView(text(trainingComputerName, 18f, Typeface.BOLD, palette.green).top(dp(5)))
+            addView(text(trainingComputerPath, 12f, Typeface.NORMAL, palette.muted).top(dp(4)))
             addView(keyValueLine("Связь агента", connectionText, connectionColor).top(dp(4)))
             addView(text("Бот ставит задания через tb.kusoft.xyz, а этот Windows-agent сам забирает их через интернет и возвращает результат.", 12f, Typeface.NORMAL, palette.muted).top(dp(8)))
         }
@@ -1052,10 +1057,9 @@ class MainActivity : Activity() {
                 val agentOnline = coordination.optBoolean("agent_online", false)
                 addView(keyValueLine("Состояние", if (agentOnline) "готов к запуску" else "ждет Windows-agent", if (agentOnline) palette.green else palette.amber).top(dp(8)))
                 if (latestJob != null) {
-                    val latestStatus = latestJob.optStringClean("status")
-                    val latestPhase = latestJob.optStringClean("phase")
                     val latestMessage = latestJob.optStringClean("message")
-                    addView(keyValueLine("Последнее обучение", trainingJobLabel(latestStatus, latestPhase), trainingJobColor(latestStatus, latestPhase)).top(dp(8)))
+                    addView(keyValueLine("Последнее обучение", trainingJobLabel(latestJob), trainingJobColor(latestJob)).top(dp(8)))
+                    addView(trainingModelDecisionLine(latestJob).top(dp(6)))
                     if (latestMessage.isNotBlank()) {
                         addView(text(latestMessage.take(160), 12f, Typeface.NORMAL, palette.muted).top(dp(6)))
                     }
@@ -1064,12 +1068,11 @@ class MainActivity : Activity() {
             }
             val job = activeJob
             val status = job.optStringClean("status")
-            val phase = job.optStringClean("phase")
             val progress = job.optInt("progress_percent", if (status == "completed") 100 else 0).coerceIn(0, 100)
             val message = job.optStringClean("message")
-            addView(keyValueLine("Состояние", trainingJobLabel(status, phase), trainingJobColor(status, phase)).top(dp(8)))
+            addView(keyValueLine("Состояние", trainingJobLabel(job), trainingJobColor(job)).top(dp(8)))
             addView(allocationBar(progress / 100.0).top(dp(10)))
-            addView(keyValueLine("Прогресс", "$progress%", trainingJobColor(status, phase)).top(dp(8)))
+            addView(keyValueLine("Прогресс", "$progress%", trainingJobColor(job)).top(dp(8)))
             if (message.isNotBlank()) {
                 addView(text(message.take(160), 12f, Typeface.NORMAL, palette.text).top(dp(8)))
             }
@@ -1241,22 +1244,19 @@ class MainActivity : Activity() {
         if (!isAuthError()) {
             return emptyState("Нет данных от API. ${lastError.ifBlank { "Проверь подключение." }}")
         }
-        val (savedLogin, savedPassword) = authParts(prefs.commandToken)
         val apiInput = input("Адрес API бота", prefs.apiBaseUrl)
-        val loginInput = input("Логин API", savedLogin)
-        val passwordInput = input("Пароль API", savedPassword).apply {
+        val tokenInput = input("API-токен", prefs.commandToken).apply {
             inputType = InputType.TYPE_CLASS_TEXT or InputType.TYPE_TEXT_VARIATION_PASSWORD
         }
-        return section("Нужен вход в API", LinearLayout(this).apply {
+        return section("Нужен API-токен", LinearLayout(this).apply {
             orientation = LinearLayout.VERTICAL
-            addView(text("Сервер tb.kusoft.xyz отвечает 401, значит он доступен, но требует авторизацию. Введите логин и пароль, приложение само отправит Basic Auth.", 13f, Typeface.NORMAL, palette.muted))
+            addView(text("Сервер доступен, но Android API-токен отсутствует или неверен. Введите отдельный токен клиента; логин и пароль сайта здесь не используются.", 13f, Typeface.NORMAL, palette.muted))
             addView(apiInput.top(dp(12)))
-            addView(loginInput.top(dp(8)))
-            addView(passwordInput.top(dp(8)))
+            addView(tokenInput.top(dp(8)))
             addView(actionRow(
                 "Подключиться" to {
                     prefs.apiBaseUrl = apiInput.text.toString()
-                    prefs.commandToken = authToken(loginInput.text.toString(), passwordInput.text.toString())
+                    prefs.commandToken = tokenInput.text.toString()
                     toast("Доступ сохранен, проверяю API")
                     refreshData(silent = false)
                 },
@@ -1642,11 +1642,40 @@ class MainActivity : Activity() {
             },
         ).joinToString("|")
 
-    private fun trainingJobLabel(status: String, phase: String): String =
-        when (status) {
+    private fun trainingModelDecision(job: JSONObject): String {
+        val explicit = job.optStringClean("model_decision")
+        if (explicit in setOf("accepted", "rejected")) return explicit
+        val summary = job.optJSONObject("summary") ?: return ""
+        return if (summary.has("accepted")) {
+            if (summary.optBoolean("accepted", false)) "accepted" else "rejected"
+        } else {
+            ""
+        }
+    }
+
+    private fun trainingModelDecisionLine(job: JSONObject): View {
+        val decision = trainingModelDecision(job)
+        val value = when (decision) {
+            "accepted" -> "кандидат принят"
+            "rejected" -> "кандидат отклонён"
+            else -> "нет решения gate"
+        }
+        val color = when (decision) {
+            "accepted" -> palette.green
+            "rejected" -> palette.amber
+            else -> palette.muted
+        }
+        return keyValueLine("Результат модели", value, color)
+    }
+
+    private fun trainingJobLabel(job: JSONObject): String {
+        val status = job.optStringClean("status")
+        val phase = job.optStringClean("phase")
+        val completedWithDecision = trainingModelDecision(job).isNotBlank()
+        return when (status) {
             "pending" -> "ждет Windows-agent"
             "completed" -> "завершено успешно"
-            "failed" -> "ошибка обучения"
+            "failed" -> if (completedWithDecision) "завершено успешно" else "ошибка обучения"
             "running" -> when (phase) {
                 "claimed" -> "задание получено"
                 "training" -> "идет обучение"
@@ -1656,14 +1685,19 @@ class MainActivity : Activity() {
             }
             else -> "готов к запуску"
         }
+    }
 
-    private fun trainingJobColor(status: String, phase: String): Int =
-        when {
-            status == "completed" -> palette.green
+    private fun trainingJobColor(job: JSONObject): Int {
+        val status = job.optStringClean("status")
+        val phase = job.optStringClean("phase")
+        val completedWithDecision = trainingModelDecision(job).isNotBlank()
+        return when {
+            status == "completed" || completedWithDecision -> palette.green
             status == "failed" -> palette.red
             status == "pending" || status == "running" || phase.isNotBlank() -> palette.amber
             else -> palette.text
         }
+    }
 
     private fun normalizedAction(raw: String?): String {
         val value = raw.orEmpty().uppercase(Locale.US)
@@ -1700,27 +1734,6 @@ class MainActivity : Activity() {
             else -> value.ifBlank { "нет данных" }
         }
 
-    private fun authParts(value: String): Pair<String, String> {
-        val trimmed = value.trim()
-        if (trimmed.isBlank() || trimmed.startsWith("Bearer ", ignoreCase = true)) {
-            return "" to ""
-        }
-        val basicPrefix = "Basic "
-        val raw = if (trimmed.startsWith(basicPrefix, ignoreCase = true)) "" else trimmed
-        val separator = raw.indexOf(':')
-        return if (separator >= 0) {
-            raw.substring(0, separator) to raw.substring(separator + 1)
-        } else {
-            raw to ""
-        }
-    }
-
-    private fun authToken(login: String, password: String): String {
-        val cleanLogin = login.trim()
-        val cleanPassword = password.trim()
-        return if (cleanLogin.isBlank() && cleanPassword.isBlank()) "" else "$cleanLogin:$cleanPassword"
-    }
-
     private fun modelLabel(value: String): String =
         when {
             value.contains("gru", ignoreCase = true) -> "PyTorch GRU"
diff --git a/android/TradeBotMonitor/app/src/main/java/xyz/kusoft/tradebotmonitor/TradeBotApi.kt b/android/TradeBotMonitor/app/src/main/java/xyz/kusoft/tradebotmonitor/TradeBotApi.kt
index 1731298..86b3e48 100644
--- a/android/TradeBotMonitor/app/src/main/java/xyz/kusoft/tradebotmonitor/TradeBotApi.kt
+++ b/android/TradeBotMonitor/app/src/main/java/xyz/kusoft/tradebotmonitor/TradeBotApi.kt
@@ -1,6 +1,5 @@
 package xyz.kusoft.tradebotmonitor
 
-import android.util.Base64
 import org.json.JSONArray
 import org.json.JSONObject
 import java.io.BufferedReader
@@ -98,7 +97,7 @@ class TradeBotApi(
         connection.disconnect()
         if (code !in 200..299) {
             if (code == HttpURLConnection.HTTP_UNAUTHORIZED) {
-                throw IllegalStateException("HTTP 401: сервер требует логин и пароль")
+                throw IllegalStateException("HTTP 401: API-токен отсутствует или неверен")
             }
             throw IllegalStateException("HTTP $code: ${text.take(240)}")
         }
@@ -282,15 +281,8 @@ class TradeBotApi(
 private fun applyAuthHeaders(connection: HttpURLConnection, token: String) {
     val value = token.trim()
     if (value.isBlank()) return
-    connection.setRequestProperty("X-TradeBot-Token", value)
-    val authorization = when {
-        value.startsWith("Basic ", ignoreCase = true) -> value
-        value.startsWith("Bearer ", ignoreCase = true) -> value
-        ":" in value -> {
-            val encoded = Base64.encodeToString(value.toByteArray(StandardCharsets.UTF_8), Base64.NO_WRAP)
-            "Basic $encoded"
-        }
-        else -> "Bearer $value"
-    }
+    val rawToken = value.removePrefix("Bearer ").removePrefix("bearer ").trim()
+    connection.setRequestProperty("X-TradeBot-Token", rawToken)
+    val authorization = "Bearer $rawToken"
     connection.setRequestProperty("Authorization", authorization)
 }
diff --git a/android/TradeBotMonitor/gradle/wrapper/gradle-wrapper.jar b/android/TradeBotMonitor/gradle/wrapper/gradle-wrapper.jar
new file mode 100644
index 0000000..d997cfc
Binary files /dev/null and b/android/TradeBotMonitor/gradle/wrapper/gradle-wrapper.jar differ
diff --git a/android/TradeBotMonitor/gradle/wrapper/gradle-wrapper.properties b/android/TradeBotMonitor/gradle/wrapper/gradle-wrapper.properties
new file mode 100644
index 0000000..c61a118
--- /dev/null
+++ b/android/TradeBotMonitor/gradle/wrapper/gradle-wrapper.properties
@@ -0,0 +1,7 @@
+distributionBase=GRADLE_USER_HOME
+distributionPath=wrapper/dists
+distributionUrl=https\://services.gradle.org/distributions/gradle-9.4.1-bin.zip
+networkTimeout=10000
+validateDistributionUrl=true
+zipStoreBase=GRADLE_USER_HOME
+zipStorePath=wrapper/dists
diff --git a/android/TradeBotMonitor/gradlew b/android/TradeBotMonitor/gradlew
new file mode 100644
index 0000000..739907d
--- /dev/null
+++ b/android/TradeBotMonitor/gradlew
@@ -0,0 +1,248 @@
+#!/bin/sh
+
+#
+# Copyright © 2015 the original authors.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#      https://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+#
+# SPDX-License-Identifier: Apache-2.0
+#
+
+##############################################################################
+#
+#   Gradle start up script for POSIX generated by Gradle.
+#
+#   Important for running:
+#
+#   (1) You need a POSIX-compliant shell to run this script. If your /bin/sh is
+#       noncompliant, but you have some other compliant shell such as ksh or
+#       bash, then to run this script, type that shell name before the whole
+#       command line, like:
+#
+#           ksh Gradle
+#
+#       Busybox and similar reduced shells will NOT work, because this script
+#       requires all of these POSIX shell features:
+#         * functions;
+#         * expansions «$var», «${var}», «${var:-default}», «${var+SET}»,
+#           «${var#prefix}», «${var%suffix}», and «$( cmd )»;
+#         * compound commands having a testable exit status, especially «case»;
+#         * various built-in commands including «command», «set», and «ulimit».
+#
+#   Important for patching:
+#
+#   (2) This script targets any POSIX shell, so it avoids extensions provided
+#       by Bash, Ksh, etc; in particular arrays are avoided.
+#
+#       The "traditional" practice of packing multiple parameters into a
+#       space-separated string is a well documented source of bugs and security
+#       problems, so this is (mostly) avoided, by progressively accumulating
+#       options in "$@", and eventually passing that to Java.
+#
+#       Where the inherited environment variables (DEFAULT_JVM_OPTS, JAVA_OPTS,
+#       and GRADLE_OPTS) rely on word-splitting, this is performed explicitly;
+#       see the in-line comments for details.
+#
+#       There are tweaks for specific operating systems such as AIX, CygWin,
+#       Darwin, MinGW, and NonStop.
+#
+#   (3) This script is generated from the Groovy template
+#       https://github.com/gradle/gradle/blob/2d6327017519d23b96af35865dc997fcb544fb40/platforms/jvm/plugins-application/src/main/resources/org/gradle/api/internal/plugins/unixStartScript.txt
+#       within the Gradle project.
+#
+#       You can find Gradle at https://github.com/gradle/gradle/.
+#
+##############################################################################
+
+# Attempt to set APP_HOME
+
+# Resolve links: $0 may be a link
+app_path=$0
+
+# Need this for daisy-chained symlinks.
+while
+    APP_HOME=${app_path%"${app_path##*/}"}  # leaves a trailing /; empty if no leading path
+    [ -h "$app_path" ]
+do
+    ls=$( ls -ld "$app_path" )
+    link=${ls#*' -> '}
+    case $link in             #(
+      /*)   app_path=$link ;; #(
+      *)    app_path=$APP_HOME$link ;;
+    esac
+done
+
+# This is normally unused
+# shellcheck disable=SC2034
+APP_BASE_NAME=${0##*/}
+# Discard cd standard output in case $CDPATH is set (https://github.com/gradle/gradle/issues/25036)
+APP_HOME=$( cd -P "${APP_HOME:-./}" > /dev/null && printf '%s\n' "$PWD" ) || exit
+
+# Use the maximum available, or set MAX_FD != -1 to use that value.
+MAX_FD=maximum
+
+warn () {
+    echo "$*"
+} >&2
+
+die () {
+    echo
+    echo "$*"
+    echo
+    exit 1
+} >&2
+
+# OS specific support (must be 'true' or 'false').
+cygwin=false
+msys=false
+darwin=false
+nonstop=false
+case "$( uname )" in                #(
+  CYGWIN* )         cygwin=true  ;; #(
+  Darwin* )         darwin=true  ;; #(
+  MSYS* | MINGW* )  msys=true    ;; #(
+  NONSTOP* )        nonstop=true ;;
+esac
+
+
+
+# Determine the Java command to use to start the JVM.
+if [ -n "$JAVA_HOME" ] ; then
+    if [ -x "$JAVA_HOME/jre/sh/java" ] ; then
+        # IBM's JDK on AIX uses strange locations for the executables
+        JAVACMD=$JAVA_HOME/jre/sh/java
+    else
+        JAVACMD=$JAVA_HOME/bin/java
+    fi
+    if [ ! -x "$JAVACMD" ] ; then
+        die "ERROR: JAVA_HOME is set to an invalid directory: $JAVA_HOME
+
+Please set the JAVA_HOME variable in your environment to match the
+location of your Java installation."
+    fi
+else
+    JAVACMD=java
+    if ! command -v java >/dev/null 2>&1
+    then
+        die "ERROR: JAVA_HOME is not set and no 'java' command could be found in your PATH.
+
+Please set the JAVA_HOME variable in your environment to match the
+location of your Java installation."
+    fi
+fi
+
+# Increase the maximum file descriptors if we can.
+if ! "$cygwin" && ! "$darwin" && ! "$nonstop" ; then
+    case $MAX_FD in #(
+      max*)
+        # In POSIX sh, ulimit -H is undefined. That's why the result is checked to see if it worked.
+        # shellcheck disable=SC2039,SC3045
+        MAX_FD=$( ulimit -H -n ) ||
+            warn "Could not query maximum file descriptor limit"
+    esac
+    case $MAX_FD in  #(
+      '' | soft) :;; #(
+      *)
+        # In POSIX sh, ulimit -n is undefined. That's why the result is checked to see if it worked.
+        # shellcheck disable=SC2039,SC3045
+        ulimit -n "$MAX_FD" ||
+            warn "Could not set maximum file descriptor limit to $MAX_FD"
+    esac
+fi
+
+# Collect all arguments for the java command, stacking in reverse order:
+#   * args from the command line
+#   * the main class name
+#   * -classpath
+#   * -D...appname settings
+#   * --module-path (only if needed)
+#   * DEFAULT_JVM_OPTS, JAVA_OPTS, and GRADLE_OPTS environment variables.
+
+# For Cygwin or MSYS, switch paths to Windows format before running java
+if "$cygwin" || "$msys" ; then
+    APP_HOME=$( cygpath --path --mixed "$APP_HOME" )
+
+    JAVACMD=$( cygpath --unix "$JAVACMD" )
+
+    # Now convert the arguments - kludge to limit ourselves to /bin/sh
+    for arg do
+        if
+            case $arg in                                #(
+              -*)   false ;;                            # don't mess with options #(
+              /?*)  t=${arg#/} t=/${t%%/*}              # looks like a POSIX filepath
+                    [ -e "$t" ] ;;                      #(
+              *)    false ;;
+            esac
+        then
+            arg=$( cygpath --path --ignore --mixed "$arg" )
+        fi
+        # Roll the args list around exactly as many times as the number of
+        # args, so each arg winds up back in the position where it started, but
+        # possibly modified.
+        #
+        # NB: a `for` loop captures its iteration list before it begins, so
+        # changing the positional parameters here affects neither the number of
+        # iterations, nor the values presented in `arg`.
+        shift                   # remove old arg
+        set -- "$@" "$arg"      # push replacement arg
+    done
+fi
+
+
+# Add default JVM options here. You can also use JAVA_OPTS and GRADLE_OPTS to pass JVM options to this script.
+DEFAULT_JVM_OPTS='"-Xmx64m" "-Xms64m"'
+
+# Collect all arguments for the java command:
+#   * DEFAULT_JVM_OPTS, JAVA_OPTS, and optsEnvironmentVar are not allowed to contain shell fragments,
+#     and any embedded shellness will be escaped.
+#   * For example: A user cannot expect ${Hostname} to be expanded, as it is an environment variable and will be
+#     treated as '${Hostname}' itself on the command line.
+
+set -- \
+        "-Dorg.gradle.appname=$APP_BASE_NAME" \
+        -jar "$APP_HOME/gradle/wrapper/gradle-wrapper.jar" \
+        "$@"
+
+# Stop when "xargs" is not available.
+if ! command -v xargs >/dev/null 2>&1
+then
+    die "xargs is not available"
+fi
+
+# Use "xargs" to parse quoted args.
+#
+# With -n1 it outputs one arg per line, with the quotes and backslashes removed.
+#
+# In Bash we could simply go:
+#
+#   readarray ARGS < <( xargs -n1 <<<"$var" ) &&
+#   set -- "${ARGS[@]}" "$@"
+#
+# but POSIX shell has neither arrays nor command substitution, so instead we
+# post-process each arg (as a line of input to sed) to backslash-escape any
+# character that might be a shell metacharacter, then use eval to reverse
+# that process (while maintaining the separation between arguments), and wrap
+# the whole thing up as a single "set" statement.
+#
+# This will of course break if any of these variables contains a newline or
+# an unmatched quote.
+#
+
+eval "set -- $(
+        printf '%s\n' "$DEFAULT_JVM_OPTS $JAVA_OPTS $GRADLE_OPTS" |
+        xargs -n1 |
+        sed ' s~[^-[:alnum:]+,./:=@_]~\\&~g; ' |
+        tr '\n' ' '
+    )" '"$@"'
+
+exec "$JAVACMD" "$@"
diff --git a/android/TradeBotMonitor/gradlew.bat b/android/TradeBotMonitor/gradlew.bat
new file mode 100644
index 0000000..c4bdd3a
--- /dev/null
+++ b/android/TradeBotMonitor/gradlew.bat
@@ -0,0 +1,93 @@
+@rem
+@rem Copyright 2015 the original author or authors.
+@rem
+@rem Licensed under the Apache License, Version 2.0 (the "License");
+@rem you may not use this file except in compliance with the License.
+@rem You may obtain a copy of the License at
+@rem
+@rem      https://www.apache.org/licenses/LICENSE-2.0
+@rem
+@rem Unless required by applicable law or agreed to in writing, software
+@rem distributed under the License is distributed on an "AS IS" BASIS,
+@rem WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+@rem See the License for the specific language governing permissions and
+@rem limitations under the License.
+@rem
+@rem SPDX-License-Identifier: Apache-2.0
+@rem
+
+@if "%DEBUG%"=="" @echo off
+@rem ##########################################################################
+@rem
+@rem  Gradle startup script for Windows
+@rem
+@rem ##########################################################################
+
+@rem Set local scope for the variables with windows NT shell
+if "%OS%"=="Windows_NT" setlocal
+
+set DIRNAME=%~dp0
+if "%DIRNAME%"=="" set DIRNAME=.
+@rem This is normally unused
+set APP_BASE_NAME=%~n0
+set APP_HOME=%DIRNAME%
+
+@rem Resolve any "." and ".." in APP_HOME to make it shorter.
+for %%i in ("%APP_HOME%") do set APP_HOME=%%~fi
+
+@rem Add default JVM options here. You can also use JAVA_OPTS and GRADLE_OPTS to pass JVM options to this script.
+set DEFAULT_JVM_OPTS="-Xmx64m" "-Xms64m"
+
+@rem Find java.exe
+if defined JAVA_HOME goto findJavaFromJavaHome
+
+set JAVA_EXE=java.exe
+%JAVA_EXE% -version >NUL 2>&1
+if %ERRORLEVEL% equ 0 goto execute
+
+echo. 1>&2
+echo ERROR: JAVA_HOME is not set and no 'java' command could be found in your PATH. 1>&2
+echo. 1>&2
+echo Please set the JAVA_HOME variable in your environment to match the 1>&2
+echo location of your Java installation. 1>&2
+
+goto fail
+
+:findJavaFromJavaHome
+set JAVA_HOME=%JAVA_HOME:"=%
+set JAVA_EXE=%JAVA_HOME%/bin/java.exe
+
+if exist "%JAVA_EXE%" goto execute
+
+echo. 1>&2
+echo ERROR: JAVA_HOME is set to an invalid directory: %JAVA_HOME% 1>&2
+echo. 1>&2
+echo Please set the JAVA_HOME variable in your environment to match the 1>&2
+echo location of your Java installation. 1>&2
+
+goto fail
+
+:execute
+@rem Setup the command line
+
+
+
+@rem Execute Gradle
+"%JAVA_EXE%" %DEFAULT_JVM_OPTS% %JAVA_OPTS% %GRADLE_OPTS% "-Dorg.gradle.appname=%APP_BASE_NAME%" -jar "%APP_HOME%\gradle\wrapper\gradle-wrapper.jar" %*
+
+:end
+@rem End local scope for the variables with windows NT shell
+if %ERRORLEVEL% equ 0 goto mainEnd
+
+:fail
+rem Set variable GRADLE_EXIT_CONSOLE if you need the _script_ return code instead of
+rem the _cmd.exe /c_ return code!
+set EXIT_CODE=%ERRORLEVEL%
+if %EXIT_CODE% equ 0 set EXIT_CODE=1
+if not ""=="%GRADLE_EXIT_CONSOLE%" exit %EXIT_CODE%
+exit /b %EXIT_CODE%
+
+:mainEnd
+if "%OS%"=="Windows_NT" endlocal
+
+:omega
diff --git a/crypto_spot_bot/strategy.py b/crypto_spot_bot/strategy.py
index 4706af1..6a87bb3 100644
--- a/crypto_spot_bot/strategy.py
+++ b/crypto_spot_bot/strategy.py
@@ -644,8 +644,12 @@ def _torch_forecast_entry_signal(
     expected_return = _safe_float(forecast.get("expected_return_percent"), 0.0)
     probability_up = _safe_float(forecast.get("probability_up"), 0.5)
     skill = _safe_float(forecast.get("skill"), 0.0)
-    min_edge = max(0.0, settings.time_series_min_edge_percent)
-    min_probability = _torch_min_probability(settings)
+    min_edge = max(0.0, _safe_float(forecast.get("calibrated_min_edge_percent"), settings.time_series_min_edge_percent))
+    min_probability = _clamp(
+        _safe_float(forecast.get("calibrated_min_probability_up"), _torch_min_probability(settings)),
+        0.5,
+        0.95,
+    )
     probe_min_edge = max(0.0, min(settings.time_series_probe_min_edge_percent, min_edge))
     probe_min_probability = round(
         _clamp(settings.time_series_probe_min_probability_up, min_probability, 0.85),
@@ -716,7 +720,7 @@ def _torch_forecast_entry_signal(
         and expected_return >= 0.0
         and probability_up >= rebound_model_probability_min
         and skill > 0.0
-        and confidence >= settings.time_series_min_confidence
+        and confidence >= _safe_float(forecast.get("calibrated_min_confidence"), settings.time_series_min_confidence)
     )
     fallback_rebound_entry_ok = bool(
         settings.time_series_rebound_fallback_enabled
@@ -725,7 +729,7 @@ def _torch_forecast_entry_signal(
         and quality_gate_ok
         and model_fresh_ok
         and not bool(forecast.get("block_entry", False))
-        and confidence >= settings.time_series_min_confidence
+        and confidence >= _safe_float(forecast.get("calibrated_min_confidence"), settings.time_series_min_confidence)
     )
     rebound_entry_ok = model_rebound_entry_ok or fallback_rebound_entry_ok
     if rebound_entry_ok and position_notional > 0:
@@ -896,8 +900,12 @@ def _torch_forecast_exit_signal(
     expected_return = _safe_float(forecast.get("expected_return_percent"), 0.0)
     probability_up = _safe_float(forecast.get("probability_up"), 0.5)
     skill = _safe_float(forecast.get("skill"), 0.0)
-    min_edge = max(0.0, settings.time_series_min_edge_percent)
-    min_probability = _torch_min_probability(settings)
+    min_edge = max(0.0, _safe_float(forecast.get("calibrated_min_edge_percent"), settings.time_series_min_edge_percent))
+    min_probability = _clamp(
+        _safe_float(forecast.get("calibrated_min_probability_up"), _torch_min_probability(settings)),
+        0.5,
+        0.95,
+    )
     estimated_exit_net_percent = _estimated_exit_net_percent(position, price, settings)
     min_exit_net_percent = _min_exit_net_percent(settings)
     entry_path = str(position.entry_diagnostics.get("entry_path", ""))
diff --git a/crypto_spot_bot/time_series.py b/crypto_spot_bot/time_series.py
index aa2065f..43e9e9b 100644
--- a/crypto_spot_bot/time_series.py
+++ b/crypto_spot_bot/time_series.py
@@ -160,6 +160,9 @@ class TimeSeriesForecast:
     model_created_at: str = ""
     model_age_hours: float | None = None
     model_fresh: bool = False
+    calibrated_min_edge_percent: float = 0.0
+    calibrated_min_probability_up: float = 0.0
+    calibrated_min_confidence: float = 0.0
 
     def as_dict(self) -> dict[str, Any]:
         return asdict(self)
@@ -196,8 +199,20 @@ class TimeSeriesForecaster:
             artifact,
             self.settings.time_series_model_max_age_hours,
         )
-        quality_gate = self._load_quality_gate()
+        calibration = self._load_quality_gate()
+        quality_gate = (
+            calibration.get("validation")
+            if isinstance(calibration.get("validation"), dict)
+            else calibration
+        )
         quality_gate_passed = _quality_gate_passed(quality_gate)
+        calibrated = _calibrated_thresholds(
+            calibration,
+            symbol,
+            edge=self.settings.time_series_min_edge_percent,
+            probability=self.settings.time_series_min_probability_up,
+            confidence=self.settings.time_series_min_confidence,
+        )
         entry = _torch_recurrent_entry(symbol, artifact)
         model = _torch_recurrent_model_name(symbol, artifact)
         clip = _clamp(_float_entry(entry or {}, "clip", 8.0), 1.0, 50.0)
@@ -249,7 +264,7 @@ class TimeSeriesForecaster:
             q90_percent = (math.exp(float(selected.get("q90", expected_return))) - 1) * 100
             skill = _clamp(_float_entry(entry, "skill", 0.0), -1.0, 1.0)
             horizon = int(selected.get("horizon", _entry_horizon(entry, self.settings.time_series_forecast_horizon)))
-            min_edge = max(0.0, self.settings.time_series_min_edge_percent)
+            min_edge = calibrated["edge"]
             confidence_adjustment = _confidence_adjustment(
                 expected_return_percent=expected_return_percent,
                 probability_up=probability_up,
@@ -299,6 +314,9 @@ class TimeSeriesForecaster:
                 model_created_at=model_created_at,
                 model_age_hours=model_age_hours,
                 model_fresh=model_fresh,
+                calibrated_min_edge_percent=calibrated["edge"],
+                calibrated_min_probability_up=calibrated["probability"],
+                calibrated_min_confidence=calibrated["confidence"],
             )
 
         direct_horizon = _is_direct_horizon(entry)
@@ -318,7 +336,7 @@ class TimeSeriesForecaster:
         expected_return_percent = (math.exp(expected_return) - 1) * 100
         probability_up = _normal_cdf(expected_return / max(uncertainty, 1e-9))
         skill = _clamp(_float_entry(entry, "skill", 0.0), -1.0, 1.0)
-        min_edge = max(0.0, self.settings.time_series_min_edge_percent)
+        min_edge = calibrated["edge"]
         confidence_adjustment = _confidence_adjustment(
             expected_return_percent=expected_return_percent,
             probability_up=probability_up,
@@ -364,6 +382,9 @@ class TimeSeriesForecaster:
             model_created_at=model_created_at,
             model_age_hours=model_age_hours,
             model_fresh=model_fresh,
+            calibrated_min_edge_percent=calibrated["edge"],
+            calibrated_min_probability_up=calibrated["probability"],
+            calibrated_min_confidence=calibrated["confidence"],
         )
 
     def _load_lstm_artifact(self) -> dict[str, Any]:
@@ -400,8 +421,7 @@ class TimeSeriesForecaster:
             data = json.loads(path.read_text(encoding="utf-8"))
         except (OSError, json.JSONDecodeError):
             data = {}
-        validation = data.get("validation") if isinstance(data, dict) else {}
-        self._quality_gate = validation if isinstance(validation, dict) else {}
+        self._quality_gate = data if isinstance(data, dict) else {}
         self._calibration_mtime = stat.st_mtime
         return self._quality_gate
 
@@ -443,6 +463,9 @@ def _empty_forecast(enabled: bool, reason: str) -> TimeSeriesForecast:
 def _quality_gate_passed(quality_gate: dict[str, Any]) -> bool | None:
     if not quality_gate:
         return None
+    validation = quality_gate.get("validation")
+    if isinstance(validation, dict):
+        return _quality_gate_passed(validation)
     if "passed" in quality_gate:
         return bool(quality_gate.get("passed"))
     status = str(quality_gate.get("status", "")).strip().lower()
@@ -453,6 +476,26 @@ def _quality_gate_passed(quality_gate: dict[str, Any]) -> bool | None:
     return None
 
 
+def _calibrated_thresholds(
+    calibration: dict[str, Any],
+    symbol: str | None,
+    *,
+    edge: float,
+    probability: float,
+    confidence: float,
+) -> dict[str, float]:
+    recommended = calibration.get("recommended") if isinstance(calibration, dict) else None
+    per_symbol = calibration.get("symbol_recommendations") if isinstance(calibration, dict) else None
+    if symbol and isinstance(per_symbol, dict) and isinstance(per_symbol.get(symbol.upper()), dict):
+        recommended = per_symbol[symbol.upper()]
+    row = recommended if isinstance(recommended, dict) else {}
+    return {
+        "edge": max(0.0, float(row.get("edge", edge) or edge)),
+        "probability": _clamp(float(row.get("probability", probability) or probability), 0.5, 0.95),
+        "confidence": _clamp(float(row.get("confidence", confidence) or confidence), 0.0, 1.0),
+    }
+
+
 def _model_freshness(artifact: dict[str, Any], max_age_hours: float) -> tuple[str, float | None, bool]:
     raw = str(artifact.get("created_at", "")).strip() if isinstance(artifact, dict) else ""
     if not raw:
@@ -525,6 +568,8 @@ def _feature_context(
 def _feature_value(name: str, candles: list[Candle], index: int, candle: Candle, context: dict[str, Any]) -> float:
     close = max(float(candle.close), 1e-12)
     previous = candles[index - 1] if index >= 1 else candle
+    if name.startswith("symbol_is_"):
+        return 1.0 if context.get("symbol") == name.removeprefix("symbol_is_").upper() else 0.0
     if name == "return_1":
         return _log_change(candle.close, previous.close)
     if name == "return_3":
@@ -990,6 +1035,31 @@ def _torch_recurrent_predict(
     model_name = _torch_recurrent_model_name(symbol, artifact)
     if not entry or not model_name:
         return None
+    ensemble_members = entry.get("ensemble_members")
+    if isinstance(ensemble_members, list) and ensemble_members:
+        predictions: list[float | dict[str, Any]] = []
+        for member in ensemble_members:
+            if not isinstance(member, dict):
+                continue
+            member_entry = {**entry, **member}
+            member_entry.pop("ensemble_members", None)
+            member_entry.pop("ensemble_size", None)
+            member_artifact: dict[str, Any] = {"type": "pytorch_recurrent_forecaster"}
+            if symbol:
+                member_artifact["symbols"] = {symbol.upper(): member_entry}
+            else:
+                member_artifact["default"] = member_entry
+            prediction = _torch_recurrent_predict(
+                returns,
+                symbol,
+                member_artifact,
+                feature_rows=feature_rows,
+                closes=closes,
+                candles=candles,
+            )
+            if isinstance(prediction, (int, float, dict)):
+                predictions.append(prediction)
+        return _average_ensemble_predictions(predictions)
     lookback = int(_clamp(_float_entry(entry, "lookback", 0.0), 4.0, 512.0))
     hidden_size = int(_clamp(_float_entry(entry, "hidden_size", 0.0), 1.0, 512.0))
     num_layers = int(_clamp(_float_entry(entry, "num_layers", 1.0), 1.0, 8.0))
@@ -1050,6 +1120,33 @@ def _torch_recurrent_predict(
     return _clamp(prediction, -cap, cap)
 
 
+def _average_ensemble_predictions(predictions: list[float | dict[str, Any]]) -> float | dict[str, Any] | None:
+    if not predictions:
+        return None
+    numeric = [float(value) for value in predictions if isinstance(value, (int, float))]
+    if numeric:
+        return sum(numeric) / len(numeric)
+    mappings = [value for value in predictions if isinstance(value, dict)]
+    if not mappings:
+        return None
+    first = mappings[0]
+    output: dict[str, Any] = {}
+    for key, value in first.items():
+        if key == "horizons" and isinstance(value, dict):
+            horizons: dict[str, Any] = {}
+            for horizon, row in value.items():
+                rows = [item.get("horizons", {}).get(horizon) for item in mappings]
+                rows = [item for item in rows if isinstance(item, dict)]
+                if rows:
+                    horizons[horizon] = _average_ensemble_predictions(rows)
+            output[key] = horizons
+            continue
+        values = [item.get(key) for item in mappings]
+        finite = [float(item) for item in values if isinstance(item, (int, float)) and math.isfinite(float(item))]
+        output[key] = sum(finite) / len(finite) if finite else value
+    return output
+
+
 def _torch_head_outputs(context: list[float], entry: dict[str, Any], hidden_size: int) -> list[float]:
     context = _apply_context_norm(context, entry)
     raw_weight = entry.get("head_weight")
diff --git a/crypto_spot_bot/training_coordination.py b/crypto_spot_bot/training_coordination.py
index 82e84dc..4969020 100644
--- a/crypto_spot_bot/training_coordination.py
+++ b/crypto_spot_bot/training_coordination.py
@@ -215,6 +215,10 @@ class TrainingCoordinator:
             job["message"] = str(payload.get("message") or "")
             if isinstance(payload.get("summary"), dict):
                 job["summary"] = payload["summary"]
+                if isinstance(payload["summary"].get("accepted"), bool):
+                    job["model_decision"] = (
+                        "accepted" if payload["summary"]["accepted"] else "rejected"
+                    )
             self._save_state(state)
             return {"ok": True, "job": job, "status": self._public_status(state)}
 
@@ -294,10 +298,11 @@ class TrainingCoordinator:
         os.replace(tmp, self.state_path)
 
     def _worker_from_payload(self, payload: dict[str, Any]) -> dict[str, Any]:
+        worker_id = str(payload.get("worker_id") or payload.get("id") or "windows-training-host").strip()
         return {
-            "id": str(payload.get("worker_id") or payload.get("id") or "windows-training-host"),
-            "name": str(payload.get("name") or "DESKTOP-TMFDL0H"),
-            "path": str(payload.get("path") or "C:\\Repos\\TradeBot"),
+            "id": worker_id,
+            "name": str(payload.get("name") or worker_id).strip(),
+            "path": str(payload.get("path") or "").strip(),
             "version": str(payload.get("version") or "1"),
             "last_seen_at": _now(),
         }
diff --git a/tests/test_calibrate_thresholds.py b/tests/test_calibrate_thresholds.py
new file mode 100644
index 0000000..0583ff8
--- /dev/null
+++ b/tests/test_calibrate_thresholds.py
@@ -0,0 +1,87 @@
+from __future__ import annotations
+
+from tools.calibrate_torch_thresholds import (
+    CalibrationResult,
+    ForecastRecord,
+    _apply_platt_calibration,
+    _choose_recommendation,
+    _fit_platt_calibration,
+    _entry_validation_skill,
+)
+
+
+def _result(*, trades: int, average: float, total: float, profit_factor: float) -> CalibrationResult:
+    return CalibrationResult(
+        edge=0.05,
+        probability=0.52,
+        confidence=0.4,
+        trades=trades,
+        wins=max(0, trades // 2),
+        win_rate=0.5,
+        total_net_percent=total,
+        average_net_percent=average,
+        max_drawdown_percent=1.0,
+        profit_factor=profit_factor,
+        score=1.0,
+    )
+
+
+def _record(index: int, probability: float, future: float) -> ForecastRecord:
+    return ForecastRecord(
+        symbol="BTCUSDT",
+        index=index,
+        timestamp=index,
+        close=100.0,
+        high=101.0,
+        low=99.0,
+        next_open=100.0,
+        next_timestamp=index + 1,
+        atr=1.0,
+        expected_percent=0.1,
+        probability_up=probability,
+        confidence=0.5,
+        skill=0.1,
+        q50_percent=0.1,
+        block_entry=False,
+        future_net_percent=future,
+        benchmark_entry=False,
+        benchmark_exit=False,
+    )
+
+
+def test_calibration_does_not_fallback_to_too_few_trades() -> None:
+    selected = _choose_recommendation(
+        [_result(trades=1, average=2.0, total=2.0, profit_factor=999.0)],
+        min_trades=30,
+    )
+
+    assert selected is None
+
+
+def test_calibration_selects_only_viable_result() -> None:
+    viable = _result(trades=30, average=0.2, total=6.0, profit_factor=1.4)
+
+    assert _choose_recommendation([viable], min_trades=30) is viable
+
+
+def test_platt_calibration_learns_probability_direction_from_train_records() -> None:
+    records = [
+        _record(index, 0.8 if index % 2 else 0.2, -1.0 if index % 2 else 1.0)
+        for index in range(100)
+    ]
+
+    calibration = _fit_platt_calibration(records)
+    calibrated = _apply_platt_calibration(
+        [_record(101, 0.8, -1.0), _record(102, 0.2, 1.0)],
+        calibration,
+    )
+
+    assert calibration["slope"] < 0
+    assert calibrated[0].probability_up < calibrated[1].probability_up
+
+
+def test_entry_quality_never_falls_back_to_holdout_skill() -> None:
+    entry = {"validation_skill": 0.12, "skill": 0.99, "holdout_skill": 0.99}
+
+    assert _entry_validation_skill(entry) == 0.12
+    assert _entry_validation_skill({"skill": 0.99, "holdout_skill": 0.99}) == 0.0
diff --git a/tests/test_time_series.py b/tests/test_time_series.py
index 9c0a516..10a1657 100644
--- a/tests/test_time_series.py
+++ b/tests/test_time_series.py
@@ -304,6 +304,60 @@ def test_time_series_forecaster_attaches_quality_gate(make_settings, tmp_path) -
     assert forecast.quality_gate["status"] == "fail"
 
 
+def test_time_series_forecaster_uses_symbol_calibration(make_settings, tmp_path) -> None:
+    artifact_path = tmp_path / "lstm_forecaster.json"
+    _write_torch_gru_artifact(artifact_path, head_bias=0.2)
+    (tmp_path / "torch_threshold_calibration.json").write_text(
+        json.dumps(
+            {
+                "validation": {"status": "pass", "passed": True},
+                "recommended": {"edge": 0.08, "probability": 0.52, "confidence": 0.4},
+                "symbol_recommendations": {
+                    "BTCUSDT": {"edge": 0.03, "probability": 0.55, "confidence": 0.45}
+                },
+            }
+        ),
+        encoding="utf-8",
+    )
+    settings = make_settings(
+        tmp_path,
+        time_series_lstm_model_path=artifact_path,
+        time_series_forecast_horizon=1,
+    )
+
+    forecast = TimeSeriesForecaster(settings).forecast(
+        _candles_from_returns([0.0001] * 140), symbol="BTCUSDT"
+    )
+
+    assert forecast.calibrated_min_edge_percent == 0.03
+    assert forecast.calibrated_min_probability_up == 0.55
+    assert forecast.calibrated_min_confidence == 0.45
+
+
+def test_time_series_forecaster_averages_ensemble_members(make_settings, tmp_path) -> None:
+    artifact_path = tmp_path / "lstm_forecaster.json"
+    _write_torch_gru_artifact(artifact_path, head_bias=0.9)
+    artifact = json.loads(artifact_path.read_text(encoding="utf-8"))
+    entry = artifact["symbols"]["BTCUSDT"]
+    entry["ensemble_members"] = [
+        {"state_dict": entry["state_dict"], "head_weight": [0.0, 0.0], "head_bias": bias}
+        for bias in (0.1, 0.3)
+    ]
+    artifact_path.write_text(json.dumps(artifact), encoding="utf-8")
+    settings = make_settings(
+        tmp_path,
+        time_series_lstm_model_path=artifact_path,
+        time_series_forecast_horizon=1,
+    )
+
+    forecast = TimeSeriesForecaster(settings).forecast(
+        _candles_from_returns([0.0001] * 140), symbol="BTCUSDT"
+    )
+
+    assert forecast.usable is True
+    assert 0.015 <= forecast.expected_return_percent <= 0.025
+
+
 def test_time_series_forecaster_reads_multifeature_direct_horizon_artifact(make_settings, tmp_path) -> None:
     artifact_path = tmp_path / "lstm_forecaster.json"
     _write_multifeature_torch_gru_artifact(artifact_path, head_bias=0.2)
diff --git a/tests/test_training_coordination.py b/tests/test_training_coordination.py
index 1686a0b..264c4f7 100644
--- a/tests/test_training_coordination.py
+++ b/tests/test_training_coordination.py
@@ -48,6 +48,42 @@ def test_training_coordinator_preserves_boolean_resume_candidate_parameter(tmp_p
     assert requested["job"]["parameters"] == {"resume_candidate": True}
 
 
+def test_training_coordinator_reports_worker_identity_from_heartbeat(tmp_path) -> None:
+    coordinator = TrainingCoordinator(tmp_path)
+
+    heartbeat = coordinator.heartbeat(
+        {
+            "worker_id": "SEVENHILL:G:\\Repos\\TradeBot",
+            "name": "SEVENHILL",
+            "path": "G:\\Repos\\TradeBot",
+        }
+    )
+
+    assert heartbeat["worker"]["name"] == "SEVENHILL"
+    assert heartbeat["worker"]["path"] == "G:\\Repos\\TradeBot"
+    assert heartbeat["status"]["worker"] == heartbeat["worker"]
+
+
+def test_training_coordinator_records_rejected_candidate_as_completed_training(tmp_path) -> None:
+    coordinator = TrainingCoordinator(tmp_path)
+    job = coordinator.request_retrain({"source": "android"})["job"]
+    coordinator.claim({"worker_id": "worker-1"})
+
+    completed = coordinator.complete(
+        job["id"],
+        {
+            "success": True,
+            "message": "training completed; candidate rejected by quality gate",
+            "summary": {"accepted": False, "reason": "candidate_failed_honest_validation"},
+        },
+    )
+
+    assert completed["job"]["status"] == "completed"
+    assert completed["job"]["phase"] == "completed"
+    assert completed["job"]["progress_percent"] == 100
+    assert completed["job"]["model_decision"] == "rejected"
+
+
 def test_training_coordinator_accepts_chunked_artifact_upload(tmp_path) -> None:
     coordinator = TrainingCoordinator(tmp_path)
     job = coordinator.request_retrain({"source": "test"})["job"]
diff --git a/tools/calibrate_torch_thresholds.py b/tools/calibrate_torch_thresholds.py
index 5dda0e1..62c45bf 100644
--- a/tools/calibrate_torch_thresholds.py
+++ b/tools/calibrate_torch_thresholds.py
@@ -6,7 +6,7 @@ import json
 import math
 import sys
 import time
-from dataclasses import dataclass
+from dataclasses import dataclass, replace
 from pathlib import Path
 from typing import Any
 
@@ -130,13 +130,15 @@ def main() -> None:
     if not records:
         raise SystemExit("No forecast records could be built for calibration.")
 
-    results = _calibrate(
+    results = _calibrate_strategy(
         records,
         edges=_float_grid(args.edge_grid),
         probabilities=_float_grid(args.probability_grid),
         confidences=_float_grid(args.confidence_grid),
         min_trades=args.min_trades,
         horizon=horizon,
+        round_trip_cost=round_trip_cost,
+        settings=settings,
     )
     if not results:
         raise SystemExit("No calibration result produced trades. Use wider grids or more history.")
@@ -156,6 +158,33 @@ def main() -> None:
         round_trip_cost=round_trip_cost,
         settings=settings,
     )
+    symbol_recommendations: dict[str, dict[str, Any]] = {}
+    for symbol in symbols:
+        symbol_records = [record for record in records if record.symbol == symbol]
+        symbol_results = _calibrate_strategy(
+            symbol_records,
+            edges=_float_grid(args.edge_grid),
+            probabilities=_float_grid(args.probability_grid),
+            confidences=_float_grid(args.confidence_grid),
+            min_trades=max(3, min(args.min_trades, len(symbol_records) // 8)),
+            horizon=horizon,
+            round_trip_cost=round_trip_cost,
+            settings=settings,
+        )
+        symbol_selected = _choose_recommendation(
+            symbol_results,
+            min_trades=max(3, min(args.min_trades, len(symbol_records) // 8)),
+        ) if symbol_results else None
+        if symbol_selected is not None:
+            symbol_recommendations[symbol] = _result_dict(symbol_selected)
+    calibration_insufficient = recommended is None
+    if recommended is None:
+        recommended = _empty_recommendation(
+            _float_grid(args.edge_grid),
+            _float_grid(args.probability_grid),
+            _float_grid(args.confidence_grid),
+        )
+        full_backtest = {**_stats([]), "trades_detail": [], "symbol_breakdown": []}
     print("\nRECOMMENDED")
     print(_result_line(recommended))
     print("\nFULL_REPLAY")
@@ -188,6 +217,16 @@ def main() -> None:
         min_profit_factor=args.min_oos_profit_factor,
         min_benchmark_edge=args.min_benchmark_edge_percent,
     )
+    deployment_recommended = recommended
+    deployment_symbol_recommendations = symbol_recommendations
+    if walk_forward.get("folds"):
+        last_fold = walk_forward["folds"][-1]
+        fold_thresholds = last_fold.get("thresholds")
+        if isinstance(fold_thresholds, dict):
+            deployment_recommended = _result_from_dict(fold_thresholds)
+        fold_symbols = last_fold.get("symbol_thresholds")
+        if isinstance(fold_symbols, dict):
+            deployment_symbol_recommendations = fold_symbols
     print("\nWALK_FORWARD")
     print(json.dumps(walk_forward["summary"], ensure_ascii=False, sort_keys=True))
     print("\nBENCHMARK")
@@ -206,7 +245,9 @@ def main() -> None:
             "artifact_sha256": artifact_sha256,
             "artifact": _artifact_summary(artifact),
             "records_by_symbol": per_symbol_counts,
-            "recommended": _result_dict(recommended),
+            "recommended": _result_dict(deployment_recommended),
+            "calibration_insufficient": calibration_insufficient,
+            "symbol_recommendations": deployment_symbol_recommendations,
             "full_replay": full_backtest,
             "walk_forward": walk_forward,
             "benchmark": benchmark,
@@ -306,7 +347,9 @@ def _forecast_records(
         return batched_records
 
     records: list[ForecastRecord] = []
-    skill = float(entry.get("skill", 0.0) or 0.0)
+    # Entry eligibility may use validation-derived quality only. Holdout metrics
+    # belong exclusively to the final quality gate and cannot influence replay.
+    skill = _entry_validation_skill(entry)
     for index in range(start, max(start, end)):
         prediction = _torch_recurrent_predict(
             _log_returns(closes[: index + 1]),
@@ -394,7 +437,7 @@ def _batch_forecast_records(
         return []
 
     records: list[ForecastRecord] = []
-    skill = float(entry.get("skill", 0.0) or 0.0)
+    skill = _entry_validation_skill(entry)
     model.eval()
     with torch.no_grad():
         for offset in range(0, len(indices), max(1, batch_size)):
@@ -460,6 +503,10 @@ def _batch_forecast_records(
 
 
 def _build_torch_model(entry: dict[str, Any], model_name: str) -> Any | None:
+    if isinstance(entry.get("ensemble_members"), list) and entry["ensemble_members"]:
+        # Ensemble inference is handled by the shared pure-Python runtime so
+        # calibration and production use the exact same averaging path.
+        return None
     if torch is None or RecurrentReturnModel is None:
         return None
     architecture = "lstm" if model_name == "torch_lstm" else "gru" if model_name == "torch_gru" else ""
@@ -590,6 +637,7 @@ def _full_backtest(
     round_trip_cost: float,
     settings: Any,
     detail_limit: int = 50,
+    symbol_thresholds: dict[str, CalibrationResult] | None = None,
 ) -> dict[str, Any]:
     positions: dict[str, dict[str, Any]] = {}
     trades: list[float] = []
@@ -600,6 +648,7 @@ def _full_backtest(
     stop_loss_exit_enabled = bool(getattr(settings, "stop_loss_exit_enabled", True))
     atr_multiplier = max(0.5, min(10.0, float(settings.atr_trailing_multiplier)))
     for record in sorted(records, key=lambda item: (item.timestamp, item.symbol)):
+        active_thresholds = (symbol_thresholds or {}).get(record.symbol, thresholds)
         position = positions.get(record.symbol)
         if position is not None:
             position["highest"] = max(position["highest"], record.high)
@@ -616,8 +665,8 @@ def _full_backtest(
                 and (stop_loss_exit_enabled or atr_stop_level > position["entry_price"])
             )
             weak_forecast = (
-                record.expected_percent < thresholds.edge
-                or record.probability_up < thresholds.probability
+                record.expected_percent < active_thresholds.edge
+                or record.probability_up < active_thresholds.probability
                 or record.skill <= 0.0
             )
             exit_reason = ""
@@ -633,7 +682,7 @@ def _full_backtest(
             elif atr_stop:
                 exit_reason = "atr_trailing_stop"
                 exit_price = float(atr_stop_level)
-            elif (record.expected_percent <= 0.0 or record.probability_up <= 0.50 or _candidate_blocks(record, thresholds.edge)):
+            elif (record.expected_percent <= 0.0 or record.probability_up <= 0.50 or _candidate_blocks(record, active_thresholds.edge)):
                 exit_reason = "forecast_negative"
             elif weak_forecast and net_percent >= 0:
                 exit_reason = "forecast_weak_profit_lock"
@@ -659,7 +708,7 @@ def _full_backtest(
 
         if record.symbol in positions:
             continue
-        if _candidate_allows(record, thresholds.edge, thresholds.probability, thresholds.confidence):
+        if _candidate_allows(record, active_thresholds.edge, active_thresholds.probability, active_thresholds.confidence):
             positions[record.symbol] = {
                 "entry_price": record.next_open,
                 "entry_index": record.index + 1,
@@ -815,24 +864,49 @@ def _walk_forward(
         test_end = timestamps[(fold + 1) * fold_size - 1] if fold < folds - 1 else timestamps[-1]
         train = [record for record in ordered if record.timestamp < test_start]
         test = [record for record in ordered if test_start <= record.timestamp <= test_end]
-        train_results = _calibrate(
-            train,
+        probability_calibration = _fit_platt_calibration(train)
+        calibrated_train = _apply_platt_calibration(train, probability_calibration)
+        calibrated_test = _apply_platt_calibration(test, probability_calibration)
+        train_results = _calibrate_strategy(
+            calibrated_train,
             edges=edges,
             probabilities=probabilities,
             confidences=confidences,
             min_trades=max(4, min_trades // 2),
             horizon=horizon,
+            round_trip_cost=round_trip_cost,
+            settings=settings,
         )
         if not train_results:
             continue
         selected = _choose_recommendation(train_results, min_trades=max(4, min_trades // 2))
+        if selected is None:
+            continue
+        symbol_thresholds: dict[str, CalibrationResult] = {}
+        train_symbols = sorted({record.symbol for record in calibrated_train})
+        symbol_min_trades = max(3, min_trades // max(2, len(train_symbols) * 2))
+        for symbol in train_symbols:
+            symbol_results = _calibrate_strategy(
+                [record for record in calibrated_train if record.symbol == symbol],
+                edges=edges,
+                probabilities=probabilities,
+                confidences=confidences,
+                min_trades=symbol_min_trades,
+                horizon=horizon,
+                round_trip_cost=round_trip_cost,
+                settings=settings,
+            )
+            symbol_selected = _choose_recommendation(symbol_results, min_trades=symbol_min_trades) if symbol_results else None
+            if symbol_selected is not None:
+                symbol_thresholds[symbol] = symbol_selected
         test_backtest = _full_backtest(
-            test,
+            calibrated_test,
             selected,
             horizon=horizon,
             round_trip_cost=round_trip_cost,
             settings=settings,
             detail_limit=0,
+            symbol_thresholds=symbol_thresholds,
         )
         test_rows = test_backtest.get("trades_detail", [])
         test_trades = [float(row.get("net_percent", 0.0) or 0.0) for row in test_rows if isinstance(row, dict)]
@@ -844,6 +918,10 @@ def _walk_forward(
                 "train_records": len(train),
                 "test_records": len(test),
                 "thresholds": _result_dict(selected),
+                "symbol_thresholds": {
+                    symbol: _result_dict(value) for symbol, value in symbol_thresholds.items()
+                },
+                "probability_calibration": probability_calibration,
                 "test": {key: value for key, value in test_backtest.items() if key != "trades_detail"},
             }
         )
@@ -980,6 +1058,11 @@ def _candidate_blocks(record: ForecastRecord, edge: float) -> bool:
     )
 
 
+def _entry_validation_skill(entry: dict[str, Any]) -> float:
+    value = entry.get("validation_skill")
+    return float(value) if isinstance(value, (int, float)) and math.isfinite(float(value)) else 0.0
+
+
 def _candidate_allows(record: ForecastRecord, edge: float, probability: float, confidence: float) -> bool:
     dynamic_confidence = _forecast_confidence(record.expected_percent, record.probability_up, record.skill, edge)
     return (
@@ -1146,6 +1229,101 @@ def _calibrate(
     return results
 
 
+def _calibrate_strategy(
+    records: list[ForecastRecord],
+    *,
+    edges: list[float],
+    probabilities: list[float],
+    confidences: list[float],
+    min_trades: int,
+    horizon: int,
+    round_trip_cost: float,
+    settings: Any,
+) -> list[CalibrationResult]:
+    results: list[CalibrationResult] = []
+    for edge in edges:
+        for probability in probabilities:
+            for confidence in confidences:
+                thresholds = CalibrationResult(
+                    edge=edge,
+                    probability=probability,
+                    confidence=confidence,
+                    trades=0,
+                    wins=0,
+                    win_rate=0.0,
+                    total_net_percent=0.0,
+                    average_net_percent=0.0,
+                    max_drawdown_percent=0.0,
+                    profit_factor=0.0,
+                    score=0.0,
+                )
+                replay = _full_backtest(
+                    records,
+                    thresholds,
+                    horizon=horizon,
+                    round_trip_cost=round_trip_cost,
+                    settings=settings,
+                    detail_limit=0,
+                )
+                trades = int(replay.get("trades", 0) or 0)
+                if trades <= 0:
+                    continue
+                wins = int(replay.get("wins", 0) or 0)
+                total = float(replay.get("total_net_percent", 0.0) or 0.0)
+                average = float(replay.get("avg_net_percent", 0.0) or 0.0)
+                drawdown = float(replay.get("max_drawdown_percent", 0.0) or 0.0)
+                profit_factor = float(replay.get("profit_factor", 0.0) or 0.0)
+                trade_factor = min(1.0, trades / max(1, min_trades))
+                score = (
+                    average * trade_factor
+                    + total * 0.015
+                    - drawdown * 0.03
+                    + (wins / trades) * 0.04
+                )
+                results.append(
+                    CalibrationResult(
+                        edge=edge,
+                        probability=probability,
+                        confidence=confidence,
+                        trades=trades,
+                        wins=wins,
+                        win_rate=wins / trades,
+                        total_net_percent=total,
+                        average_net_percent=average,
+                        max_drawdown_percent=drawdown,
+                        profit_factor=profit_factor,
+                        score=score,
+                    )
+                )
+    results.sort(
+        key=lambda item: (
+            item.score,
+            item.average_net_percent,
+            item.total_net_percent,
+            item.profit_factor,
+            item.trades,
+        ),
+        reverse=True,
+    )
+    return results
+
+
+def _result_from_dict(value: dict[str, Any]) -> CalibrationResult:
+    return CalibrationResult(
+        edge=float(value.get("edge", 0.1) or 0.1),
+        probability=float(value.get("probability", 0.7) or 0.7),
+        confidence=float(value.get("confidence", 0.4) or 0.4),
+        trades=int(value.get("trades", 0) or 0),
+        wins=int(value.get("wins", 0) or 0),
+        win_rate=float(value.get("win_rate", 0.0) or 0.0),
+        total_net_percent=float(value.get("total_net_percent", 0.0) or 0.0),
+        average_net_percent=float(value.get("average_net_percent", 0.0) or 0.0),
+        max_drawdown_percent=float(value.get("max_drawdown_percent", 0.0) or 0.0),
+        profit_factor=float(value.get("profit_factor", 0.0) or 0.0),
+        score=float(value.get("score", 0.0) or 0.0),
+    )
+
+
 def _selected_trades(
     records: list[ForecastRecord],
     edge: float,
@@ -1164,7 +1342,7 @@ def _selected_trades(
     return trades
 
 
-def _choose_recommendation(results: list[CalibrationResult], *, min_trades: int) -> CalibrationResult:
+def _choose_recommendation(results: list[CalibrationResult], *, min_trades: int) -> CalibrationResult | None:
     viable = [
         result
         for result in results
@@ -1173,7 +1351,67 @@ def _choose_recommendation(results: list[CalibrationResult], *, min_trades: int)
         and result.total_net_percent > 0
         and result.profit_factor >= 1.05
     ]
-    return viable[0] if viable else results[0]
+    return viable[0] if viable else None
+
+
+def _empty_recommendation(
+    edges: list[float], probabilities: list[float], confidences: list[float]
+) -> CalibrationResult:
+    return CalibrationResult(
+        edge=max(edges or [1.0]),
+        probability=max(probabilities or [0.95]),
+        confidence=max(confidences or [1.0]),
+        trades=0,
+        wins=0,
+        win_rate=0.0,
+        total_net_percent=0.0,
+        average_net_percent=0.0,
+        max_drawdown_percent=0.0,
+        profit_factor=0.0,
+        score=-1.0,
+    )
+
+
+def _fit_platt_calibration(records: list[ForecastRecord]) -> dict[str, float]:
+    samples = [
+        (
+            math.log(_clamp(record.probability_up, 1e-5, 1.0 - 1e-5) / (1.0 - _clamp(record.probability_up, 1e-5, 1.0 - 1e-5))),
+            1.0 if record.future_net_percent > 0 else 0.0,
+        )
+        for record in records
+    ]
+    if len(samples) < 30:
+        return {"slope": 1.0, "intercept": 0.0, "samples": float(len(samples))}
+    slope = 1.0
+    intercept = 0.0
+    learning_rate = 0.05
+    for _ in range(300):
+        grad_slope = 0.0
+        grad_intercept = 0.0
+        for logit, target in samples:
+            probability = 1.0 / (1.0 + math.exp(-_clamp(slope * logit + intercept, -30.0, 30.0)))
+            error = probability - target
+            grad_slope += error * logit
+            grad_intercept += error
+        grad_slope = grad_slope / len(samples) + 0.001 * (slope - 1.0)
+        grad_intercept /= len(samples)
+        slope -= learning_rate * grad_slope
+        intercept -= learning_rate * grad_intercept
+    return {"slope": round(slope, 8), "intercept": round(intercept, 8), "samples": float(len(samples))}
+
+
+def _apply_platt_calibration(
+    records: list[ForecastRecord], calibration: dict[str, float]
+) -> list[ForecastRecord]:
+    slope = float(calibration.get("slope", 1.0))
+    intercept = float(calibration.get("intercept", 0.0))
+    output: list[ForecastRecord] = []
+    for record in records:
+        probability = _clamp(record.probability_up, 1e-5, 1.0 - 1e-5)
+        logit = math.log(probability / (1.0 - probability))
+        calibrated = 1.0 / (1.0 + math.exp(-_clamp(slope * logit + intercept, -30.0, 30.0)))
+        output.append(replace(record, probability_up=calibrated))
+    return output
 
 
 def _choose_replay_recommendation(
@@ -1185,8 +1423,10 @@ def _choose_replay_recommendation(
     horizon: int,
     round_trip_cost: float,
     settings: Any,
-) -> tuple[CalibrationResult, dict[str, Any]]:
+) -> tuple[CalibrationResult | None, dict[str, Any]]:
     fallback = _choose_recommendation(results, min_trades=min_trades)
+    if fallback is None:
+        return None, {**_stats([]), "trades_detail": [], "symbol_breakdown": []}
     fallback_replay = _full_backtest(records, fallback, horizon=horizon, round_trip_cost=round_trip_cost, settings=settings)
     if min_full_replay_trades <= 0:
         return fallback, fallback_replay
@@ -1207,7 +1447,7 @@ def _choose_replay_recommendation(
         viable.append((result, replay))
 
     if not viable:
-        return fallback, fallback_replay
+        return None, fallback_replay
     viable.sort(
         key=lambda item: (
             item[0].score,
diff --git a/tools/install_windows_training_agent.ps1 b/tools/install_windows_training_agent.ps1
index 5e83f0c..162b616 100644
--- a/tools/install_windows_training_agent.ps1
+++ b/tools/install_windows_training_agent.ps1
@@ -6,6 +6,7 @@ param(
     [int]$PollSeconds = 10,
     [int]$WatchdogMinutes = 5,
     [string]$RepoRoot = "",
+    [string]$CredentialPath = "",
     [switch]$StartNow,
     [switch]$KeepLegacyRetrainer
 )
@@ -15,105 +16,153 @@ $ErrorActionPreference = "Stop"
 if (-not $RepoRoot) {
     $RepoRoot = (Resolve-Path (Join-Path $PSScriptRoot "..")).Path
 }
-$Agent = Join-Path $RepoRoot "tools\windows_training_agent.py"
-if (-not (Test-Path $Agent)) {
-    throw "Windows training agent not found: $Agent"
+if (-not $CredentialPath) {
+    $CredentialPath = Join-Path $env:LOCALAPPDATA "TradeBot\training-agent.token"
 }
 
-function Resolve-Python {
-    $venvPython = Join-Path $RepoRoot ".venv\Scripts\python.exe"
-    if (Test-Path $venvPython) {
-        return $venvPython
-    }
-
-    $userPython = Join-Path $env:LOCALAPPDATA "Programs\TradeBotPython312\python.exe"
-    if (Test-Path $userPython) {
-        return $userPython
-    }
-
-    foreach ($candidate in @("python.exe", "python")) {
-        $command = Get-Command $candidate -ErrorAction SilentlyContinue
-        if ($command) {
-            return $command.Source
-        }
-    }
-    throw "Python was not found. Create .venv or install Python 3.12."
-}
-
-function Resolve-WindowlessPython {
-    $python = Resolve-Python
-    $pythonw = Join-Path (Split-Path -Parent $python) "pythonw.exe"
-    if (Test-Path $pythonw) {
-        return $pythonw
-    }
-    return $python
+$runner = Join-Path $RepoRoot "tools\run_windows_training_agent.ps1"
+if (-not (Test-Path -LiteralPath $runner)) {
+    throw "Windows training agent runner not found: $runner"
 }
 
+$credentialDirectory = Split-Path -Parent $CredentialPath
+New-Item -ItemType Directory -Path $credentialDirectory -Force | Out-Null
 if ($ApiAuth) {
-    [Environment]::SetEnvironmentVariable("TRADEBOT_API_AUTH", $ApiAuth, "User")
-    $env:TRADEBOT_API_AUTH = $ApiAuth
+    $ApiAuth.Trim() |
+        ConvertTo-SecureString -AsPlainText -Force |
+        ConvertFrom-SecureString |
+        Set-Content -LiteralPath $CredentialPath -Encoding UTF8
 }
+if (-not (Test-Path -LiteralPath $CredentialPath)) {
+    throw "ApiAuth is required for the first installation."
+}
+
+# Remove the legacy plaintext secret from the user environment. The new runner
+# decrypts the DPAPI-protected credential only inside the agent process tree.
+[Environment]::SetEnvironmentVariable("TRADEBOT_API_AUTH", $null, "User")
+Remove-Item Env:TRADEBOT_API_AUTH -ErrorAction SilentlyContinue
 [Environment]::SetEnvironmentVariable("TRADEBOT_API_BASE_URL", $ApiBaseUrl, "User")
 [Environment]::SetEnvironmentVariable("TRADEBOT_TRAINING_WORKER_NAME", $env:COMPUTERNAME, "User")
-$env:TRADEBOT_API_BASE_URL = $ApiBaseUrl
-$env:TRADEBOT_TRAINING_WORKER_NAME = $env:COMPUTERNAME
 
 if (-not $KeepLegacyRetrainer) {
     foreach ($legacyName in @("TradeBot PyTorch Forecaster Retrainer", "TradeBot LSTM Retrainer")) {
-        $legacyTask = Get-ScheduledTask -TaskName $legacyName -ErrorAction SilentlyContinue
-        if ($legacyTask) {
-            Unregister-ScheduledTask -TaskName $legacyName -Confirm:$false
-            Write-Host "Removed legacy scheduled task '$legacyName'."
+        try {
+            $legacyTask = Get-ScheduledTask -TaskName $legacyName -ErrorAction SilentlyContinue
+            if ($legacyTask) {
+                Unregister-ScheduledTask -TaskName $legacyName -Confirm:$false
+                Write-Host "Removed legacy scheduled task '$legacyName'."
+            }
+        }
+        catch {
+            Write-Warning "Could not remove legacy scheduled task '$legacyName': $($_.Exception.Message)"
         }
     }
 }
 
-$python = Resolve-WindowlessPython
 $currentUser = [System.Security.Principal.WindowsIdentity]::GetCurrent().Name
-$arguments = @(
-    "-u",
-    "`"$Agent`"",
-    "--repo-root", "`"$RepoRoot`"",
-    "--api-base-url", "`"$ApiBaseUrl`"",
-    "--poll-seconds", $PollSeconds.ToString()
+$principal = New-Object System.Security.Principal.WindowsPrincipal(
+    [System.Security.Principal.WindowsIdentity]::GetCurrent()
+)
+$isAdministrator = $principal.IsInRole([System.Security.Principal.WindowsBuiltInRole]::Administrator)
+$powershell = (Get-Command powershell.exe -ErrorAction Stop).Source
+$runnerArguments = @(
+    "-NoProfile",
+    "-WindowStyle", "Hidden",
+    "-ExecutionPolicy", "Bypass",
+    "-File", "`"$runner`"",
+    "-RepoRoot", "`"$RepoRoot`"",
+    "-ApiBaseUrl", "`"$ApiBaseUrl`"",
+    "-CredentialPath", "`"$CredentialPath`"",
+    "-WorkerName", "`"$env:COMPUTERNAME`"",
+    "-PollSeconds", $PollSeconds.ToString()
 ) -join " "
 
-$action = New-ScheduledTaskAction -Execute $python -Argument $arguments -WorkingDirectory $RepoRoot
-$trigger = @(
-    New-ScheduledTaskTrigger -AtLogOn -User $currentUser
-    New-ScheduledTaskTrigger -AtStartup
-    New-ScheduledTaskTrigger `
-        -Once `
-        -At (Get-Date).AddMinutes(1) `
-        -RepetitionInterval (New-TimeSpan -Minutes $WatchdogMinutes) `
-        -RepetitionDuration (New-TimeSpan -Days 3650)
-)
-$principal = New-ScheduledTaskPrincipal `
-    -UserId $currentUser `
-    -LogonType Interactive `
-    -RunLevel Limited
-$settings = New-ScheduledTaskSettingsSet `
-    -StartWhenAvailable `
-    -MultipleInstances IgnoreNew `
-    -AllowStartIfOnBatteries `
-    -DontStopIfGoingOnBatteries `
-    -RestartCount 999 `
-    -RestartInterval (New-TimeSpan -Minutes 1) `
-    -ExecutionTimeLimit (New-TimeSpan -Days 30)
+$startupShortcut = Join-Path ([Environment]::GetFolderPath("Startup")) "$TaskName.lnk"
+$runKey = "HKCU:\Software\Microsoft\Windows\CurrentVersion\Run"
+Remove-ItemProperty -Path $runKey -Name "TradeBotWindowsTrainingAgent" -ErrorAction SilentlyContinue
 
-Register-ScheduledTask `
-    -TaskName $TaskName `
-    -Action $action `
-    -Trigger $trigger `
-    -Principal $principal `
-    -Settings $settings `
-    -Description "Keeps the TradeBot Windows training agent online and polls the public bot API for retrain jobs." `
-    -Force | Out-Null
+$installMode = "startup shortcut"
+if ($isAdministrator) {
+    if (Test-Path -LiteralPath $startupShortcut) {
+        Remove-Item -LiteralPath $startupShortcut -Force
+    }
+    $action = New-ScheduledTaskAction -Execute $powershell -Argument $runnerArguments -WorkingDirectory $RepoRoot
+    $trigger = @(
+        New-ScheduledTaskTrigger -AtLogOn -User $currentUser
+        New-ScheduledTaskTrigger -AtStartup
+        New-ScheduledTaskTrigger `
+            -Once `
+            -At (Get-Date).AddMinutes(1) `
+            -RepetitionInterval (New-TimeSpan -Minutes $WatchdogMinutes) `
+            -RepetitionDuration (New-TimeSpan -Days 3650)
+    )
+    $taskPrincipal = New-ScheduledTaskPrincipal `
+        -UserId $currentUser `
+        -LogonType Interactive `
+        -RunLevel Limited
+    $settings = New-ScheduledTaskSettingsSet `
+        -StartWhenAvailable `
+        -MultipleInstances IgnoreNew `
+        -AllowStartIfOnBatteries `
+        -DontStopIfGoingOnBatteries `
+        -RestartCount 999 `
+        -RestartInterval (New-TimeSpan -Minutes 1) `
+        -ExecutionTimeLimit (New-TimeSpan -Days 30)
 
-if ($StartNow) {
-    Start-ScheduledTask -TaskName $TaskName
+    Register-ScheduledTask `
+        -TaskName $TaskName `
+        -Action $action `
+        -Trigger $trigger `
+        -Principal $taskPrincipal `
+        -Settings $settings `
+        -Description "Keeps the TradeBot Windows training agent online and polls the bot API for retrain jobs." `
+        -Force | Out-Null
+    $installMode = "scheduled task"
+}
+else {
+    try {
+        $existingTask = Get-ScheduledTask -TaskName $TaskName -ErrorAction SilentlyContinue
+        if ($existingTask) {
+            Unregister-ScheduledTask -TaskName $TaskName -Confirm:$false
+        }
+    }
+    catch {
+        Write-Warning "Could not remove an existing elevated task: $($_.Exception.Message)"
+    }
+
+    $shell = New-Object -ComObject WScript.Shell
+    $shortcut = $shell.CreateShortcut($startupShortcut)
+    $shortcut.TargetPath = $powershell
+    $shortcut.Arguments = $runnerArguments
+    $shortcut.WorkingDirectory = $RepoRoot
+    $shortcut.WindowStyle = 7
+    $shortcut.Description = "TradeBot Windows Training Agent"
+    $shortcut.Save()
 }
 
-Write-Host "Registered scheduled task '$TaskName' for Windows startup, logon, and watchdog restarts."
+Get-CimInstance Win32_Process |
+    Where-Object {
+        $_.ProcessId -ne $PID -and
+        $_.CommandLine -and
+        ($_.CommandLine -match [regex]::Escape("windows_training_agent.py") -or
+         $_.CommandLine -match [regex]::Escape("run_windows_training_agent.ps1"))
+    } |
+    ForEach-Object { Stop-Process -Id $_.ProcessId -Force -ErrorAction SilentlyContinue }
+
+if ($StartNow) {
+    if ($installMode -eq "scheduled task") {
+        Start-ScheduledTask -TaskName $TaskName
+    }
+    else {
+        Start-Process `
+            -FilePath $powershell `
+            -ArgumentList $runnerArguments `
+            -WorkingDirectory $RepoRoot `
+            -WindowStyle Hidden | Out-Null
+    }
+}
+
+Write-Host "Installed '$TaskName' using $installMode."
 Write-Host "Agent API: $ApiBaseUrl"
-Write-Host "Agent script: $Agent"
+Write-Host "Encrypted credential: $CredentialPath"
+Write-Host "Agent runner: $runner"
diff --git a/tools/run_torch_retrain.ps1 b/tools/run_torch_retrain.ps1
index ac843a3..6f4fd53 100644
--- a/tools/run_torch_retrain.ps1
+++ b/tools/run_torch_retrain.ps1
@@ -12,6 +12,10 @@ param(
     [string]$Features = "",
     [string]$ContextSymbols = "",
     [int]$Seed = 0,
+    [string]$EnsembleSeeds = "",
+    [int]$SelectionFolds = 0,
+    [double]$LearningRate = 0,
+    [double]$WeightDecay = 0,
     [int]$Epochs = 0,
     [int]$Patience = 0,
     [int]$HoldoutWindow = 0,
@@ -132,18 +136,22 @@ function Sync-AcceptedArtifactsToPi {
 
 if (-not $Symbols -and $env:TORCH_RETRAIN_SYMBOLS) { $Symbols = $env:TORCH_RETRAIN_SYMBOLS }
 if ($Limit -le 0) {
-    $Limit = if ($env:TORCH_RETRAIN_LIMIT) { [int]$env:TORCH_RETRAIN_LIMIT } else { 3000 }
+    $Limit = if ($env:TORCH_RETRAIN_LIMIT) { [int]$env:TORCH_RETRAIN_LIMIT } else { 6000 }
 }
-if (-not $Lookbacks) { $Lookbacks = if ($env:TORCH_RETRAIN_LOOKBACKS) { $env:TORCH_RETRAIN_LOOKBACKS } else { "64" } }
+if (-not $Lookbacks) { $Lookbacks = if ($env:TORCH_RETRAIN_LOOKBACKS) { $env:TORCH_RETRAIN_LOOKBACKS } else { "32,64,128" } }
 if (-not $Architectures) { $Architectures = if ($env:TORCH_RETRAIN_ARCHITECTURES) { $env:TORCH_RETRAIN_ARCHITECTURES } else { "lstm,gru" } }
 if (-not $HiddenSizes) { $HiddenSizes = if ($env:TORCH_RETRAIN_HIDDEN_SIZES) { $env:TORCH_RETRAIN_HIDDEN_SIZES } else { "64,96" } }
 if (-not $Layers) { $Layers = if ($env:TORCH_RETRAIN_LAYERS) { $env:TORCH_RETRAIN_LAYERS } else { "2" } }
-if (-not $Dropouts) { $Dropouts = if ($env:TORCH_RETRAIN_DROPOUTS) { $env:TORCH_RETRAIN_DROPOUTS } else { "0.15" } }
-if ($Horizon -le 0 -and $env:TORCH_RETRAIN_HORIZON) { $Horizon = [int]$env:TORCH_RETRAIN_HORIZON }
-if (-not $Horizons -and $env:TORCH_RETRAIN_HORIZONS) { $Horizons = $env:TORCH_RETRAIN_HORIZONS }
+if (-not $Dropouts) { $Dropouts = if ($env:TORCH_RETRAIN_DROPOUTS) { $env:TORCH_RETRAIN_DROPOUTS } else { "0.20" } }
+if ($Horizon -le 0) { $Horizon = if ($env:TORCH_RETRAIN_HORIZON) { [int]$env:TORCH_RETRAIN_HORIZON } else { 12 } }
+if (-not $Horizons) { $Horizons = if ($env:TORCH_RETRAIN_HORIZONS) { $env:TORCH_RETRAIN_HORIZONS } else { "3,6,12,24" } }
 if (-not $Features -and $env:TORCH_RETRAIN_FEATURES) { $Features = $env:TORCH_RETRAIN_FEATURES }
 if (-not $ContextSymbols -and $env:TORCH_RETRAIN_CONTEXT_SYMBOLS) { $ContextSymbols = $env:TORCH_RETRAIN_CONTEXT_SYMBOLS }
 if ($Seed -le 0 -and $env:TORCH_RETRAIN_SEED) { $Seed = [int]$env:TORCH_RETRAIN_SEED }
+if (-not $EnsembleSeeds) { $EnsembleSeeds = if ($env:TORCH_RETRAIN_ENSEMBLE_SEEDS) { $env:TORCH_RETRAIN_ENSEMBLE_SEEDS } else { "7,19,43" } }
+if ($SelectionFolds -le 0) { $SelectionFolds = if ($env:TORCH_RETRAIN_SELECTION_FOLDS) { [int]$env:TORCH_RETRAIN_SELECTION_FOLDS } else { 3 } }
+if ($LearningRate -le 0) { $LearningRate = if ($env:TORCH_RETRAIN_LEARNING_RATE) { [double]$env:TORCH_RETRAIN_LEARNING_RATE } else { 0.0007 } }
+if ($WeightDecay -le 0) { $WeightDecay = if ($env:TORCH_RETRAIN_WEIGHT_DECAY) { [double]$env:TORCH_RETRAIN_WEIGHT_DECAY } else { 0.0005 } }
 if ($Epochs -le 0) { $Epochs = if ($env:TORCH_RETRAIN_EPOCHS) { [int]$env:TORCH_RETRAIN_EPOCHS } else { 70 } }
 if ($Patience -le 0) { $Patience = if ($env:TORCH_RETRAIN_PATIENCE) { [int]$env:TORCH_RETRAIN_PATIENCE } else { 8 } }
 if ($HoldoutWindow -le 0) { $HoldoutWindow = if ($env:TORCH_RETRAIN_HOLDOUT_WINDOW) { [int]$env:TORCH_RETRAIN_HOLDOUT_WINDOW } else { 240 } }
@@ -182,6 +190,10 @@ try {
         "--epochs", $Epochs.ToString(),
         "--patience", $Patience.ToString(),
         "--holdout-window", $HoldoutWindow.ToString(),
+        "--ensemble-seeds", $EnsembleSeeds,
+        "--selection-folds", $SelectionFolds.ToString(),
+        "--learning-rate", $LearningRate.ToString([Globalization.CultureInfo]::InvariantCulture),
+        "--weight-decay", $WeightDecay.ToString([Globalization.CultureInfo]::InvariantCulture),
         "--output", $CandidateFile
     )
     if ($Symbols) { $trainerArgs += @("--symbols", $Symbols) }
@@ -222,8 +234,8 @@ try {
     $calibrationBaseArgs = @(
         "-u",
         "tools\calibrate_torch_thresholds.py",
-        "--limit", "3000",
-        "--calibration-window", "1200",
+        "--limit", $Limit.ToString(),
+        "--calibration-window", ([Math]::Min(2400, [Math]::Max(1200, [int]($Limit / 2)))).ToString(),
         "--min-trades", "60",
         "--walk-forward-folds", "8",
         "--confidence-grid", "0.40"
diff --git a/tools/run_windows_training_agent.ps1 b/tools/run_windows_training_agent.ps1
new file mode 100644
index 0000000..5ec375a
--- /dev/null
+++ b/tools/run_windows_training_agent.ps1
@@ -0,0 +1,82 @@
+[CmdletBinding()]
+param(
+    [string]$ApiBaseUrl = "https://tb.kusoft.xyz",
+    [string]$RepoRoot = "",
+    [string]$CredentialPath = "",
+    [string]$WorkerName = $env:COMPUTERNAME,
+    [int]$PollSeconds = 10,
+    [int]$RestartDelaySeconds = 10
+)
+
+$ErrorActionPreference = "Stop"
+
+if (-not $RepoRoot) {
+    $RepoRoot = (Resolve-Path (Join-Path $PSScriptRoot "..")).Path
+}
+if (-not $CredentialPath) {
+    $CredentialPath = Join-Path $env:LOCALAPPDATA "TradeBot\training-agent.token"
+}
+
+$agent = Join-Path $RepoRoot "tools\windows_training_agent.py"
+if (-not (Test-Path -LiteralPath $agent)) {
+    throw "Windows training agent not found: $agent"
+}
+if (-not (Test-Path -LiteralPath $CredentialPath)) {
+    throw "Encrypted training credential not found: $CredentialPath"
+}
+
+function Resolve-Python {
+    $venvPython = Join-Path $RepoRoot ".venv\Scripts\python.exe"
+    if (Test-Path -LiteralPath $venvPython) {
+        return $venvPython
+    }
+
+    $userPython = Join-Path $env:LOCALAPPDATA "Programs\TradeBotPython312\python.exe"
+    if (Test-Path -LiteralPath $userPython) {
+        return $userPython
+    }
+
+    foreach ($candidate in @("python.exe", "python")) {
+        $command = Get-Command $candidate -ErrorAction SilentlyContinue
+        if ($command) {
+            return $command.Source
+        }
+    }
+    throw "Python was not found. Create .venv or install Python 3.12."
+}
+
+$createdNew = $false
+$mutex = [System.Threading.Mutex]::new($false, "Local\TradeBotWindowsTrainingAgent", [ref]$createdNew)
+if (-not $createdNew) {
+    $mutex.Dispose()
+    exit 0
+}
+
+$encryptedToken = (Get-Content -LiteralPath $CredentialPath -Raw -Encoding UTF8).Trim()
+$secureToken = $encryptedToken | ConvertTo-SecureString
+$tokenPointer = [Runtime.InteropServices.Marshal]::SecureStringToBSTR($secureToken)
+try {
+    $env:TRADEBOT_API_AUTH = [Runtime.InteropServices.Marshal]::PtrToStringBSTR($tokenPointer)
+    $python = Resolve-Python
+    $workerId = "${WorkerName}:$RepoRoot"
+    $arguments = @(
+        "-u",
+        $agent,
+        "--repo-root", $RepoRoot,
+        "--api-base-url", $ApiBaseUrl,
+        "--worker-id", $workerId,
+        "--worker-name", $WorkerName,
+        "--poll-seconds", [Math]::Max(5, $PollSeconds).ToString()
+    )
+
+    while ($true) {
+        & $python @arguments
+        Start-Sleep -Seconds ([Math]::Max(5, $RestartDelaySeconds))
+    }
+}
+finally {
+    $env:TRADEBOT_API_AUTH = $null
+    [Runtime.InteropServices.Marshal]::ZeroFreeBSTR($tokenPointer)
+    $mutex.ReleaseMutex()
+    $mutex.Dispose()
+}
diff --git a/tools/train_torch_recurrent_forecaster.py b/tools/train_torch_recurrent_forecaster.py
index b6805d9..08f14d1 100644
--- a/tools/train_torch_recurrent_forecaster.py
+++ b/tools/train_torch_recurrent_forecaster.py
@@ -125,6 +125,9 @@ def main() -> None:
     decision_horizon = args.horizon if args.horizon > 0 else max(1, settings.time_series_forecast_horizon)
     target_horizons = _horizons(args.horizons, decision_horizon)
     feature_names = _feature_names_arg(args.features)
+    if args.pooled:
+        feature_names.extend(f"symbol_is_{symbol}" for symbol in symbols)
+    ensemble_seeds = _ints(args.ensemble_seeds) or [args.seed]
     round_trip_cost = max(0.0, 2.0 * (float(settings.taker_fee_rate) + float(settings.slippage_rate)))
     _progress(
         f"training started: symbols={len(symbols)} interval={interval} "
@@ -151,9 +154,82 @@ def main() -> None:
         "feature_names": feature_names,
         "feature_count": len(feature_names),
         "device": str(device),
+        "ensemble_seeds": ensemble_seeds,
+        "selection_folds": args.selection_folds,
         "symbols": {},
     }
 
+    if args.pooled:
+        artifact["version"] = 5
+        artifact["pooled_multi_asset"] = True
+        artifact["symbol_embedding"] = "learned_one_hot_projection"
+        artifact["symbols"] = _train_pooled_symbols(
+            client=client,
+            symbols=symbols,
+            interval=interval,
+            limit=args.limit,
+            validation_window=args.validation_window,
+            holdout_window=args.holdout_window,
+            target_horizons=target_horizons,
+            decision_horizon=decision_horizon,
+            feature_names=feature_names,
+            round_trip_cost=round_trip_cost,
+            context_symbols=_strings(args.context_symbols),
+            architectures=_strings(args.architectures),
+            lookbacks=_ints(args.lookbacks),
+            hidden_sizes=_ints(args.hidden_sizes),
+            layers_values=_ints(args.layers),
+            dropouts=_floats(args.dropouts),
+            epochs=args.epochs,
+            patience=args.patience,
+            batch_size=args.batch_size,
+            learning_rate=args.learning_rate,
+            weight_decay=args.weight_decay,
+            clip=args.clip,
+            attention_pooling=args.attention_pooling,
+            context_norm=args.context_norm,
+            device=device,
+            seeds=ensemble_seeds,
+            selection_folds=args.selection_folds,
+        )
+    else:
+        artifact["symbols"] = _train_independent_symbols(
+            client=client,
+            symbols=symbols,
+            interval=interval,
+            args=args,
+            target_horizons=target_horizons,
+            decision_horizon=decision_horizon,
+            feature_names=feature_names,
+            round_trip_cost=round_trip_cost,
+            device=device,
+            ensemble_seeds=ensemble_seeds,
+        )
+
+    for symbol, result in artifact["symbols"].items():
+        _progress(
+            f"{symbol}: model={result['model']} lookback={result['lookback']} "
+            f"features={result['input_size']} hidden={result['hidden_size']} "
+            f"layers={result['num_layers']} horizons={','.join(map(str, result['target_horizons']))} "
+            f"mae={result['validation_mae_percent']:.5f}% "
+            f"baseline={result['baseline_mae_percent']:.5f}% "
+            f"skill={result['skill']:.4f} dir={result['directional_accuracy']:.3f} "
+            f"p_brier={result['probability_brier']:.4f}"
+        )
+
+    output.parent.mkdir(parents=True, exist_ok=True)
+    tmp_output = output.with_name(f"{output.name}.tmp")
+    tmp_output.write_text(json.dumps(artifact, ensure_ascii=False, indent=2) + "\n", encoding="utf-8")
+    tmp_output.replace(output)
+    _progress(f"saved {output}")
+
+
+def _train_independent_symbols(
+    *, client: BybitClient, symbols: list[str], interval: str, args: argparse.Namespace,
+    target_horizons: list[int], decision_horizon: int, feature_names: list[str],
+    round_trip_cost: float, device: torch.device, ensemble_seeds: list[int],
+) -> dict[str, Any]:
+    results: dict[str, Any] = {}
     total_symbols = len(symbols)
     for index, symbol in enumerate(symbols, start=1):
         _progress(f"{symbol}: training started ({index}/{total_symbols})")
@@ -183,27 +259,221 @@ def main() -> None:
             attention_pooling=args.attention_pooling,
             context_norm=args.context_norm,
             device=device,
-            seed=args.seed,
+            seeds=ensemble_seeds,
+            selection_folds=args.selection_folds,
         )
         if result is None:
             _progress(f"{symbol}: skipped, not enough candles or train/validation samples")
             continue
-        artifact["symbols"][symbol] = result
-        _progress(
-            f"{symbol}: model={result['model']} lookback={result['lookback']} "
-            f"features={result['input_size']} hidden={result['hidden_size']} "
-            f"layers={result['num_layers']} horizons={','.join(map(str, result['target_horizons']))} "
-            f"mae={result['validation_mae_percent']:.5f}% "
-            f"baseline={result['baseline_mae_percent']:.5f}% "
-            f"skill={result['skill']:.4f} dir={result['directional_accuracy']:.3f} "
-            f"p_brier={result['probability_brier']:.4f}"
-        )
+        results[symbol] = result
+    return results
 
-    output.parent.mkdir(parents=True, exist_ok=True)
-    tmp_output = output.with_name(f"{output.name}.tmp")
-    tmp_output.write_text(json.dumps(artifact, ensure_ascii=False, indent=2) + "\n", encoding="utf-8")
-    tmp_output.replace(output)
-    _progress(f"saved {output}")
+
+def _train_pooled_symbols(
+    *, client: BybitClient, symbols: list[str], interval: str, limit: int,
+    validation_window: int, holdout_window: int, target_horizons: list[int],
+    decision_horizon: int, feature_names: list[str], round_trip_cost: float,
+    context_symbols: list[str], architectures: list[str], lookbacks: list[int],
+    hidden_sizes: list[int], layers_values: list[int], dropouts: list[float],
+    epochs: int, patience: int, batch_size: int, learning_rate: float,
+    weight_decay: float, clip: float, attention_pooling: bool, context_norm: bool,
+    device: torch.device, seeds: list[int], selection_folds: int,
+) -> dict[str, Any]:
+    market_candles: dict[str, list[Candle]] = {}
+    for symbol in sorted({item.upper() for item in symbols + context_symbols}):
+        rows = _historical_klines(client, symbol, interval, limit)
+        add_indicators(rows)
+        market_candles[symbol] = rows
+        _progress(f"{symbol}: pooled data loaded ({len(rows)} candles)")
+    trend_by_symbol: dict[str, list[Candle]] = {}
+    for symbol in symbols:
+        rows = _historical_klines(client, symbol, "D", min(max(260, limit // 24 + 260), 1000))
+        add_indicators(rows)
+        trend_by_symbol[symbol] = rows
+
+    best: dict[str, Any] | None = None
+    best_prepared: dict[str, PreparedData] = {}
+    for lookback in lookbacks:
+        prepared_by_symbol: dict[str, PreparedData] = {}
+        for symbol in symbols:
+            prepared = _prepare_data(
+                candles=market_candles[symbol],
+                feature_names=feature_names,
+                lookback=lookback,
+                target_horizons=target_horizons,
+                decision_horizon=decision_horizon,
+                round_trip_cost=round_trip_cost,
+                market_candles=market_candles,
+                trend_candles=trend_by_symbol[symbol],
+                validation_window=validation_window,
+                holdout_window=holdout_window,
+                clip=clip,
+                device=device,
+            )
+            if prepared is not None:
+                prepared_by_symbol[symbol] = prepared
+        if len(prepared_by_symbol) < 2:
+            continue
+        for architecture in architectures:
+            if architecture not in {"lstm", "gru"}:
+                continue
+            for hidden_size in hidden_sizes:
+                for num_layers in layers_values:
+                    for dropout in dropouts:
+                        if num_layers <= 1 and dropout != 0.0:
+                            continue
+                        _progress(
+                            f"pooled: fitting {architecture} lookback={lookback} hidden={hidden_size} "
+                            f"layers={num_layers} dropout={dropout} symbols={len(prepared_by_symbol)}"
+                        )
+                        members = [
+                            _fit_pooled_candidate(
+                                prepared_by_symbol=prepared_by_symbol,
+                                architecture=architecture,
+                                input_size=len(feature_names),
+                                output_size=len(target_horizons) * len(OUTPUT_LAYOUT),
+                                hidden_size=hidden_size,
+                                num_layers=num_layers,
+                                dropout=dropout,
+                                epochs=epochs,
+                                patience=patience,
+                                batch_size=batch_size,
+                                learning_rate=learning_rate,
+                                weight_decay=weight_decay,
+                                clip=clip,
+                                attention_pooling=attention_pooling,
+                                context_norm=context_norm,
+                                device=device,
+                                seed=member_seed,
+                                selection_folds=selection_folds,
+                            )
+                            for member_seed in seeds
+                        ]
+                        candidate = _ensemble_candidate(members, seeds)
+                        candidate.update(
+                            model=f"torch_{architecture}", architecture=architecture,
+                            lookback=lookback, hidden_size=hidden_size, num_layers=num_layers,
+                            dropout=dropout if num_layers > 1 else 0.0,
+                            attention_pooling=attention_pooling, context_norm=context_norm,
+                            input_size=len(feature_names), output_size=len(target_horizons) * len(OUTPUT_LAYOUT),
+                        )
+                        if best is None or _candidate_score(candidate) < _candidate_score(best):
+                            best = candidate
+                            best_prepared = prepared_by_symbol
+    if best is None:
+        return {}
+
+    results: dict[str, Any] = {}
+    symbol_metrics = best.get("symbol_metrics", {})
+    common = {key: value for key, value in best.items() if key != "symbol_metrics"}
+    for symbol, prepared in best_prepared.items():
+        metrics = symbol_metrics.get(symbol, {}) if isinstance(symbol_metrics, dict) else {}
+        baseline = sum(abs(row[prepared.decision_horizon_index]) for row in prepared.validation_targets) / len(prepared.validation_targets)
+        validation_mae = float(metrics.get("validation_mae", baseline))
+        results[symbol] = {
+            **common, **metrics,
+            "pooled_multi_asset": True,
+            "target_horizon": prepared.decision_horizon,
+            "target_horizons": prepared.target_horizons,
+            "direct_horizon": True,
+            "target_transform": "net_return_over_volatility",
+            "round_trip_cost": round(round_trip_cost, 10),
+            "output_layout": list(OUTPUT_LAYOUT),
+            "feature_names": feature_names,
+            "feature_means": prepared.feature_means,
+            "feature_scales": prepared.feature_scales,
+            "target_means": prepared.target_means,
+            "target_scales": prepared.target_scales,
+            "target_mean": prepared.target_means[prepared.decision_horizon_index],
+            "target_scale": prepared.target_scales[prepared.decision_horizon_index],
+            "clip": clip,
+            "validation_mae_percent": validation_mae * 100,
+            "baseline_mae_percent": baseline * 100,
+            "validation_skill": (baseline - validation_mae) / baseline if baseline > 0 else 0.0,
+            # Runtime and calibration may use validation skill. Untouched
+            # holdout skill is report-only and must never gate individual entries.
+            "skill": (baseline - validation_mae) / baseline if baseline > 0 else 0.0,
+            "train_samples": prepared.train_samples,
+            "validation_samples": prepared.validation_samples,
+            "holdout_samples": prepared.holdout_samples,
+            "holdout_start_timestamp": prepared.holdout_start_timestamp,
+        }
+    return results
+
+
+def _fit_pooled_candidate(
+    *, prepared_by_symbol: dict[str, PreparedData], architecture: str, input_size: int,
+    output_size: int, hidden_size: int, num_layers: int, dropout: float, epochs: int,
+    patience: int, batch_size: int, learning_rate: float, weight_decay: float,
+    clip: float, attention_pooling: bool, context_norm: bool, device: torch.device,
+    seed: int, selection_folds: int,
+) -> dict[str, Any]:
+    _seed(seed)
+    model = RecurrentReturnModel(
+        architecture=architecture, input_size=input_size, hidden_size=hidden_size,
+        num_layers=num_layers, dropout=dropout, output_size=output_size,
+        attention_pooling=attention_pooling, context_norm=context_norm,
+    ).to(device)
+    optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate, weight_decay=weight_decay)
+    loader = DataLoader(
+        TensorDataset(
+            torch.cat([row.train_x for row in prepared_by_symbol.values()]),
+            torch.cat([row.train_y for row in prepared_by_symbol.values()]),
+            torch.cat([row.train_up for row in prepared_by_symbol.values()]),
+        ),
+        batch_size=max(1, batch_size), shuffle=True,
+        generator=torch.Generator(device="cpu").manual_seed(seed),
+    )
+    best_state: dict[str, torch.Tensor] | None = None
+    best_score = math.inf
+    stale = 0
+    best_epoch = 0
+    for epoch in range(1, max(1, epochs) + 1):
+        model.train()
+        for batch_x, batch_y, batch_up in loader:
+            optimizer.zero_grad(set_to_none=True)
+            loss = _forecast_loss(model(batch_x), batch_y, batch_up, len(next(iter(prepared_by_symbol.values())).target_horizons))
+            loss.backward()
+            nn.utils.clip_grad_norm_(model.parameters(), max_norm=1.0)
+            optimizer.step()
+        symbol_rows = {
+            symbol: _validation_metrics(model, prepared, clip)
+            for symbol, prepared in prepared_by_symbol.items()
+        }
+        score = sum(float(row["validation_mae"]) for row in symbol_rows.values()) / len(symbol_rows)
+        if score + 1e-12 < best_score:
+            best_score = score
+            best_epoch = epoch
+            best_state = {key: value.detach().cpu().clone() for key, value in model.state_dict().items()}
+            stale = 0
+        else:
+            stale += 1
+            if stale >= max(1, patience):
+                break
+    if best_state:
+        model.load_state_dict(best_state)
+    per_symbol: dict[str, dict[str, Any]] = {}
+    for symbol, prepared in prepared_by_symbol.items():
+        metrics = _validation_metrics(model, prepared, clip)
+        metrics.update(_validation_stability_metrics(model, prepared, clip, selection_folds))
+        metrics.update(_holdout_metrics(model, prepared, clip))
+        per_symbol[symbol] = metrics
+    aggregate: dict[str, Any] = {"symbol_metrics": per_symbol}
+    for name in (
+        "validation_mae", "directional_accuracy", "buy_precision", "probability_brier",
+        "holdout_skill", "validation_fold_mae_std", "validation_trade_mean",
+        "validation_trade_win_rate",
+    ):
+        values = [float(row[name]) for row in per_symbol.values() if isinstance(row.get(name), (int, float))]
+        aggregate[name] = sum(values) / len(values) if values else 0.0
+    aggregate.update(
+        best_epoch=best_epoch, epochs_trained=best_epoch + stale,
+        state_dict=_export_recurrent_state(model),
+        head_weight=_round_nested(model.head.weight.detach().cpu().tolist()),
+        head_bias=_round_list(model.head.bias.detach().cpu().tolist()),
+        **_export_context_state(model),
+    )
+    return aggregate
 
 
 def _progress(message: str) -> None:
@@ -236,6 +506,9 @@ def _parse_args() -> argparse.Namespace:
     parser.add_argument("--attention-pooling", action=argparse.BooleanOptionalAction, default=True, help="Use exportable attention pooling over recurrent states.")
     parser.add_argument("--context-norm", action=argparse.BooleanOptionalAction, default=True, help="Use exportable LayerNorm before the forecast head.")
     parser.add_argument("--seed", type=int, default=7, help="Random seed.")
+    parser.add_argument("--ensemble-seeds", default="7,19,43", help="Comma-separated seeds averaged at inference time.")
+    parser.add_argument("--selection-folds", type=int, default=3, help="Validation slices used to penalize unstable candidates.")
+    parser.add_argument("--pooled", action=argparse.BooleanOptionalAction, default=True, help="Train shared multi-asset recurrent weights with learned symbol one-hot projection.")
     parser.add_argument("--threads", type=int, default=0, help="Torch CPU threads; 0 keeps torch default.")
     parser.add_argument("--device", default="auto", help="auto, cpu, cuda, or mps.")
     parser.add_argument("--output", default="", help="Output JSON path. Defaults to TIME_SERIES_LSTM_MODEL_PATH.")
@@ -277,7 +550,8 @@ def _train_symbol(
     attention_pooling: bool,
     context_norm: bool,
     device: torch.device,
-    seed: int,
+    seeds: list[int],
+    selection_folds: int,
 ) -> dict[str, Any] | None:
     candles = _historical_klines(client, symbol, interval, limit)
     add_indicators(candles)
@@ -339,25 +613,30 @@ def _train_symbol(
                             f"lookback={lookback} hidden={hidden_size} "
                             f"layers={num_layers} dropout={dropout}"
                         )
-                        candidate = _fit_candidate(
-                            prepared=prepared,
-                            architecture=architecture,
-                            input_size=len(feature_names),
-                            output_size=len(target_horizons) * len(OUTPUT_LAYOUT),
-                            hidden_size=hidden_size,
-                            num_layers=num_layers,
-                            dropout=dropout,
-                            epochs=epochs,
-                            patience=patience,
-                            batch_size=batch_size,
-                            learning_rate=learning_rate,
-                            weight_decay=weight_decay,
-                            clip=clip,
-                            attention_pooling=attention_pooling,
-                            context_norm=context_norm,
-                            device=device,
-                            seed=seed,
-                        )
+                        members = [
+                            _fit_candidate(
+                                prepared=prepared,
+                                architecture=architecture,
+                                input_size=len(feature_names),
+                                output_size=len(target_horizons) * len(OUTPUT_LAYOUT),
+                                hidden_size=hidden_size,
+                                num_layers=num_layers,
+                                dropout=dropout,
+                                epochs=epochs,
+                                patience=patience,
+                                batch_size=batch_size,
+                                learning_rate=learning_rate,
+                                weight_decay=weight_decay,
+                                clip=clip,
+                                attention_pooling=attention_pooling,
+                                context_norm=context_norm,
+                                device=device,
+                                seed=member_seed,
+                                selection_folds=selection_folds,
+                            )
+                            for member_seed in seeds
+                        ]
+                        candidate = _ensemble_candidate(members, seeds)
                         validation_mae = float(candidate["validation_mae"])
                         skill = (baseline_mae - validation_mae) / baseline_mae if baseline_mae > 0 else 0.0
                         row = {
@@ -408,7 +687,7 @@ def _train_symbol(
     if best is None:
         return None
     best["validation_skill"] = best.get("skill", 0.0)
-    best["skill"] = best.get("holdout_skill", 0.0)
+    best["skill"] = best["validation_skill"]
     best.pop("validation_mae", None)
     return best
 
@@ -484,7 +763,7 @@ def _prepare_data(
     if len(train_samples) < 24 or len(validation_samples) < 8 or len(holdout_samples) < 16:
         return None
 
-    feature_means, feature_scales = _feature_stats(train_samples, len(feature_names))
+    feature_means, feature_scales = _feature_stats(train_samples, feature_names)
     target_means, target_scales = _target_stats(train_samples, len(target_horizons))
     decision_horizon = decision_horizon if decision_horizon in target_horizons else min(
         target_horizons,
@@ -545,7 +824,8 @@ def _prepare_data(
     )
 
 
-def _feature_stats(samples: list[TrainingSample], input_size: int) -> tuple[list[float], list[float]]:
+def _feature_stats(samples: list[TrainingSample], feature_names: list[str]) -> tuple[list[float], list[float]]:
+    input_size = len(feature_names)
     columns = [[] for _ in range(input_size)]
     for sample in samples:
         window = sample.window
@@ -554,7 +834,11 @@ def _feature_stats(samples: list[TrainingSample], input_size: int) -> tuple[list
                 columns[index].append(float(row[index] if index < len(row) else 0.0))
     means: list[float] = []
     scales: list[float] = []
-    for values in columns:
+    for index, values in enumerate(columns):
+        if feature_names[index].startswith("symbol_is_"):
+            means.append(0.0)
+            scales.append(1.0)
+            continue
         if not values:
             means.append(0.0)
             scales.append(1.0)
@@ -641,6 +925,7 @@ def _fit_candidate(
     context_norm: bool,
     device: torch.device,
     seed: int,
+    selection_folds: int,
 ) -> dict[str, Any]:
     _seed(seed)
     model = RecurrentReturnModel(
@@ -676,6 +961,7 @@ def _fit_candidate(
             optimizer.step()
 
         metrics = _validation_metrics(model, prepared, clip)
+        metrics.update(_validation_stability_metrics(model, prepared, clip, selection_folds))
         if metrics["validation_mae"] + 1e-12 < best_metrics["validation_mae"]:
             best_metrics = metrics
             best_epoch = epoch
@@ -701,6 +987,66 @@ def _fit_candidate(
     }
 
 
+def _ensemble_candidate(members: list[dict[str, Any]], seeds: list[int]) -> dict[str, Any]:
+    if not members:
+        raise ValueError("ensemble requires at least one member")
+    result = dict(members[0])
+    metric_names = (
+        "validation_mae",
+        "directional_accuracy",
+        "buy_precision",
+        "probability_brier",
+        "holdout_mae",
+        "holdout_baseline_mae",
+        "holdout_skill",
+        "holdout_directional_accuracy",
+        "holdout_buy_precision",
+        "holdout_probability_brier",
+        "validation_fold_mae_std",
+        "validation_fold_mae_worst",
+        "validation_trade_mean",
+        "validation_trade_win_rate",
+    )
+    for name in metric_names:
+        values = [float(member[name]) for member in members if isinstance(member.get(name), (int, float))]
+        if values:
+            result[name] = sum(values) / len(values)
+    export_names = (
+        "state_dict",
+        "head_weight",
+        "head_bias",
+        "attention_weight",
+        "attention_bias",
+        "context_norm_weight",
+        "context_norm_bias",
+    )
+    result["ensemble_members"] = [
+        {name: member[name] for name in export_names if name in member}
+        | {"seed": seeds[index] if index < len(seeds) else index}
+        for index, member in enumerate(members)
+    ]
+    result["ensemble_size"] = len(members)
+    symbol_names = sorted(
+        {
+            symbol
+            for member in members
+            for symbol in (member.get("symbol_metrics") or {})
+        }
+    )
+    if symbol_names:
+        result["symbol_metrics"] = {}
+        for symbol in symbol_names:
+            rows = [member.get("symbol_metrics", {}).get(symbol, {}) for member in members]
+            keys = {key for row in rows if isinstance(row, dict) for key in row}
+            averaged: dict[str, float] = {}
+            for key in keys:
+                values = [float(row[key]) for row in rows if isinstance(row.get(key), (int, float))]
+                if values:
+                    averaged[key] = sum(values) / len(values)
+            result["symbol_metrics"][symbol] = averaged
+    return result
+
+
 def _validation_metrics(model: nn.Module, prepared: PreparedData, clip: float) -> dict[str, float]:
     return _evaluation_metrics(
         model,
@@ -712,6 +1058,39 @@ def _validation_metrics(model: nn.Module, prepared: PreparedData, clip: float) -
     )
 
 
+def _validation_stability_metrics(
+    model: nn.Module,
+    prepared: PreparedData,
+    clip: float,
+    folds: int,
+) -> dict[str, float]:
+    fold_count = max(1, min(int(folds), len(prepared.validation_targets)))
+    fold_size = max(1, len(prepared.validation_targets) // fold_count)
+    maes: list[float] = []
+    for fold in range(fold_count):
+        start = fold * fold_size
+        end = len(prepared.validation_targets) if fold == fold_count - 1 else min(len(prepared.validation_targets), start + fold_size)
+        if end <= start:
+            continue
+        metrics = _evaluation_metrics(
+            model,
+            values=prepared.validation_x[start:end],
+            targets=prepared.validation_targets[start:end],
+            volatility_scales=prepared.validation_volatility_scales[start:end],
+            prepared=prepared,
+            clip=clip,
+        )
+        maes.append(float(metrics["validation_mae"]))
+    if not maes:
+        return {"validation_fold_mae_std": 0.0, "validation_fold_mae_worst": math.inf}
+    mean = sum(maes) / len(maes)
+    variance = sum((value - mean) ** 2 for value in maes) / len(maes)
+    return {
+        "validation_fold_mae_std": math.sqrt(variance),
+        "validation_fold_mae_worst": max(maes),
+    }
+
+
 def _holdout_metrics(model: nn.Module, prepared: PreparedData, clip: float) -> dict[str, Any]:
     metrics = _evaluation_metrics(
         model,
@@ -790,6 +1169,13 @@ def _evaluation_metrics(
         if prediction > 0
     ]
     buy_wins = [actual for actual in buy_predictions if actual > 0]
+    ranked = sorted(
+        zip(decision_predictions, [row[decision] for row in probabilities], decision_targets),
+        key=lambda item: item[0] * max(0.0, item[1] - 0.5),
+        reverse=True,
+    )
+    selected = ranked[: max(8, len(ranked) // 5)]
+    selected_targets = [row[2] for row in selected]
     by_horizon = {}
     baseline_by_horizon = {}
     for horizon_index, horizon in enumerate(prepared.target_horizons):
@@ -815,6 +1201,12 @@ def _evaluation_metrics(
         "directional_accuracy": len(correct) / len(non_zero) if non_zero else 0.0,
         "buy_precision": len(buy_wins) / len(buy_predictions) if buy_predictions else 0.0,
         "probability_brier": sum(probability_errors) / len(probability_errors) if probability_errors else 1.0,
+        "validation_trade_mean": sum(selected_targets) / len(selected_targets) if selected_targets else 0.0,
+        "validation_trade_win_rate": (
+            sum(1 for value in selected_targets if value > 0) / len(selected_targets)
+            if selected_targets
+            else 0.0
+        ),
     }
 
 
@@ -824,9 +1216,13 @@ def _candidate_score(row: dict[str, Any]) -> float:
     directional = float(row.get("directional_accuracy", 0.0))
     buy_precision = float(row.get("buy_precision", 0.0))
     probability_brier = float(row.get("probability_brier", 1.0))
-    return mae * (1.0 - max(0.0, skill) * 0.05) * (1.0 - max(0.0, directional - 0.5) * 0.03) * (
+    fold_std = max(0.0, float(row.get("validation_fold_mae_std", 0.0)))
+    stability_penalty = 1.0 + min(1.0, fold_std / max(mae, 1e-9)) * 0.25
+    trade_mean = float(row.get("validation_trade_mean", 0.0))
+    trade_penalty = max(0.0, -trade_mean) * 2.0 - max(0.0, trade_mean) * 0.5
+    return mae * stability_penalty * (1.0 - max(0.0, skill) * 0.05) * (1.0 - max(0.0, directional - 0.5) * 0.03) * (
         1.0 - max(0.0, buy_precision - 0.5) * 0.02
-    ) * (1.0 + max(0.0, probability_brier - 0.25) * 0.02)
+    ) * (1.0 + max(0.0, probability_brier - 0.25) * 0.02) + trade_penalty
 
 
 def _forecast_loss(outputs: torch.Tensor, targets: torch.Tensor, up_targets: torch.Tensor, horizon_count: int) -> torch.Tensor:
@@ -842,7 +1238,19 @@ def _forecast_loss(outputs: torch.Tensor, targets: torch.Tensor, up_targets: tor
     probabilities = torch.sigmoid(logits)
     pt = probabilities * up_targets + (1.0 - probabilities) * (1.0 - up_targets)
     focal = ((1.0 - pt) ** 2.0 * bce).mean()
-    return mean_loss + 0.35 * sum(quantile_losses) / len(quantile_losses) + 0.15 * focal
+    soft_long = torch.sigmoid(values[:, :, 0] * 2.0) * probabilities
+    after_cost_utility = -(soft_long * targets).mean()
+    prediction_centered = values[:, :, 0] - values[:, :, 0].mean(dim=0, keepdim=True)
+    target_centered = targets - targets.mean(dim=0, keepdim=True)
+    cosine = nn.functional.cosine_similarity(prediction_centered, target_centered, dim=0).mean()
+    ranking_loss = 1.0 - cosine
+    return (
+        mean_loss
+        + 0.35 * sum(quantile_losses) / len(quantile_losses)
+        + 0.15 * focal
+        + 0.10 * after_cost_utility
+        + 0.05 * ranking_loss
+    )
 
 
 def _export_recurrent_state(model: RecurrentReturnModel) -> dict[str, Any]:
diff --git a/tools/windows_training_agent.py b/tools/windows_training_agent.py
index f949d7a..a9f00b8 100644
--- a/tools/windows_training_agent.py
+++ b/tools/windows_training_agent.py
@@ -64,19 +64,20 @@ def poll_once(args: argparse.Namespace, repo_root: Path, runtime_dir: Path, log_
     try:
         run_retrain(args, job_id, job, repo_root, log_path)
         summary = read_json(runtime_dir / "torch_retrain_guard.json")
-        if summary.get("accepted") is not True:
-            raise RuntimeError(
-                "candidate rejected by untouched-holdout guard: "
-                + str(summary.get("reason") or "validation failed")
-            )
-        report_progress(args, job_id, "running", "uploading", 72, "Обучение завершено, загружаю артефакты")
-        for name in ARTIFACT_NAMES:
-            path = runtime_dir / name
-            if path.is_file():
-                upload_artifact(args, job_id, path, log_path)
+        accepted = summary.get("accepted") is True
+        if accepted:
+            report_progress(args, job_id, "running", "uploading", 72, "Обучение завершено, загружаю артефакты")
+            for name in ARTIFACT_NAMES:
+                path = runtime_dir / name
+                if path.is_file():
+                    upload_artifact(args, job_id, path, log_path)
+            message = "training completed; candidate accepted"
+            log(log_path, f"Completed retrain job {job_id}; candidate accepted")
+        else:
+            reason = str(summary.get("reason") or "validation failed")
+            message = f"training completed; candidate rejected by quality gate: {reason}"
+            log(log_path, f"Completed retrain job {job_id}; candidate rejected: {reason}")
         success = True
-        message = "training completed"
-        log(log_path, f"Completed retrain job {job_id}")
     except Exception as exc:  # noqa: BLE001 - report failure to the bot.
         message = str(exc)
         log(log_path, f"Job {job_id} failed: {message}")
@@ -108,6 +109,10 @@ def run_retrain(args: argparse.Namespace, job_id: str, job: dict[str, Any], repo
         "dropouts": "-Dropouts",
         "epochs": "-Epochs",
         "holdout_window": "-HoldoutWindow",
+        "ensemble_seeds": "-EnsembleSeeds",
+        "selection_folds": "-SelectionFolds",
+        "learning_rate": "-LearningRate",
+        "weight_decay": "-WeightDecay",
     }
     for key, ps_arg in arg_map.items():
         value = parameters.get(key)