diff --git a/Cargo.lock b/Cargo.lock
index bb839dc6d..4274d0ea4 100644
--- a/Cargo.lock
+++ b/Cargo.lock
@@ -2318,6 +2318,7 @@ dependencies = [
  "opentelemetry-otlp",
  "opentelemetry_sdk",
  "prometheus",
+ "questdb-rs",
  "rand 0.8.5",
  "redis",
  "rust_decimal",
@@ -3283,6 +3284,18 @@ dependencies = [
  "libloading",
 ]
 
+[[package]]
+name = "dns-lookup"
+version = "2.1.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "cf5597a4b7fe5275fc9dcf88ce26326bc8e4cb87d0130f33752d4c5f717793cf"
+dependencies = [
+ "cfg-if",
+ "libc",
+ "socket2 0.6.0",
+ "windows-sys 0.60.2",
+]
+
 [[package]]
 name = "doc-comment"
 version = "0.3.3"
@@ -4880,6 +4893,15 @@ dependencies = [
  "web-time",
 ]
 
+[[package]]
+name = "indoc"
+version = "2.0.7"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "79cf5c93f93228cf8efb3ba362535fb11199ac548a09ce117c9b1adc3030d706"
+dependencies = [
+ "rustversion",
+]
+
 [[package]]
 name = "influxdb"
 version = "0.7.2"
@@ -7288,6 +7310,36 @@ dependencies = [
  "winapi",
 ]
 
+[[package]]
+name = "questdb-confstr"
+version = "0.1.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "7aceffde1cbf8e67f34cdfd70d2436396176d6ff648fa719e0231fb9856ef3e9"
+
+[[package]]
+name = "questdb-rs"
+version = "4.0.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "882489d3cc6b44ff276ed90b483afd44ae1d1d3cafd3ecceff58b83a2fdc9299"
+dependencies = [
+ "base64ct",
+ "dns-lookup",
+ "indoc",
+ "itoa",
+ "libc",
+ "questdb-confstr",
+ "ring",
+ "rustls 0.22.4",
+ "rustls-pemfile 2.2.0",
+ "rustls-pki-types",
+ "ryu",
+ "serde",
+ "serde_json",
+ "slugify",
+ "socket2 0.5.10",
+ "winapi",
+]
+
 [[package]]
 name = "quick-error"
 version = "1.2.3"
@@ -7732,7 +7784,7 @@ dependencies = [
  "percent-encoding",
  "pin-project-lite",
  "rustls 0.21.12",
- "rustls-pemfile",
+ "rustls-pemfile 1.0.4",
  "serde",
  "serde_json",
  "serde_urlencoded",
@@ -8200,7 +8252,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "a9aace74cb666635c918e9c12bc0d348266037aa8eb599b5cba565709a8dff00"
 dependencies = [
  "openssl-probe",
- "rustls-pemfile",
+ "rustls-pemfile 1.0.4",
  "schannel",
  "security-framework 2.11.1",
 ]
@@ -8226,6 +8278,15 @@ dependencies = [
  "base64 0.21.7",
 ]
 
+[[package]]
+name = "rustls-pemfile"
+version = "2.2.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "dce314e5fee3f39953d46bb63bb8a46d40c2f8fb7cc5a3b6cab2bde9721d6e50"
+dependencies = [
+ "rustls-pki-types",
+]
+
 [[package]]
 name = "rustls-pki-types"
 version = "1.12.0"
@@ -8799,6 +8860,15 @@ dependencies = [
  "time",
 ]
 
+[[package]]
+name = "slugify"
+version = "0.1.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "d6b8cf203d2088b831d7558f8e5151bfa420c57a34240b28cee29d0ae5f2ac8b"
+dependencies = [
+ "unidecode",
+]
+
 [[package]]
 name = "smallvec"
 version = "1.15.1"
@@ -10819,6 +10889,12 @@ version = "0.2.6"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "ebc1c04c71510c7f702b52b7c350734c9ff1295c464a03335b00bb84fc54f853"
 
+[[package]]
+name = "unidecode"
+version = "0.3.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "402bb19d8e03f1d1a7450e2bd613980869438e0666331be3e073089124aa1adc"
+
 [[package]]
 name = "universal-hash"
 version = "0.5.1"
diff --git a/Cargo.toml b/Cargo.toml
index 58bcef197..d60e01aa0 100644
--- a/Cargo.toml
+++ b/Cargo.toml
@@ -222,6 +222,7 @@ dashmap = { version = "6.0", features = ["serde"] }
 bytes = "1.5"
 smallvec = { version = "1.11", features = ["serde", "const_generics"] }
 prometheus = "0.14"
+questdb-rs = { version = "4", default-features = false }
 
 # MINIMAL numerical libraries only - ALL ML/GPU frameworks REMOVED from workspace
 nalgebra = { version = "0.33", features = ["serde", "rand"] }
diff --git a/common/Cargo.toml b/common/Cargo.toml
index 3ff6e8ad0..1c3f920c3 100644
--- a/common/Cargo.toml
+++ b/common/Cargo.toml
@@ -60,6 +60,7 @@ config = { path = "../config" }
 uuid = { workspace = true, features = ["v4", "serde"] }
 once_cell.workspace = true
 rand.workspace = true
+questdb-rs = { workspace = true, optional = true }
 
 # Metrics (Wave 5)
 prometheus.workspace = true
@@ -76,6 +77,7 @@ tempfile = "3.8"
 [features]
 default = ["database"]
 database = ["sqlx"]
+questdb = ["questdb-rs"]
 
 [[bench]]
 name = "ml_strategy_bench"
diff --git a/common/src/lib.rs b/common/src/lib.rs
index 357defbec..84eaf7236 100644
--- a/common/src/lib.rs
+++ b/common/src/lib.rs
@@ -39,6 +39,8 @@ pub mod thresholds;
 pub mod tls;
 pub mod traits;
 pub mod types;
+#[cfg(feature = "questdb")]
+pub mod questdb;
 
 // Re-export database types for external use
 pub use database::{DatabaseConfig, DatabaseError, DatabasePool};
diff --git a/common/src/questdb.rs b/common/src/questdb.rs
new file mode 100644
index 000000000..360eac592
--- /dev/null
+++ b/common/src/questdb.rs
@@ -0,0 +1,302 @@
+//! QuestDB client with ring buffer for graceful degradation
+//!
+//! Non-critical path: if QuestDB is unavailable, metrics buffer locally
+//! and flush when connection is restored. Trading never depends on QuestDB.
+
+use std::collections::VecDeque;
+use std::sync::atomic::{AtomicBool, Ordering};
+use std::sync::Arc;
+use std::time::{Duration, Instant};
+use tokio::sync::Mutex;
+use tracing::{debug, info, warn};
+
+/// Maximum entries in the ring buffer before oldest are dropped
+const DEFAULT_BUFFER_CAPACITY: usize = 10_000;
+
+/// Health check interval
+const HEALTH_CHECK_INTERVAL: Duration = Duration::from_secs(30);
+
+/// A metric entry waiting to be flushed to QuestDB
+#[derive(Debug, Clone)]
+pub struct MetricEntry {
+    pub table: String,
+    pub columns: Vec<(String, MetricValue)>,
+    pub timestamp_ns: i64,
+    pub created_at: Instant,
+}
+
+/// Supported metric value types for ILP
+#[derive(Debug, Clone)]
+pub enum MetricValue {
+    Symbol(String),
+    Float(f64),
+    Int(i64),
+    Bool(bool),
+}
+
+/// QuestDB connection configuration
+#[derive(Debug, Clone)]
+pub struct QuestDBConfig {
+    /// ILP host for writes (default: localhost)
+    pub ilp_host: String,
+    /// ILP port for writes (default: 9009)
+    pub ilp_port: u16,
+    /// PostgreSQL host for queries (default: localhost)
+    pub pg_host: String,
+    /// PostgreSQL port for queries (default: 8812)
+    pub pg_port: u16,
+    /// Ring buffer capacity (default: 10,000)
+    pub buffer_capacity: usize,
+}
+
+impl Default for QuestDBConfig {
+    fn default() -> Self {
+        Self {
+            ilp_host: "localhost".into(),
+            ilp_port: 9009,
+            pg_host: "localhost".into(),
+            pg_port: 8812,
+            buffer_capacity: DEFAULT_BUFFER_CAPACITY,
+        }
+    }
+}
+
+/// QuestDB client with graceful degradation via ring buffer
+pub struct QuestDBClient {
+    config: QuestDBConfig,
+    buffer: Arc<Mutex<VecDeque<MetricEntry>>>,
+    connected: Arc<AtomicBool>,
+    last_health_check: Arc<Mutex<Instant>>,
+}
+
+impl QuestDBClient {
+    /// Create a new QuestDB client (does not connect immediately)
+    pub fn new(config: QuestDBConfig) -> Self {
+        Self {
+            buffer: Arc::new(Mutex::new(VecDeque::with_capacity(config.buffer_capacity))),
+            connected: Arc::new(AtomicBool::new(false)),
+            last_health_check: Arc::new(Mutex::new(Instant::now())),
+            config,
+        }
+    }
+
+    /// Check if QuestDB is currently connected
+    pub fn is_connected(&self) -> bool {
+        self.connected.load(Ordering::Relaxed)
+    }
+
+    /// Get current buffer size (metrics waiting to be flushed)
+    pub async fn buffer_size(&self) -> usize {
+        self.buffer.lock().await.len()
+    }
+
+    /// Get age of oldest buffered entry
+    pub async fn buffer_age(&self) -> Duration {
+        let buf = self.buffer.lock().await;
+        buf.front()
+            .map(|e| e.created_at.elapsed())
+            .unwrap_or(Duration::ZERO)
+    }
+
+    /// Write a metric entry (buffers if QuestDB is unavailable)
+    pub async fn write(&self, entry: MetricEntry) {
+        let mut buf = self.buffer.lock().await;
+
+        // If buffer is full, drop oldest entry
+        if buf.len() >= self.config.buffer_capacity {
+            buf.pop_front();
+            debug!("QuestDB buffer full, dropped oldest entry");
+        }
+
+        buf.push_back(entry);
+
+        // Try to flush if connected
+        if self.connected.load(Ordering::Relaxed) {
+            drop(buf); // Release lock before flush
+            if let Err(e) = self.try_flush().await {
+                warn!("QuestDB flush failed: {}", e);
+                self.connected.store(false, Ordering::Relaxed);
+            }
+        }
+    }
+
+    /// Attempt to flush buffered entries to QuestDB via ILP
+    async fn try_flush(&self) -> Result<(), Box<dyn std::error::Error + Send + Sync>> {
+        let mut buf = self.buffer.lock().await;
+        if buf.is_empty() {
+            return Ok(());
+        }
+
+        let mut sender = questdb::ingress::SenderBuilder::new(
+            questdb::ingress::Protocol::Tcp,
+            &self.config.ilp_host,
+            self.config.ilp_port,
+        )
+        .build()?;
+
+        let mut ilp_buf = questdb::ingress::Buffer::new();
+        let entries_to_flush: Vec<MetricEntry> = buf.drain(..).collect();
+        let count = entries_to_flush.len();
+
+        for entry in &entries_to_flush {
+            ilp_buf.table(entry.table.as_str())?;
+            for (name, value) in &entry.columns {
+                match value {
+                    MetricValue::Symbol(s) => {
+                        ilp_buf.symbol(name.as_str(), s.as_str())?;
+                    }
+                    MetricValue::Float(f) => {
+                        ilp_buf.column_f64(name.as_str(), *f)?;
+                    }
+                    MetricValue::Int(i) => {
+                        ilp_buf.column_i64(name.as_str(), *i)?;
+                    }
+                    MetricValue::Bool(b) => {
+                        ilp_buf.column_bool(name.as_str(), *b)?;
+                    }
+                }
+            }
+            ilp_buf.at(questdb::ingress::TimestampNanos::new(entry.timestamp_ns))?;
+        }
+
+        sender.flush(&mut ilp_buf)?;
+        info!("Flushed {} entries to QuestDB", count);
+
+        Ok(())
+    }
+
+    /// Periodic health check — call from a background task
+    pub async fn health_check(&self) -> bool {
+        let mut last = self.last_health_check.lock().await;
+        if last.elapsed() < HEALTH_CHECK_INTERVAL {
+            return self.connected.load(Ordering::Relaxed);
+        }
+        *last = Instant::now();
+        drop(last);
+
+        // Try connecting via ILP
+        let connected = match questdb::ingress::SenderBuilder::new(
+            questdb::ingress::Protocol::Tcp,
+            &self.config.ilp_host,
+            self.config.ilp_port,
+        )
+        .build()
+        {
+            Ok(_sender) => {
+                if !self.connected.load(Ordering::Relaxed) {
+                    info!("QuestDB connection restored");
+                }
+                true
+            }
+            Err(e) => {
+                if self.connected.load(Ordering::Relaxed) {
+                    warn!("QuestDB connection lost: {}", e);
+                }
+                false
+            }
+        };
+
+        self.connected.store(connected, Ordering::Relaxed);
+
+        // If reconnected, try flushing buffer
+        if connected {
+            if let Err(e) = self.try_flush().await {
+                warn!("QuestDB reconnect flush failed: {}", e);
+            }
+        }
+
+        connected
+    }
+
+    /// Get ILP connection string for the configured host
+    pub fn ilp_address(&self) -> String {
+        format!("{}:{}", self.config.ilp_host, self.config.ilp_port)
+    }
+
+    /// Get PostgreSQL connection string for queries
+    pub fn pg_connection_string(&self) -> String {
+        format!(
+            "postgresql://admin:quest@{}:{}/qdb",
+            self.config.pg_host, self.config.pg_port
+        )
+    }
+}
+
+impl std::fmt::Debug for QuestDBClient {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        f.debug_struct("QuestDBClient")
+            .field("connected", &self.connected.load(Ordering::Relaxed))
+            .field("config", &self.config)
+            .finish()
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[tokio::test]
+    async fn test_client_creation() {
+        let client = QuestDBClient::new(QuestDBConfig::default());
+        assert!(!client.is_connected());
+        assert_eq!(client.buffer_size().await, 0);
+    }
+
+    #[tokio::test]
+    async fn test_buffer_stores_entries() {
+        let client = QuestDBClient::new(QuestDBConfig::default());
+        let entry = MetricEntry {
+            table: "test".into(),
+            columns: vec![("value".into(), MetricValue::Float(1.0))],
+            timestamp_ns: 1_234_567_890,
+            created_at: Instant::now(),
+        };
+        client.write(entry).await;
+        assert_eq!(client.buffer_size().await, 1);
+    }
+
+    #[tokio::test]
+    async fn test_buffer_capacity_limit() {
+        let config = QuestDBConfig {
+            buffer_capacity: 3,
+            ..Default::default()
+        };
+        let client = QuestDBClient::new(config);
+        for i in 0..5 {
+            let entry = MetricEntry {
+                table: "test".into(),
+                columns: vec![("i".into(), MetricValue::Int(i))],
+                timestamp_ns: i,
+                created_at: Instant::now(),
+            };
+            client.write(entry).await;
+        }
+        // Buffer capacity is 3, so oldest 2 should be dropped
+        assert_eq!(client.buffer_size().await, 3);
+    }
+
+    #[tokio::test]
+    async fn test_buffer_age_empty() {
+        let client = QuestDBClient::new(QuestDBConfig::default());
+        assert_eq!(client.buffer_age().await, Duration::ZERO);
+    }
+
+    #[tokio::test]
+    async fn test_default_config() {
+        let config = QuestDBConfig::default();
+        assert_eq!(config.ilp_host, "localhost");
+        assert_eq!(config.ilp_port, 9009);
+        assert_eq!(config.pg_host, "localhost");
+        assert_eq!(config.pg_port, 8812);
+        assert_eq!(config.buffer_capacity, 10_000);
+    }
+
+    #[test]
+    fn test_pg_connection_string() {
+        let client = QuestDBClient::new(QuestDBConfig::default());
+        assert_eq!(
+            client.pg_connection_string(),
+            "postgresql://admin:quest@localhost:8812/qdb"
+        );
+    }
+}
diff --git a/docker-compose.yml b/docker-compose.yml
index e2c1fe6f7..dea3009ff 100644
--- a/docker-compose.yml
+++ b/docker-compose.yml
@@ -41,6 +41,27 @@ services:
     networks:
       - foxhunt-network
 
+  # QuestDB - High-performance time-series DB for ML metrics and feedback loop
+  questdb:
+    image: questdb/questdb:8.2.3
+    container_name: foxhunt-questdb
+    ports:
+      - "9009:9009"    # ILP ingestion (Influx Line Protocol)
+      - "8812:8812"    # PostgreSQL wire protocol (SQL queries)
+      - "9003:9003"    # HTTP REST API + Web Console
+    volumes:
+      - questdb_data:/var/lib/questdb
+    environment:
+      - QDB_PG_ENABLED=true
+      - QDB_LINE_TCP_NET_BIND_TO=0.0.0.0:9009
+    healthcheck:
+      test: ["CMD-SHELL", "wget --no-verbose --tries=1 -O /dev/null http://localhost:9003/exec?query=SELECT%201 || exit 1"]
+      interval: 10s
+      timeout: 5s
+      retries: 5
+    networks:
+      - foxhunt-network
+
   # InfluxDB - Time-series data for HFT metrics
   influxdb:
     image: influxdb:2.7-alpine
@@ -190,6 +211,9 @@ services:
       # mTLS Validation Options
       - MTLS_ENABLE_REVOCATION_CHECK=${MTLS_ENABLE_REVOCATION_CHECK:-false}
       - MTLS_CRL_URL=${MTLS_CRL_URL:-}
+      # QuestDB for ML feedback loop metrics
+      - QUESTDB_ILP_HOST=questdb:9009
+      - QUESTDB_PG_URL=postgresql://admin:quest@questdb:8812/qdb
       - KILL_SWITCH_SOCKET_PATH=/tmp/kill_switch.sock
       - GRPC_PORT=50051
       - RUST_LOG=info
@@ -477,6 +501,7 @@ services:
 volumes:
   postgres_data:
   redis_data:
+  questdb_data:
   influxdb_data:
   vault_data:
   prometheus_data:
diff --git a/ml/src/ensemble/conviction_gates.rs b/ml/src/ensemble/conviction_gates.rs
new file mode 100644
index 000000000..3113b235c
--- /dev/null
+++ b/ml/src/ensemble/conviction_gates.rs
@@ -0,0 +1,566 @@
+//! 7-Gate Conviction System
+//!
+//! Evaluates ensemble predictions through a series of quality gates before
+//! allowing trade execution. Gates are evaluated in order; any failure
+//! results in HOLD.
+
+use serde::{Deserialize, Serialize};
+
+/// Trading session windows (Eastern Time)
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)]
+pub enum TradingSession {
+    PreMarket,
+    Regular,
+    AfterHours,
+}
+
+/// Configuration for the 7-gate conviction system
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct ConvictionGateConfig {
+    pub model_health_threshold: f64,
+    pub allowed_sessions: Vec<TradingSession>,
+    pub min_confidence: f64,
+    pub max_disagreement: f64,
+    pub min_quorum: f64,
+    pub regime_tightening_factor: f64,
+    pub high_vol_threshold: f64,
+    pub conviction_scaling_enabled: bool,
+}
+
+impl Default for ConvictionGateConfig {
+    fn default() -> Self {
+        Self {
+            model_health_threshold: 0.70,
+            allowed_sessions: vec![TradingSession::Regular],
+            min_confidence: 0.60,
+            max_disagreement: 0.40,
+            min_quorum: 0.60,
+            regime_tightening_factor: 0.80,
+            high_vol_threshold: 0.03,
+            conviction_scaling_enabled: true,
+        }
+    }
+}
+
+/// Which gate rejected the decision (if any)
+#[derive(Debug, Clone, PartialEq)]
+pub enum GateRejection {
+    ModelHealth {
+        healthy_ratio: f64,
+        threshold: f64,
+    },
+    TimeOfDay {
+        current_session: TradingSession,
+    },
+    Confidence {
+        confidence: f64,
+        threshold: f64,
+    },
+    Agreement {
+        disagreement: f64,
+        threshold: f64,
+    },
+    Quorum {
+        quorum_ratio: f64,
+        threshold: f64,
+    },
+    Regime {
+        adjusted_confidence: f64,
+        adjusted_threshold: f64,
+    },
+}
+
+/// Result of passing all conviction gates
+#[derive(Debug, Clone)]
+pub struct GatePassResult {
+    pub conviction_score: f64,
+    pub gate_details: Vec<GateEvaluation>,
+}
+
+/// Individual gate evaluation result
+#[derive(Debug, Clone)]
+pub struct GateEvaluation {
+    pub gate_name: String,
+    pub value: f64,
+    pub threshold: f64,
+    pub passed: bool,
+}
+
+/// Outcome of evaluating all gates
+#[derive(Debug, Clone)]
+pub enum ConvictionGateOutcome {
+    Passed(GatePassResult),
+    Rejected(GateRejection),
+}
+
+/// Input data for gate evaluation
+#[derive(Debug, Clone)]
+pub struct GateInput {
+    pub confidence: f64,
+    pub disagreement_rate: f64,
+    pub quorum_ratio: f64,
+    pub healthy_models: usize,
+    pub total_models: usize,
+    pub current_session: TradingSession,
+    pub regime_volatility: f64,
+}
+
+/// Evaluates ensemble decisions through 7 conviction gates
+#[derive(Debug, Clone)]
+pub struct ConvictionGateEvaluator {
+    config: ConvictionGateConfig,
+}
+
+impl ConvictionGateEvaluator {
+    pub fn new(config: ConvictionGateConfig) -> Self {
+        Self { config }
+    }
+
+    pub fn config(&self) -> &ConvictionGateConfig {
+        &self.config
+    }
+
+    pub fn config_mut(&mut self) -> &mut ConvictionGateConfig {
+        &mut self.config
+    }
+
+    /// Evaluate all 7 gates in order. Returns on first rejection.
+    pub fn evaluate(&self, input: &GateInput) -> ConvictionGateOutcome {
+        let mut gate_details = Vec::with_capacity(7);
+
+        // Gate 1: Model Health - healthy_ratio >= threshold
+        let healthy_ratio = if input.total_models == 0 {
+            0.0
+        } else {
+            input.healthy_models as f64 / input.total_models as f64
+        };
+
+        let gate1_passed = healthy_ratio >= self.config.model_health_threshold;
+        gate_details.push(GateEvaluation {
+            gate_name: "model_health".to_string(),
+            value: healthy_ratio,
+            threshold: self.config.model_health_threshold,
+            passed: gate1_passed,
+        });
+        if !gate1_passed {
+            return ConvictionGateOutcome::Rejected(GateRejection::ModelHealth {
+                healthy_ratio,
+                threshold: self.config.model_health_threshold,
+            });
+        }
+
+        // Gate 2: Time-of-Day - current_session in allowed_sessions
+        let gate2_passed = self.config.allowed_sessions.contains(&input.current_session);
+        gate_details.push(GateEvaluation {
+            gate_name: "time_of_day".to_string(),
+            value: if gate2_passed { 1.0 } else { 0.0 },
+            threshold: 1.0,
+            passed: gate2_passed,
+        });
+        if !gate2_passed {
+            return ConvictionGateOutcome::Rejected(GateRejection::TimeOfDay {
+                current_session: input.current_session,
+            });
+        }
+
+        // Gate 6 prep: If regime_volatility > high_vol_threshold, tighten thresholds
+        let high_vol = input.regime_volatility > self.config.high_vol_threshold;
+        let (adj_min_confidence, adj_max_disagreement, adj_min_quorum) = if high_vol {
+            let factor = self.config.regime_tightening_factor;
+            let adj_conf = (self.config.min_confidence / factor).min(0.95);
+            let adj_disagree = (self.config.max_disagreement * factor).max(0.05);
+            let adj_quorum = (self.config.min_quorum / factor).min(0.95);
+            (adj_conf, adj_disagree, adj_quorum)
+        } else {
+            (
+                self.config.min_confidence,
+                self.config.max_disagreement,
+                self.config.min_quorum,
+            )
+        };
+
+        // Gate 3: Confidence >= adj_min_confidence
+        let gate3_passed = input.confidence >= adj_min_confidence;
+        gate_details.push(GateEvaluation {
+            gate_name: "confidence".to_string(),
+            value: input.confidence,
+            threshold: adj_min_confidence,
+            passed: gate3_passed,
+        });
+        if !gate3_passed {
+            return ConvictionGateOutcome::Rejected(GateRejection::Confidence {
+                confidence: input.confidence,
+                threshold: adj_min_confidence,
+            });
+        }
+
+        // Gate 4: Disagreement <= adj_max_disagreement
+        let gate4_passed = input.disagreement_rate <= adj_max_disagreement;
+        gate_details.push(GateEvaluation {
+            gate_name: "agreement".to_string(),
+            value: input.disagreement_rate,
+            threshold: adj_max_disagreement,
+            passed: gate4_passed,
+        });
+        if !gate4_passed {
+            return ConvictionGateOutcome::Rejected(GateRejection::Agreement {
+                disagreement: input.disagreement_rate,
+                threshold: adj_max_disagreement,
+            });
+        }
+
+        // Gate 5: Quorum >= adj_min_quorum
+        let gate5_passed = input.quorum_ratio >= adj_min_quorum;
+        gate_details.push(GateEvaluation {
+            gate_name: "quorum".to_string(),
+            value: input.quorum_ratio,
+            threshold: adj_min_quorum,
+            passed: gate5_passed,
+        });
+        if !gate5_passed {
+            return ConvictionGateOutcome::Rejected(GateRejection::Quorum {
+                quorum_ratio: input.quorum_ratio,
+                threshold: adj_min_quorum,
+            });
+        }
+
+        // Gate 6: Regime (already applied via threshold adjustments)
+        // Record the regime gate evaluation
+        let regime_value = if high_vol {
+            input.confidence
+        } else {
+            input.confidence
+        };
+        let regime_threshold = adj_min_confidence;
+        gate_details.push(GateEvaluation {
+            gate_name: "regime".to_string(),
+            value: regime_value,
+            threshold: regime_threshold,
+            passed: true, // Already enforced via adjusted thresholds in gates 3-5
+        });
+
+        // Gate 7: Conviction Sizing
+        let conviction_score = if self.config.conviction_scaling_enabled {
+            input.confidence
+                * (1.0 - input.disagreement_rate)
+                * input.quorum_ratio
+                * healthy_ratio
+        } else {
+            1.0
+        };
+        gate_details.push(GateEvaluation {
+            gate_name: "conviction_sizing".to_string(),
+            value: conviction_score,
+            threshold: 0.0, // No minimum threshold for conviction sizing
+            passed: true,
+        });
+
+        ConvictionGateOutcome::Passed(GatePassResult {
+            conviction_score,
+            gate_details,
+        })
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn default_passing_input() -> GateInput {
+        GateInput {
+            confidence: 0.80,
+            disagreement_rate: 0.10,
+            quorum_ratio: 0.80,
+            healthy_models: 8,
+            total_models: 10,
+            current_session: TradingSession::Regular,
+            regime_volatility: 0.01,
+        }
+    }
+
+    #[test]
+    fn test_default_config_values() {
+        let config = ConvictionGateConfig::default();
+        assert!((config.model_health_threshold - 0.70).abs() < f64::EPSILON);
+        assert_eq!(config.allowed_sessions, vec![TradingSession::Regular]);
+        assert!((config.min_confidence - 0.60).abs() < f64::EPSILON);
+        assert!((config.max_disagreement - 0.40).abs() < f64::EPSILON);
+        assert!((config.min_quorum - 0.60).abs() < f64::EPSILON);
+        assert!((config.regime_tightening_factor - 0.80).abs() < f64::EPSILON);
+        assert!((config.high_vol_threshold - 0.03).abs() < f64::EPSILON);
+        assert!(config.conviction_scaling_enabled);
+    }
+
+    #[test]
+    fn test_gate_rejection_variants() {
+        let r1 = GateRejection::ModelHealth {
+            healthy_ratio: 0.5,
+            threshold: 0.7,
+        };
+        let r2 = GateRejection::ModelHealth {
+            healthy_ratio: 0.5,
+            threshold: 0.7,
+        };
+        assert_eq!(r1, r2);
+
+        let r3 = GateRejection::TimeOfDay {
+            current_session: TradingSession::PreMarket,
+        };
+        assert_ne!(r1, r3);
+    }
+
+    #[test]
+    fn test_gate1_model_health_passes() {
+        let evaluator = ConvictionGateEvaluator::new(ConvictionGateConfig::default());
+        let input = GateInput {
+            healthy_models: 8,
+            total_models: 10,
+            ..default_passing_input()
+        };
+        let outcome = evaluator.evaluate(&input);
+        assert!(matches!(outcome, ConvictionGateOutcome::Passed(_)));
+    }
+
+    #[test]
+    fn test_gate1_model_health_rejects() {
+        let evaluator = ConvictionGateEvaluator::new(ConvictionGateConfig::default());
+        let input = GateInput {
+            healthy_models: 5,
+            total_models: 10,
+            ..default_passing_input()
+        };
+        let outcome = evaluator.evaluate(&input);
+        if let ConvictionGateOutcome::Rejected(GateRejection::ModelHealth {
+            healthy_ratio,
+            threshold,
+        }) = outcome
+        {
+            assert!((healthy_ratio - 0.5).abs() < f64::EPSILON);
+            assert!((threshold - 0.70).abs() < f64::EPSILON);
+        } else {
+            panic!("Expected ModelHealth rejection, got {:?}", outcome);
+        }
+    }
+
+    #[test]
+    fn test_gate2_time_of_day_rejects_premarket() {
+        let evaluator = ConvictionGateEvaluator::new(ConvictionGateConfig::default());
+        let input = GateInput {
+            current_session: TradingSession::PreMarket,
+            ..default_passing_input()
+        };
+        let outcome = evaluator.evaluate(&input);
+        if let ConvictionGateOutcome::Rejected(GateRejection::TimeOfDay { current_session }) =
+            outcome
+        {
+            assert_eq!(current_session, TradingSession::PreMarket);
+        } else {
+            panic!("Expected TimeOfDay rejection, got {:?}", outcome);
+        }
+    }
+
+    #[test]
+    fn test_gate3_confidence_rejects_low() {
+        let evaluator = ConvictionGateEvaluator::new(ConvictionGateConfig::default());
+        let input = GateInput {
+            confidence: 0.45,
+            ..default_passing_input()
+        };
+        let outcome = evaluator.evaluate(&input);
+        if let ConvictionGateOutcome::Rejected(GateRejection::Confidence {
+            confidence,
+            threshold,
+        }) = outcome
+        {
+            assert!((confidence - 0.45).abs() < f64::EPSILON);
+            assert!((threshold - 0.60).abs() < f64::EPSILON);
+        } else {
+            panic!("Expected Confidence rejection, got {:?}", outcome);
+        }
+    }
+
+    #[test]
+    fn test_gate4_agreement_rejects_high_disagreement() {
+        let evaluator = ConvictionGateEvaluator::new(ConvictionGateConfig::default());
+        let input = GateInput {
+            disagreement_rate: 0.55,
+            ..default_passing_input()
+        };
+        let outcome = evaluator.evaluate(&input);
+        if let ConvictionGateOutcome::Rejected(GateRejection::Agreement {
+            disagreement,
+            threshold,
+        }) = outcome
+        {
+            assert!((disagreement - 0.55).abs() < f64::EPSILON);
+            assert!((threshold - 0.40).abs() < f64::EPSILON);
+        } else {
+            panic!("Expected Agreement rejection, got {:?}", outcome);
+        }
+    }
+
+    #[test]
+    fn test_gate5_quorum_rejects_low() {
+        let evaluator = ConvictionGateEvaluator::new(ConvictionGateConfig::default());
+        let input = GateInput {
+            quorum_ratio: 0.40,
+            ..default_passing_input()
+        };
+        let outcome = evaluator.evaluate(&input);
+        if let ConvictionGateOutcome::Rejected(GateRejection::Quorum {
+            quorum_ratio,
+            threshold,
+        }) = outcome
+        {
+            assert!((quorum_ratio - 0.40).abs() < f64::EPSILON);
+            assert!((threshold - 0.60).abs() < f64::EPSILON);
+        } else {
+            panic!("Expected Quorum rejection, got {:?}", outcome);
+        }
+    }
+
+    #[test]
+    fn test_conviction_score_calculation() {
+        let evaluator = ConvictionGateEvaluator::new(ConvictionGateConfig::default());
+        let input = GateInput {
+            confidence: 0.80,
+            disagreement_rate: 0.10,
+            quorum_ratio: 0.80,
+            healthy_models: 9,
+            total_models: 10,
+            current_session: TradingSession::Regular,
+            regime_volatility: 0.01,
+        };
+        let outcome = evaluator.evaluate(&input);
+        if let ConvictionGateOutcome::Passed(result) = outcome {
+            // score = 0.80 * (1.0 - 0.10) * 0.80 * (9/10) = 0.80 * 0.90 * 0.80 * 0.90 = 0.5184
+            let expected = 0.80 * 0.90 * 0.80 * 0.90;
+            assert!(
+                (result.conviction_score - expected).abs() < 1e-10,
+                "Expected {}, got {}",
+                expected,
+                result.conviction_score
+            );
+        } else {
+            panic!("Expected Passed outcome, got {:?}", outcome);
+        }
+    }
+
+    #[test]
+    fn test_gate6_regime_tightens_in_high_vol() {
+        let evaluator = ConvictionGateEvaluator::new(ConvictionGateConfig::default());
+        // With default config: min_confidence=0.60, tightening=0.80
+        // High vol: adj_min_confidence = 0.60 / 0.80 = 0.75
+        // 0.70 confidence < 0.75 threshold => rejected
+        let input = GateInput {
+            confidence: 0.70,
+            regime_volatility: 0.05,
+            ..default_passing_input()
+        };
+        let outcome = evaluator.evaluate(&input);
+        if let ConvictionGateOutcome::Rejected(GateRejection::Confidence {
+            confidence,
+            threshold,
+        }) = outcome
+        {
+            assert!((confidence - 0.70).abs() < f64::EPSILON);
+            assert!((threshold - 0.75).abs() < f64::EPSILON);
+        } else {
+            panic!(
+                "Expected Confidence rejection due to regime tightening, got {:?}",
+                outcome
+            );
+        }
+    }
+
+    #[test]
+    fn test_gate6_regime_no_effect_in_low_vol() {
+        let evaluator = ConvictionGateEvaluator::new(ConvictionGateConfig::default());
+        // Low vol: thresholds unchanged, min_confidence=0.60
+        // 0.65 confidence >= 0.60 => passes
+        let input = GateInput {
+            confidence: 0.65,
+            regime_volatility: 0.01,
+            ..default_passing_input()
+        };
+        let outcome = evaluator.evaluate(&input);
+        assert!(
+            matches!(outcome, ConvictionGateOutcome::Passed(_)),
+            "Expected Passed in low vol, got {:?}",
+            outcome
+        );
+    }
+
+    #[test]
+    fn test_zero_models_rejects() {
+        let evaluator = ConvictionGateEvaluator::new(ConvictionGateConfig::default());
+        let input = GateInput {
+            healthy_models: 0,
+            total_models: 0,
+            ..default_passing_input()
+        };
+        let outcome = evaluator.evaluate(&input);
+        if let ConvictionGateOutcome::Rejected(GateRejection::ModelHealth {
+            healthy_ratio,
+            threshold,
+        }) = outcome
+        {
+            assert!((healthy_ratio - 0.0).abs() < f64::EPSILON);
+            assert!((threshold - 0.70).abs() < f64::EPSILON);
+        } else {
+            panic!("Expected ModelHealth rejection for 0/0 models, got {:?}", outcome);
+        }
+    }
+
+    #[test]
+    fn test_conviction_scaling_disabled() {
+        let mut config = ConvictionGateConfig::default();
+        config.conviction_scaling_enabled = false;
+        let evaluator = ConvictionGateEvaluator::new(config);
+        let input = default_passing_input();
+        let outcome = evaluator.evaluate(&input);
+        if let ConvictionGateOutcome::Passed(result) = outcome {
+            assert!(
+                (result.conviction_score - 1.0).abs() < f64::EPSILON,
+                "Expected 1.0 when scaling disabled, got {}",
+                result.conviction_score
+            );
+        } else {
+            panic!("Expected Passed outcome, got {:?}", outcome);
+        }
+    }
+
+    #[test]
+    fn test_gate_details_count() {
+        let evaluator = ConvictionGateEvaluator::new(ConvictionGateConfig::default());
+        let input = default_passing_input();
+        let outcome = evaluator.evaluate(&input);
+        if let ConvictionGateOutcome::Passed(result) = outcome {
+            assert_eq!(
+                result.gate_details.len(),
+                7,
+                "Expected 7 gate evaluations, got {}",
+                result.gate_details.len()
+            );
+        } else {
+            panic!("Expected Passed outcome, got {:?}", outcome);
+        }
+    }
+
+    #[test]
+    fn test_after_hours_allowed_when_configured() {
+        let mut config = ConvictionGateConfig::default();
+        config.allowed_sessions.push(TradingSession::AfterHours);
+        let evaluator = ConvictionGateEvaluator::new(config);
+        let input = GateInput {
+            current_session: TradingSession::AfterHours,
+            ..default_passing_input()
+        };
+        let outcome = evaluator.evaluate(&input);
+        assert!(
+            matches!(outcome, ConvictionGateOutcome::Passed(_)),
+            "Expected Passed for AfterHours when configured, got {:?}",
+            outcome
+        );
+    }
+}
diff --git a/ml/src/ensemble/coordinator.rs b/ml/src/ensemble/coordinator.rs
index 796cbd537..83b53a1ff 100644
--- a/ml/src/ensemble/coordinator.rs
+++ b/ml/src/ensemble/coordinator.rs
@@ -4,9 +4,13 @@
 //! from multiple ML models (DQN, PPO, TFT, MAMBA-2, Liquid, TLOB) for production trading decisions.
 //! Supports dynamic weighting based on performance and model diversity metrics.
 
+use crate::ensemble::conviction_gates::{
+    ConvictionGateConfig, ConvictionGateEvaluator, ConvictionGateOutcome, GateInput, TradingSession,
+};
 use crate::ensemble::inference_adapter::{FeatureVector, ModelInferenceAdapter};
 use crate::ensemble::{EnsembleDecision, ModelVote, ModelWeight, TradingAction};
 use crate::{Features, MLError, MLResult, ModelPrediction};
+use chrono::{Timelike, Utc};
 use std::collections::HashMap;
 use std::sync::Arc;
 use tokio::sync::RwLock;
@@ -34,6 +38,9 @@ pub struct EnsembleCoordinator {
 
     /// Real model inference adapters for production predictions
     adapters: Vec<Box<dyn ModelInferenceAdapter>>,
+
+    /// Conviction gate evaluator (optional — None means no gating)
+    conviction_gates: Option<ConvictionGateEvaluator>,
 }
 
 impl std::fmt::Debug for EnsembleCoordinator {
@@ -77,9 +84,28 @@ impl EnsembleCoordinator {
             model_weights: Arc::new(RwLock::new(HashMap::new())),
             config,
             adapters: Vec::new(),
+            conviction_gates: Some(ConvictionGateEvaluator::new(ConvictionGateConfig {
+                // Allow all sessions by default — trading service narrows to Regular
+                allowed_sessions: vec![
+                    TradingSession::PreMarket,
+                    TradingSession::Regular,
+                    TradingSession::AfterHours,
+                ],
+                ..ConvictionGateConfig::default()
+            })),
         }
     }
 
+    /// Replace conviction gate configuration
+    pub fn set_conviction_gates(&mut self, config: ConvictionGateConfig) {
+        self.conviction_gates = Some(ConvictionGateEvaluator::new(config));
+    }
+
+    /// Get mutable reference to conviction gate evaluator
+    pub fn conviction_gates_mut(&mut self) -> Option<&mut ConvictionGateEvaluator> {
+        self.conviction_gates.as_mut()
+    }
+
     /// Add a real model inference adapter to the ensemble.
     /// When an adapter's model_name matches a registered model, its predictions
     /// are used for ensemble inference.
@@ -129,6 +155,50 @@ impl EnsembleCoordinator {
             .aggregate(predictions, &*self.model_weights.read().await)
             .await?;
 
+        // Apply conviction gates if configured
+        if let Some(ref gates) = self.conviction_gates {
+            let gate_input = GateInput {
+                confidence: decision.confidence,
+                disagreement_rate: decision.disagreement_rate,
+                quorum_ratio: Self::calculate_quorum_ratio(&decision),
+                healthy_models: self.adapters.iter().filter(|a| a.is_ready()).count(),
+                total_models: self.adapters.len(),
+                current_session: Self::current_trading_session(),
+                regime_volatility: Self::extract_regime_volatility(features),
+            };
+
+            match gates.evaluate(&gate_input) {
+                ConvictionGateOutcome::Passed(pass_result) => {
+                    info!(
+                        "Conviction gates passed: score={:.3}, {} gates evaluated",
+                        pass_result.conviction_score,
+                        pass_result.gate_details.len()
+                    );
+                    let mut gated_decision = decision;
+                    gated_decision.metadata.insert(
+                        "conviction_score".into(),
+                        serde_json::Value::from(pass_result.conviction_score),
+                    );
+                    return Ok(gated_decision);
+                }
+                ConvictionGateOutcome::Rejected(rejection) => {
+                    info!("Conviction gate rejected: {:?} — forcing HOLD", rejection);
+                    let mut hold_decision = EnsembleDecision::new(
+                        TradingAction::Hold,
+                        decision.confidence,
+                        0.0,
+                        decision.disagreement_rate,
+                        decision.model_votes,
+                    );
+                    hold_decision.metadata.insert(
+                        "gate_rejection".into(),
+                        serde_json::Value::String(format!("{:?}", rejection)),
+                    );
+                    return Ok(hold_decision);
+                }
+            }
+        }
+
         info!(
             "Ensemble decision: {:?}, confidence: {:.3}, disagreement: {:.3}",
             decision.action, decision.confidence, decision.disagreement_rate
@@ -292,6 +362,46 @@ impl EnsembleCoordinator {
 
         Ok(())
     }
+
+    /// Calculate quorum ratio (fraction of models agreeing on majority direction)
+    fn calculate_quorum_ratio(decision: &EnsembleDecision) -> f64 {
+        if decision.model_votes.is_empty() {
+            return 0.0;
+        }
+        let majority_action = &decision.action;
+        let agreeing = decision
+            .model_votes
+            .values()
+            .filter(|v| TradingAction::from_signal(v.signal, 0.3) == *majority_action)
+            .count();
+        agreeing as f64 / decision.model_votes.len() as f64
+    }
+
+    /// Determine current trading session (Eastern Time, simplified UTC-5)
+    fn current_trading_session() -> TradingSession {
+        let now = Utc::now();
+        let et_hour = (now.hour() + 24 - 5) % 24;
+        let et_minute = now.minute();
+        let et_time = et_hour * 60 + et_minute;
+
+        match et_time {
+            t if t < 240 => TradingSession::AfterHours, // 00:00-04:00 ET
+            t if t < 570 => TradingSession::PreMarket,  // 04:00-09:30 ET
+            t if t < 960 => TradingSession::Regular,     // 09:30-16:00 ET
+            _ => TradingSession::AfterHours,             // 16:00-24:00 ET
+        }
+    }
+
+    /// Extract regime volatility from feature vector
+    /// Regime features are at indices 48-50 in the 51-dim standard feature vector
+    fn extract_regime_volatility(features: &Features) -> f64 {
+        features
+            .values
+            .get(48)
+            .copied()
+            .unwrap_or(0.01)
+            .abs()
+    }
 }
 
 impl Default for EnsembleCoordinator {
diff --git a/ml/src/ensemble/gate_optimizer.rs b/ml/src/ensemble/gate_optimizer.rs
new file mode 100644
index 000000000..73fda9c66
--- /dev/null
+++ b/ml/src/ensemble/gate_optimizer.rs
@@ -0,0 +1,454 @@
+//! Gate threshold optimizer for conviction gates
+//!
+//! Adjusts conviction gate thresholds based on win-rate per confidence bucket.
+//! Same safety rails pattern as the weight optimizer: bounded adjustments,
+//! cooldown, freeze/unfreeze.
+
+use super::conviction_gates::ConvictionGateConfig;
+use std::time::{Duration, Instant};
+use tracing::{info, warn};
+
+/// Maximum threshold change per optimization cycle
+const MAX_THRESHOLD_STEP: f64 = 0.03;
+/// Minimum allowed threshold value
+const MIN_THRESHOLD: f64 = 0.30;
+/// Maximum allowed threshold value
+const MAX_THRESHOLD: f64 = 0.90;
+/// Minimum trades per bucket before adjustment
+const MIN_BUCKET_TRADES: u64 = 50;
+
+/// Win-rate metrics per confidence bucket
+#[derive(Debug, Clone)]
+pub struct GateBucketMetrics {
+    /// Confidence bucket lower bound (e.g. 0.60)
+    pub confidence_lower: f64,
+    /// Confidence bucket upper bound (e.g. 0.70)
+    pub confidence_upper: f64,
+    /// Win rate in this bucket (0.0-1.0)
+    pub win_rate: f64,
+    /// Number of trades in this bucket
+    pub trade_count: u64,
+    /// Average P&L per trade in this bucket
+    pub avg_pnl: f64,
+}
+
+/// Gate optimizer configuration
+#[derive(Debug, Clone)]
+pub struct GateOptimizerConfig {
+    pub max_step: f64,
+    pub min_threshold: f64,
+    pub max_threshold: f64,
+    pub min_bucket_trades: u64,
+    pub cooldown: Duration,
+    /// Target win rate — thresholds tighten if below, loosen if above
+    pub target_win_rate: f64,
+}
+
+impl Default for GateOptimizerConfig {
+    fn default() -> Self {
+        Self {
+            max_step: MAX_THRESHOLD_STEP,
+            min_threshold: MIN_THRESHOLD,
+            max_threshold: MAX_THRESHOLD,
+            min_bucket_trades: MIN_BUCKET_TRADES,
+            cooldown: Duration::from_secs(24 * 3600),
+            target_win_rate: 0.55,
+        }
+    }
+}
+
+/// Proposed threshold change
+#[derive(Debug, Clone)]
+pub struct ThresholdAdjustment {
+    pub field_name: String,
+    pub old_value: f64,
+    pub new_value: f64,
+    pub reason: String,
+}
+
+/// Result of a gate optimization cycle
+#[derive(Debug, Clone)]
+pub enum GateOptimizationResult {
+    /// Thresholds adjusted
+    Adjusted(Vec<ThresholdAdjustment>),
+    /// Not enough data
+    InsufficientData { total_trades: u64, required: u64 },
+    /// Still in cooldown
+    Cooldown { remaining: Duration },
+    /// Frozen by kill switch
+    KillSwitchActive,
+}
+
+/// Gate threshold optimizer
+#[derive(Debug)]
+pub struct GateOptimizer {
+    config: GateOptimizerConfig,
+    last_adjustment: Option<Instant>,
+    frozen: bool,
+}
+
+impl GateOptimizer {
+    pub fn new(config: GateOptimizerConfig) -> Self {
+        Self {
+            config,
+            last_adjustment: None,
+            frozen: false,
+        }
+    }
+
+    /// Freeze all adjustments (kill switch)
+    pub fn freeze(&mut self) {
+        self.frozen = true;
+        warn!("Gate optimizer frozen by kill switch");
+    }
+
+    /// Unfreeze adjustments
+    pub fn unfreeze(&mut self) {
+        self.frozen = false;
+        info!("Gate optimizer unfrozen");
+    }
+
+    pub fn is_frozen(&self) -> bool {
+        self.frozen
+    }
+
+    /// Run one optimization cycle
+    ///
+    /// Analyzes win-rate per confidence bucket and adjusts min_confidence threshold.
+    /// If win rate near the current threshold is below target, threshold increases
+    /// (more selective). If well above target, threshold decreases (more permissive).
+    pub fn optimize(
+        &mut self,
+        gate_config: &ConvictionGateConfig,
+        buckets: &[GateBucketMetrics],
+    ) -> GateOptimizationResult {
+        if self.frozen {
+            return GateOptimizationResult::KillSwitchActive;
+        }
+
+        // Check cooldown
+        if let Some(last) = self.last_adjustment {
+            let elapsed = last.elapsed();
+            if elapsed < self.config.cooldown {
+                return GateOptimizationResult::Cooldown {
+                    remaining: self.config.cooldown - elapsed,
+                };
+            }
+        }
+
+        // Check minimum data
+        let total_trades: u64 = buckets.iter().map(|b| b.trade_count).sum();
+        let min_required = self.config.min_bucket_trades * 3; // At least 3 buckets worth
+        if total_trades < min_required {
+            return GateOptimizationResult::InsufficientData {
+                total_trades,
+                required: min_required,
+            };
+        }
+
+        let mut adjustments = Vec::new();
+
+        // Analyze min_confidence threshold
+        // Find the bucket containing the current threshold
+        let threshold_bucket = buckets.iter().find(|b| {
+            b.confidence_lower <= gate_config.min_confidence
+                && gate_config.min_confidence < b.confidence_upper
+                && b.trade_count >= self.config.min_bucket_trades
+        });
+
+        if let Some(bucket) = threshold_bucket {
+            let win_rate_delta = bucket.win_rate - self.config.target_win_rate;
+
+            // If win rate is too low near threshold → tighten (increase threshold)
+            // If win rate is high → loosen (decrease threshold)
+            let direction = if win_rate_delta < -0.05 {
+                // Win rate below target by > 5pp → tighten
+                1.0
+            } else if win_rate_delta > 0.10 {
+                // Win rate above target by > 10pp → loosen
+                -1.0
+            } else {
+                0.0 // In acceptable range
+            };
+
+            if direction != 0.0 {
+                let step = (win_rate_delta.abs() * 0.1)
+                    .min(self.config.max_step)
+                    .max(0.005);
+                let new_confidence = (gate_config.min_confidence + direction * step)
+                    .clamp(self.config.min_threshold, self.config.max_threshold);
+
+                if (new_confidence - gate_config.min_confidence).abs() > 1e-6 {
+                    adjustments.push(ThresholdAdjustment {
+                        field_name: "min_confidence".into(),
+                        old_value: gate_config.min_confidence,
+                        new_value: new_confidence,
+                        reason: format!(
+                            "bucket win_rate={:.3}, target={:.3}, delta={:.3}",
+                            bucket.win_rate, self.config.target_win_rate, win_rate_delta
+                        ),
+                    });
+                }
+            }
+        }
+
+        // Analyze max_disagreement threshold
+        // If overall win rate on low-disagreement trades is high, can loosen
+        let low_disagree_buckets: Vec<&GateBucketMetrics> = buckets
+            .iter()
+            .filter(|b| b.trade_count >= self.config.min_bucket_trades)
+            .collect();
+
+        if !low_disagree_buckets.is_empty() {
+            let weighted_win_rate: f64 = low_disagree_buckets
+                .iter()
+                .map(|b| b.win_rate * b.trade_count as f64)
+                .sum::<f64>()
+                / low_disagree_buckets
+                    .iter()
+                    .map(|b| b.trade_count as f64)
+                    .sum::<f64>();
+
+            if weighted_win_rate < self.config.target_win_rate - 0.05 {
+                // Poor overall performance → tighten disagreement (lower max)
+                let new_max = (gate_config.max_disagreement - 0.01)
+                    .clamp(0.10, 0.60);
+
+                if (new_max - gate_config.max_disagreement).abs() > 1e-6 {
+                    adjustments.push(ThresholdAdjustment {
+                        field_name: "max_disagreement".into(),
+                        old_value: gate_config.max_disagreement,
+                        new_value: new_max,
+                        reason: format!(
+                            "weighted win_rate={:.3} below target {:.3}",
+                            weighted_win_rate, self.config.target_win_rate
+                        ),
+                    });
+                }
+            }
+        }
+
+        if !adjustments.is_empty() {
+            self.last_adjustment = Some(Instant::now());
+            info!(
+                "Gate optimizer adjusted {} thresholds",
+                adjustments.len()
+            );
+        }
+
+        GateOptimizationResult::Adjusted(adjustments)
+    }
+
+    /// Apply adjustments to a ConvictionGateConfig (returns modified copy)
+    pub fn apply(config: &ConvictionGateConfig, adjustments: &[ThresholdAdjustment]) -> ConvictionGateConfig {
+        let mut new_config = config.clone();
+        for adj in adjustments {
+            match adj.field_name.as_str() {
+                "min_confidence" => new_config.min_confidence = adj.new_value,
+                "max_disagreement" => new_config.max_disagreement = adj.new_value,
+                "min_quorum" => new_config.min_quorum = adj.new_value,
+                _ => {}
+            }
+        }
+        new_config
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn make_buckets(data: &[(f64, f64, f64, u64)]) -> Vec<GateBucketMetrics> {
+        data.iter()
+            .map(|(lower, upper, win_rate, trades)| GateBucketMetrics {
+                confidence_lower: *lower,
+                confidence_upper: *upper,
+                win_rate: *win_rate,
+                trade_count: *trades,
+                avg_pnl: 0.0,
+            })
+            .collect()
+    }
+
+    #[test]
+    fn test_insufficient_data() {
+        let mut opt = GateOptimizer::new(GateOptimizerConfig::default());
+        let config = ConvictionGateConfig::default();
+        let buckets = make_buckets(&[(0.50, 0.60, 0.55, 10), (0.60, 0.70, 0.60, 10)]);
+        let result = opt.optimize(&config, &buckets);
+        assert!(matches!(
+            result,
+            GateOptimizationResult::InsufficientData { .. }
+        ));
+    }
+
+    #[test]
+    fn test_cooldown_enforced() {
+        let mut opt = GateOptimizer::new(GateOptimizerConfig::default());
+        let config = ConvictionGateConfig::default();
+        let buckets = make_buckets(&[
+            (0.50, 0.60, 0.45, 100),
+            (0.60, 0.70, 0.40, 100),
+            (0.70, 0.80, 0.55, 100),
+        ]);
+
+        let _ = opt.optimize(&config, &buckets);
+        let result2 = opt.optimize(&config, &buckets);
+        assert!(matches!(result2, GateOptimizationResult::Cooldown { .. }));
+    }
+
+    #[test]
+    fn test_kill_switch() {
+        let mut opt = GateOptimizer::new(GateOptimizerConfig::default());
+        opt.freeze();
+        let config = ConvictionGateConfig::default();
+        let buckets = make_buckets(&[(0.50, 0.60, 0.45, 200)]);
+        let result = opt.optimize(&config, &buckets);
+        assert!(matches!(result, GateOptimizationResult::KillSwitchActive));
+    }
+
+    #[test]
+    fn test_tightens_on_low_win_rate() {
+        let mut config_opt = GateOptimizerConfig::default();
+        config_opt.cooldown = Duration::ZERO;
+        let mut opt = GateOptimizer::new(config_opt);
+        let config = ConvictionGateConfig::default(); // min_confidence = 0.60
+
+        // Win rate at threshold bucket is poor (0.40 < 0.55 target)
+        let buckets = make_buckets(&[
+            (0.50, 0.60, 0.40, 100),
+            (0.60, 0.70, 0.40, 100), // Threshold bucket
+            (0.70, 0.80, 0.60, 100),
+        ]);
+
+        let result = opt.optimize(&config, &buckets);
+        if let GateOptimizationResult::Adjusted(adjustments) = result {
+            let confidence_adj = adjustments
+                .iter()
+                .find(|a| a.field_name == "min_confidence");
+            assert!(
+                confidence_adj.is_some(),
+                "Expected min_confidence adjustment"
+            );
+            if let Some(adj) = confidence_adj {
+                assert!(
+                    adj.new_value > adj.old_value,
+                    "Expected threshold to increase (tighten) on low win rate"
+                );
+            }
+        }
+    }
+
+    #[test]
+    fn test_loosens_on_high_win_rate() {
+        let mut config_opt = GateOptimizerConfig::default();
+        config_opt.cooldown = Duration::ZERO;
+        let mut opt = GateOptimizer::new(config_opt);
+        let config = ConvictionGateConfig::default(); // min_confidence = 0.60
+
+        // Win rate at threshold bucket is very good (0.75 >> 0.55 target)
+        let buckets = make_buckets(&[
+            (0.50, 0.60, 0.70, 100),
+            (0.60, 0.70, 0.75, 100), // Threshold bucket
+            (0.70, 0.80, 0.80, 100),
+        ]);
+
+        let result = opt.optimize(&config, &buckets);
+        if let GateOptimizationResult::Adjusted(adjustments) = result {
+            let confidence_adj = adjustments
+                .iter()
+                .find(|a| a.field_name == "min_confidence");
+            if let Some(adj) = confidence_adj {
+                assert!(
+                    adj.new_value < adj.old_value,
+                    "Expected threshold to decrease (loosen) on high win rate"
+                );
+            }
+        }
+    }
+
+    #[test]
+    fn test_bounds_enforced() {
+        let mut config_opt = GateOptimizerConfig::default();
+        config_opt.cooldown = Duration::ZERO;
+        let mut opt = GateOptimizer::new(config_opt);
+
+        // Config with threshold already near maximum
+        let mut config = ConvictionGateConfig::default();
+        config.min_confidence = 0.89;
+
+        // Very low win rate to force tightening
+        let buckets = make_buckets(&[
+            (0.80, 0.90, 0.30, 100),
+            (0.89, 0.95, 0.30, 100), // Threshold bucket
+            (0.70, 0.80, 0.40, 100),
+        ]);
+
+        let result = opt.optimize(&config, &buckets);
+        if let GateOptimizationResult::Adjusted(adjustments) = result {
+            for adj in &adjustments {
+                assert!(
+                    adj.new_value >= MIN_THRESHOLD,
+                    "Below minimum: {}",
+                    adj.new_value
+                );
+                assert!(
+                    adj.new_value <= MAX_THRESHOLD,
+                    "Above maximum: {}",
+                    adj.new_value
+                );
+            }
+        }
+    }
+
+    #[test]
+    fn test_no_change_in_acceptable_range() {
+        let mut config_opt = GateOptimizerConfig::default();
+        config_opt.cooldown = Duration::ZERO;
+        let mut opt = GateOptimizer::new(config_opt);
+        let config = ConvictionGateConfig::default();
+
+        // Win rate is in acceptable range (target ± tolerance)
+        let buckets = make_buckets(&[
+            (0.50, 0.60, 0.56, 100),
+            (0.60, 0.70, 0.58, 100), // Just above target, within tolerance
+            (0.70, 0.80, 0.60, 100),
+        ]);
+
+        let result = opt.optimize(&config, &buckets);
+        if let GateOptimizationResult::Adjusted(adjustments) = result {
+            let confidence_adj = adjustments
+                .iter()
+                .find(|a| a.field_name == "min_confidence");
+            assert!(
+                confidence_adj.is_none(),
+                "Expected no min_confidence adjustment when in acceptable range"
+            );
+        }
+    }
+
+    #[test]
+    fn test_apply_adjustments() {
+        let config = ConvictionGateConfig::default();
+        let adjustments = vec![
+            ThresholdAdjustment {
+                field_name: "min_confidence".into(),
+                old_value: 0.60,
+                new_value: 0.65,
+                reason: "test".into(),
+            },
+            ThresholdAdjustment {
+                field_name: "max_disagreement".into(),
+                old_value: 0.40,
+                new_value: 0.35,
+                reason: "test".into(),
+            },
+        ];
+
+        let new_config = GateOptimizer::apply(&config, &adjustments);
+        assert!((new_config.min_confidence - 0.65).abs() < 1e-10);
+        assert!((new_config.max_disagreement - 0.35).abs() < 1e-10);
+        // Unchanged fields preserved
+        assert!((new_config.min_quorum - 0.60).abs() < 1e-10);
+    }
+}
diff --git a/ml/src/ensemble/mod.rs b/ml/src/ensemble/mod.rs
index 24ab0c40e..29f7ed862 100644
--- a/ml/src/ensemble/mod.rs
+++ b/ml/src/ensemble/mod.rs
@@ -21,6 +21,9 @@ pub mod inference_adapter;
 pub mod inference_ensemble;
 pub mod signal;
 pub mod adapters;
+pub mod conviction_gates;
+pub mod weight_optimizer;
+pub mod gate_optimizer;
 
 // Re-export key types that are used across ensemble modules
 pub use ab_testing::{
@@ -48,6 +51,18 @@ pub use metrics::{
 };
 pub use training_integration::EnsembleTrainingIntegration;
 pub use inference_adapter::{EnsemblePrediction, FeatureVector, ModelInferenceAdapter, PredictionMeta};
+pub use conviction_gates::{
+    ConvictionGateConfig, ConvictionGateEvaluator, ConvictionGateOutcome, GateEvaluation,
+    GateInput, GatePassResult, GateRejection, TradingSession,
+};
+pub use weight_optimizer::{
+    ModelRollingMetrics, OptimizationResult, WeightAdjustment, WeightOptimizer,
+    WeightOptimizerConfig,
+};
+pub use gate_optimizer::{
+    GateBucketMetrics, GateOptimizationResult, GateOptimizer, GateOptimizerConfig,
+    ThresholdAdjustment,
+};
 
 /// Errors that can occur in ensemble operations
 #[derive(Error, Debug)]
diff --git a/ml/src/ensemble/weight_optimizer.rs b/ml/src/ensemble/weight_optimizer.rs
new file mode 100644
index 000000000..abc8b3b5c
--- /dev/null
+++ b/ml/src/ensemble/weight_optimizer.rs
@@ -0,0 +1,426 @@
+//! Autonomous weight optimizer for ensemble models
+//!
+//! Adjusts model weights based on rolling Sharpe ratios using
+//! exponentially-weighted moving average. Bounded by safety rails.
+
+use std::collections::HashMap;
+use std::time::{Duration, Instant};
+use tracing::{info, warn};
+
+/// Minimum weight for any model (prevents zeroing out)
+const MIN_WEIGHT: f64 = 0.05;
+/// Maximum weight for any model (prevents dominance)
+const MAX_WEIGHT: f64 = 0.40;
+/// Maximum weight change per optimization cycle
+const MAX_STEP: f64 = 0.03;
+/// Minimum number of trades before first adjustment
+const MIN_OBSERVATIONS: u64 = 100;
+/// Cooldown between adjustments
+const COOLDOWN: Duration = Duration::from_secs(24 * 3600); // 24 hours
+/// Grace period for newly deployed models
+const GRACE_PERIOD: Duration = Duration::from_secs(7 * 24 * 3600); // 7 days
+
+/// Per-model rolling performance metrics (fed from QuestDB queries)
+#[derive(Debug, Clone)]
+pub struct ModelRollingMetrics {
+    pub model_id: String,
+    pub sharpe_30d: f64,
+    pub win_rate_30d: f64,
+    pub prediction_accuracy: f64,
+    pub trade_count: u64,
+    pub deployed_at: Instant,
+}
+
+/// Weight optimizer configuration
+#[derive(Debug, Clone)]
+pub struct WeightOptimizerConfig {
+    pub ema_alpha: f64,
+    pub min_weight: f64,
+    pub max_weight: f64,
+    pub max_step: f64,
+    pub min_observations: u64,
+    pub cooldown: Duration,
+    pub grace_period: Duration,
+}
+
+impl Default for WeightOptimizerConfig {
+    fn default() -> Self {
+        Self {
+            ema_alpha: 0.1,
+            min_weight: MIN_WEIGHT,
+            max_weight: MAX_WEIGHT,
+            max_step: MAX_STEP,
+            min_observations: MIN_OBSERVATIONS,
+            cooldown: COOLDOWN,
+            grace_period: GRACE_PERIOD,
+        }
+    }
+}
+
+/// Proposed weight changes from the optimizer
+#[derive(Debug, Clone)]
+pub struct WeightAdjustment {
+    pub model_id: String,
+    pub old_weight: f64,
+    pub new_weight: f64,
+    pub reason: String,
+}
+
+/// Result of an optimization cycle
+#[derive(Debug, Clone)]
+pub enum OptimizationResult {
+    /// Weights adjusted successfully
+    Adjusted(Vec<WeightAdjustment>),
+    /// Not enough observations yet
+    InsufficientData { total_trades: u64, required: u64 },
+    /// Still in cooldown from last adjustment
+    Cooldown { remaining: Duration },
+    /// Kill switch is active, no adjustments allowed
+    KillSwitchActive,
+}
+
+/// Autonomous weight optimizer
+#[derive(Debug)]
+pub struct WeightOptimizer {
+    config: WeightOptimizerConfig,
+    last_adjustment: Option<Instant>,
+    ema_sharpe: HashMap<String, f64>,
+    frozen: bool,
+}
+
+impl WeightOptimizer {
+    pub fn new(config: WeightOptimizerConfig) -> Self {
+        Self {
+            config,
+            last_adjustment: None,
+            ema_sharpe: HashMap::new(),
+            frozen: false,
+        }
+    }
+
+    /// Freeze all adjustments (kill switch)
+    pub fn freeze(&mut self) {
+        self.frozen = true;
+        warn!("Weight optimizer frozen by kill switch");
+    }
+
+    /// Unfreeze adjustments (human acknowledgment)
+    pub fn unfreeze(&mut self) {
+        self.frozen = false;
+        info!("Weight optimizer unfrozen");
+    }
+
+    pub fn is_frozen(&self) -> bool {
+        self.frozen
+    }
+
+    /// Run one optimization cycle
+    ///
+    /// Takes current weights and rolling metrics, returns proposed adjustments.
+    pub fn optimize(
+        &mut self,
+        current_weights: &HashMap<String, f64>,
+        metrics: &[ModelRollingMetrics],
+    ) -> OptimizationResult {
+        if self.frozen {
+            return OptimizationResult::KillSwitchActive;
+        }
+
+        // Check cooldown
+        if let Some(last) = self.last_adjustment {
+            let elapsed = last.elapsed();
+            if elapsed < self.config.cooldown {
+                return OptimizationResult::Cooldown {
+                    remaining: self.config.cooldown - elapsed,
+                };
+            }
+        }
+
+        // Check minimum observations
+        let total_trades: u64 = metrics.iter().map(|m| m.trade_count).sum();
+        if total_trades < self.config.min_observations {
+            return OptimizationResult::InsufficientData {
+                total_trades,
+                required: self.config.min_observations,
+            };
+        }
+
+        // Filter out models in grace period
+        let eligible: Vec<&ModelRollingMetrics> = metrics
+            .iter()
+            .filter(|m| m.deployed_at.elapsed() >= self.config.grace_period)
+            .collect();
+
+        if eligible.is_empty() {
+            return OptimizationResult::InsufficientData {
+                total_trades,
+                required: self.config.min_observations,
+            };
+        }
+
+        // Update EMA of Sharpe ratios
+        for m in &eligible {
+            let prev = self
+                .ema_sharpe
+                .get(&m.model_id)
+                .copied()
+                .unwrap_or(m.sharpe_30d);
+            let new_ema =
+                self.config.ema_alpha * m.sharpe_30d + (1.0 - self.config.ema_alpha) * prev;
+            self.ema_sharpe.insert(m.model_id.clone(), new_ema);
+        }
+
+        // Calculate raw weights from EMA Sharpe (shift to positive range)
+        let min_ema = self
+            .ema_sharpe
+            .values()
+            .cloned()
+            .fold(f64::INFINITY, f64::min);
+        let shift = if min_ema < 0.0 {
+            min_ema.abs() + 0.1
+        } else {
+            0.0
+        };
+
+        let raw_weights: HashMap<String, f64> = self
+            .ema_sharpe
+            .iter()
+            .map(|(id, &ema)| (id.clone(), (ema + shift).max(0.01)))
+            .collect();
+
+        let total_raw: f64 = raw_weights.values().sum();
+        if total_raw <= 0.0 {
+            return OptimizationResult::InsufficientData {
+                total_trades,
+                required: self.config.min_observations,
+            };
+        }
+
+        // Normalize and clamp
+        let mut target_weights: HashMap<String, f64> = raw_weights
+            .iter()
+            .map(|(id, &raw)| {
+                let normalized = raw / total_raw;
+                let clamped = normalized.clamp(self.config.min_weight, self.config.max_weight);
+                (id.clone(), clamped)
+            })
+            .collect();
+
+        // Re-normalize after clamping
+        let total_clamped: f64 = target_weights.values().sum();
+        if total_clamped > 0.0 {
+            for w in target_weights.values_mut() {
+                *w /= total_clamped;
+            }
+        }
+
+        // Apply max step limit
+        let mut adjustments = Vec::new();
+        for (model_id, &target) in &target_weights {
+            let current = current_weights.get(model_id).copied().unwrap_or(target);
+            let delta = (target - current).clamp(-self.config.max_step, self.config.max_step);
+            let new_weight =
+                (current + delta).clamp(self.config.min_weight, self.config.max_weight);
+
+            if (new_weight - current).abs() > 1e-6 {
+                adjustments.push(WeightAdjustment {
+                    model_id: model_id.clone(),
+                    old_weight: current,
+                    new_weight,
+                    reason: format!(
+                        "EMA Sharpe: {:.3}, target: {:.3}",
+                        self.ema_sharpe.get(model_id).unwrap_or(&0.0),
+                        target
+                    ),
+                });
+            }
+        }
+
+        if !adjustments.is_empty() {
+            self.last_adjustment = Some(Instant::now());
+            info!("Weight optimizer adjusted {} models", adjustments.len());
+        }
+
+        OptimizationResult::Adjusted(adjustments)
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn make_metrics(models: &[(&str, f64, u64)]) -> Vec<ModelRollingMetrics> {
+        models
+            .iter()
+            .map(|(id, sharpe, trades)| ModelRollingMetrics {
+                model_id: id.to_string(),
+                sharpe_30d: *sharpe,
+                win_rate_30d: 0.55,
+                prediction_accuracy: 0.60,
+                trade_count: *trades,
+                deployed_at: Instant::now() - Duration::from_secs(30 * 24 * 3600),
+            })
+            .collect()
+    }
+
+    fn equal_weights(models: &[&str]) -> HashMap<String, f64> {
+        let w = 1.0 / models.len() as f64;
+        models.iter().map(|id| (id.to_string(), w)).collect()
+    }
+
+    #[test]
+    fn test_insufficient_data() {
+        let mut opt = WeightOptimizer::new(WeightOptimizerConfig::default());
+        let weights = equal_weights(&["dqn", "ppo"]);
+        let metrics = make_metrics(&[("dqn", 1.0, 10), ("ppo", 0.5, 10)]);
+        let result = opt.optimize(&weights, &metrics);
+        assert!(matches!(
+            result,
+            OptimizationResult::InsufficientData { .. }
+        ));
+    }
+
+    #[test]
+    fn test_cooldown_enforced() {
+        let mut opt = WeightOptimizer::new(WeightOptimizerConfig::default());
+        let weights = equal_weights(&["dqn", "ppo"]);
+        let metrics = make_metrics(&[("dqn", 1.5, 200), ("ppo", 0.5, 200)]);
+
+        // First optimization succeeds
+        let result = opt.optimize(&weights, &metrics);
+        assert!(matches!(result, OptimizationResult::Adjusted(_)));
+
+        // Second immediately after should be cooldown
+        let result2 = opt.optimize(&weights, &metrics);
+        assert!(matches!(result2, OptimizationResult::Cooldown { .. }));
+    }
+
+    #[test]
+    fn test_kill_switch_freezes() {
+        let mut opt = WeightOptimizer::new(WeightOptimizerConfig::default());
+        opt.freeze();
+        let weights = equal_weights(&["dqn"]);
+        let metrics = make_metrics(&[("dqn", 1.0, 200)]);
+        let result = opt.optimize(&weights, &metrics);
+        assert!(matches!(result, OptimizationResult::KillSwitchActive));
+    }
+
+    #[test]
+    fn test_unfreeze_allows_optimization() {
+        let mut opt = WeightOptimizer::new(WeightOptimizerConfig::default());
+        opt.freeze();
+        assert!(opt.is_frozen());
+        opt.unfreeze();
+        assert!(!opt.is_frozen());
+    }
+
+    #[test]
+    fn test_weights_bounded() {
+        let mut config = WeightOptimizerConfig::default();
+        config.cooldown = Duration::ZERO;
+        let mut opt = WeightOptimizer::new(config);
+        let weights = equal_weights(&["dqn", "ppo", "tft"]);
+        let metrics = make_metrics(&[("dqn", 5.0, 200), ("ppo", -2.0, 200), ("tft", 0.5, 200)]);
+
+        let result = opt.optimize(&weights, &metrics);
+        if let OptimizationResult::Adjusted(adjustments) = result {
+            for adj in &adjustments {
+                assert!(
+                    adj.new_weight >= MIN_WEIGHT,
+                    "Weight below minimum: {}",
+                    adj.new_weight
+                );
+                assert!(
+                    adj.new_weight <= MAX_WEIGHT,
+                    "Weight above maximum: {}",
+                    adj.new_weight
+                );
+            }
+        }
+    }
+
+    #[test]
+    fn test_max_step_enforced() {
+        let mut config = WeightOptimizerConfig::default();
+        config.cooldown = Duration::ZERO;
+        let mut opt = WeightOptimizer::new(config);
+
+        // Both weights within [MIN_WEIGHT, MAX_WEIGHT] range
+        let mut weights = HashMap::new();
+        weights.insert("dqn".into(), 0.30);
+        weights.insert("ppo".into(), 0.30);
+        weights.insert("tft".into(), 0.40);
+
+        let metrics = make_metrics(&[("dqn", 5.0, 200), ("ppo", -1.0, 200), ("tft", 0.5, 200)]);
+        let result = opt.optimize(&weights, &metrics);
+
+        if let OptimizationResult::Adjusted(adjustments) = result {
+            for adj in &adjustments {
+                let delta = (adj.new_weight - adj.old_weight).abs();
+                // Step should be bounded by MAX_STEP, unless weight bounds force a correction
+                assert!(
+                    delta <= MAX_STEP + 1e-6 || adj.new_weight == MIN_WEIGHT || adj.new_weight == MAX_WEIGHT,
+                    "Step too large: {} for {} (old={}, new={})",
+                    delta,
+                    adj.model_id,
+                    adj.old_weight,
+                    adj.new_weight
+                );
+            }
+        }
+    }
+
+    #[test]
+    fn test_grace_period_excludes_new_models() {
+        let mut config = WeightOptimizerConfig::default();
+        config.cooldown = Duration::ZERO;
+        let mut opt = WeightOptimizer::new(config);
+
+        let weights = equal_weights(&["dqn", "new_model"]);
+        let mut metrics = make_metrics(&[("dqn", 1.0, 200)]);
+        // New model deployed 1 day ago (within 7-day grace period)
+        metrics.push(ModelRollingMetrics {
+            model_id: "new_model".into(),
+            sharpe_30d: 2.0,
+            win_rate_30d: 0.70,
+            prediction_accuracy: 0.80,
+            trade_count: 50,
+            deployed_at: Instant::now() - Duration::from_secs(24 * 3600),
+        });
+
+        let result = opt.optimize(&weights, &metrics);
+        if let OptimizationResult::Adjusted(adjustments) = result {
+            // new_model should not be in adjustments (grace period)
+            assert!(adjustments.iter().all(|a| a.model_id != "new_model"));
+        }
+    }
+
+    #[test]
+    fn test_weights_sum_approximately_one() {
+        let mut config = WeightOptimizerConfig::default();
+        config.cooldown = Duration::ZERO;
+        let mut opt = WeightOptimizer::new(config);
+        let weights = equal_weights(&["a", "b", "c", "d"]);
+        let metrics = make_metrics(&[
+            ("a", 1.0, 200),
+            ("b", 1.5, 200),
+            ("c", 0.5, 200),
+            ("d", 0.8, 200),
+        ]);
+
+        let result = opt.optimize(&weights, &metrics);
+        if let OptimizationResult::Adjusted(adjustments) = result {
+            let mut new_weights = weights.clone();
+            for adj in &adjustments {
+                new_weights.insert(adj.model_id.clone(), adj.new_weight);
+            }
+            let sum: f64 = new_weights.values().sum();
+            // Due to step limits, sum may not be exactly 1.0 but should be close
+            assert!(
+                (sum - 1.0).abs() < 0.2,
+                "Weights sum to {} (expected ~1.0)",
+                sum
+            );
+        }
+    }
+}
diff --git a/ml/src/lib.rs b/ml/src/lib.rs
index 08948b663..638a5191f 100644
--- a/ml/src/lib.rs
+++ b/ml/src/lib.rs
@@ -834,6 +834,7 @@ pub mod random_model;
 pub mod model_registry;
 pub mod data_pipeline;
 pub mod asset_selection;
+pub mod registry; // Operational maturity: model lifecycle (Candidate → Staging → Production → Archived)
 
 // ========== MISSING TYPES STUBS ==========
 
diff --git a/ml/src/registry/mod.rs b/ml/src/registry/mod.rs
new file mode 100644
index 000000000..8ff80b6f3
--- /dev/null
+++ b/ml/src/registry/mod.rs
@@ -0,0 +1,524 @@
+//! Model Registry for lifecycle management
+//!
+//! Tracks training runs, model versions, and promotions through
+//! Candidate → Staging → Production → Archived lifecycle.
+
+use serde::{Deserialize, Serialize};
+use std::collections::HashMap;
+use std::time::SystemTime;
+
+/// Model lifecycle stage
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash, Serialize, Deserialize)]
+pub enum ModelStage {
+    /// Initial state after training
+    Candidate,
+    /// Passed validation, running canary
+    Staging,
+    /// Active in production ensemble
+    Production,
+    /// Replaced by newer version
+    Archived,
+}
+
+impl std::fmt::Display for ModelStage {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        match self {
+            Self::Candidate => write!(f, "candidate"),
+            Self::Staging => write!(f, "staging"),
+            Self::Production => write!(f, "production"),
+            Self::Archived => write!(f, "archived"),
+        }
+    }
+}
+
+/// Record of a training run
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct TrainingRun {
+    /// Unique run identifier
+    pub run_id: String,
+    /// Experiment name (e.g. "dqn-v3-sharpe-opt")
+    pub experiment_name: String,
+    /// Model type (DQN, PPO, TFT, etc.)
+    pub model_type: String,
+    /// Hyperparameters as JSON
+    pub hyperparameters: serde_json::Value,
+    /// Git commit hash at time of training
+    pub git_commit: String,
+    /// Hash of training data for reproducibility
+    pub data_hash: String,
+    /// When training started
+    pub started_at: SystemTime,
+    /// When training finished (None if still running)
+    pub finished_at: Option<SystemTime>,
+}
+
+/// Metrics recorded for a model version
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct ModelMetrics {
+    /// Validation Sharpe ratio
+    pub sharpe_ratio: f64,
+    /// Validation accuracy
+    pub accuracy: f64,
+    /// Validation win rate
+    pub win_rate: f64,
+    /// Maximum drawdown on validation set
+    pub max_drawdown: f64,
+    /// Any additional metrics
+    pub extra: HashMap<String, f64>,
+}
+
+/// A versioned model in the registry
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct ModelVersion {
+    /// Unique version identifier
+    pub version_id: String,
+    /// Associated training run
+    pub run_id: String,
+    /// Model type
+    pub model_type: String,
+    /// Path to model artifact (safetensors file)
+    pub artifact_path: String,
+    /// Current lifecycle stage
+    pub stage: ModelStage,
+    /// Validation metrics
+    pub metrics: Option<ModelMetrics>,
+    /// When this version was registered
+    pub registered_at: SystemTime,
+    /// Who/what promoted this version
+    pub promoted_by: Option<String>,
+}
+
+/// Side-by-side run comparison
+#[derive(Debug, Clone)]
+pub struct RunComparison {
+    pub runs: Vec<(TrainingRun, Option<ModelMetrics>)>,
+}
+
+/// Model registry trait (async for database implementations)
+#[async_trait::async_trait]
+pub trait ModelRegistryTrait: Send + Sync {
+    /// Log a new training run
+    async fn log_run(&self, run: TrainingRun) -> Result<(), RegistryError>;
+
+    /// Log metrics for a model version
+    async fn log_metrics(
+        &self,
+        version_id: &str,
+        metrics: ModelMetrics,
+    ) -> Result<(), RegistryError>;
+
+    /// Register a model version
+    async fn register_version(&self, version: ModelVersion) -> Result<(), RegistryError>;
+
+    /// Promote a model to a new stage
+    async fn promote(
+        &self,
+        version_id: &str,
+        to_stage: ModelStage,
+        promoted_by: &str,
+    ) -> Result<(), RegistryError>;
+
+    /// Get the current production model for a given type
+    async fn get_production_model(
+        &self,
+        model_type: &str,
+    ) -> Result<Option<ModelVersion>, RegistryError>;
+
+    /// Revert to previous production version
+    async fn revert(&self, model_type: &str) -> Result<ModelVersion, RegistryError>;
+
+    /// Compare metrics across runs
+    async fn compare_runs(&self, run_ids: &[String]) -> Result<RunComparison, RegistryError>;
+}
+
+/// Registry errors
+#[derive(Debug, thiserror::Error)]
+pub enum RegistryError {
+    #[error("Version not found: {0}")]
+    VersionNotFound(String),
+    #[error("Run not found: {0}")]
+    RunNotFound(String),
+    #[error("Invalid stage transition: {from} → {to}")]
+    InvalidTransition { from: String, to: String },
+    #[error("No previous version to revert to for model type: {0}")]
+    NoPreviousVersion(String),
+}
+
+/// In-memory model registry for testing
+#[derive(Debug, Default)]
+pub struct InMemoryModelRegistry {
+    runs: tokio::sync::RwLock<HashMap<String, TrainingRun>>,
+    versions: tokio::sync::RwLock<HashMap<String, ModelVersion>>,
+    metrics: tokio::sync::RwLock<HashMap<String, ModelMetrics>>,
+}
+
+impl InMemoryModelRegistry {
+    pub fn new() -> Self {
+        Self::default()
+    }
+}
+
+#[async_trait::async_trait]
+impl ModelRegistryTrait for InMemoryModelRegistry {
+    async fn log_run(&self, run: TrainingRun) -> Result<(), RegistryError> {
+        self.runs.write().await.insert(run.run_id.clone(), run);
+        Ok(())
+    }
+
+    async fn log_metrics(
+        &self,
+        version_id: &str,
+        metrics: ModelMetrics,
+    ) -> Result<(), RegistryError> {
+        // Also update the version's metrics
+        let mut versions = self.versions.write().await;
+        if let Some(version) = versions.get_mut(version_id) {
+            version.metrics = Some(metrics.clone());
+        }
+        self.metrics
+            .write()
+            .await
+            .insert(version_id.to_string(), metrics);
+        Ok(())
+    }
+
+    async fn register_version(&self, version: ModelVersion) -> Result<(), RegistryError> {
+        self.versions
+            .write()
+            .await
+            .insert(version.version_id.clone(), version);
+        Ok(())
+    }
+
+    async fn promote(
+        &self,
+        version_id: &str,
+        to_stage: ModelStage,
+        promoted_by: &str,
+    ) -> Result<(), RegistryError> {
+        let mut versions = self.versions.write().await;
+        let version = versions
+            .get_mut(version_id)
+            .ok_or_else(|| RegistryError::VersionNotFound(version_id.to_string()))?;
+
+        // Validate transition
+        let valid = matches!(
+            (version.stage, to_stage),
+            (ModelStage::Candidate, ModelStage::Staging)
+                | (ModelStage::Staging, ModelStage::Production)
+                | (ModelStage::Production, ModelStage::Archived)
+                | (ModelStage::Staging, ModelStage::Archived)
+                | (ModelStage::Candidate, ModelStage::Archived)
+        );
+
+        if !valid {
+            return Err(RegistryError::InvalidTransition {
+                from: version.stage.to_string(),
+                to: to_stage.to_string(),
+            });
+        }
+
+        // If promoting to Production, archive the current production model of same type
+        if to_stage == ModelStage::Production {
+            let model_type = version.model_type.clone();
+            let current_prod: Vec<String> = versions
+                .iter()
+                .filter(|(id, v)| {
+                    v.model_type == model_type
+                        && v.stage == ModelStage::Production
+                        && *id != version_id
+                })
+                .map(|(id, _)| id.clone())
+                .collect();
+
+            // Must drop the version borrow before modifying others
+            let version = versions.get_mut(version_id).expect("just checked");
+            version.stage = to_stage;
+            version.promoted_by = Some(promoted_by.to_string());
+
+            for old_id in current_prod {
+                if let Some(old_version) = versions.get_mut(&old_id) {
+                    old_version.stage = ModelStage::Archived;
+                }
+            }
+        } else {
+            version.stage = to_stage;
+            version.promoted_by = Some(promoted_by.to_string());
+        }
+
+        Ok(())
+    }
+
+    async fn get_production_model(
+        &self,
+        model_type: &str,
+    ) -> Result<Option<ModelVersion>, RegistryError> {
+        let versions = self.versions.read().await;
+        let prod = versions
+            .values()
+            .find(|v| v.model_type == model_type && v.stage == ModelStage::Production)
+            .cloned();
+        Ok(prod)
+    }
+
+    async fn revert(&self, model_type: &str) -> Result<ModelVersion, RegistryError> {
+        let mut versions = self.versions.write().await;
+
+        // Find the most recently archived version of this type
+        let archived: Option<String> = versions
+            .iter()
+            .filter(|(_, v)| v.model_type == model_type && v.stage == ModelStage::Archived)
+            .max_by_key(|(_, v)| v.registered_at)
+            .map(|(id, _)| id.clone());
+
+        let archived_id =
+            archived.ok_or_else(|| RegistryError::NoPreviousVersion(model_type.to_string()))?;
+
+        // Archive current production
+        let current_prod: Vec<String> = versions
+            .iter()
+            .filter(|(_, v)| v.model_type == model_type && v.stage == ModelStage::Production)
+            .map(|(id, _)| id.clone())
+            .collect();
+
+        for id in current_prod {
+            if let Some(v) = versions.get_mut(&id) {
+                v.stage = ModelStage::Archived;
+            }
+        }
+
+        // Promote archived to production
+        let version = versions
+            .get_mut(&archived_id)
+            .ok_or_else(|| RegistryError::VersionNotFound(archived_id.clone()))?;
+        version.stage = ModelStage::Production;
+        version.promoted_by = Some("revert".to_string());
+
+        Ok(version.clone())
+    }
+
+    async fn compare_runs(&self, run_ids: &[String]) -> Result<RunComparison, RegistryError> {
+        let runs = self.runs.read().await;
+        let metrics = self.metrics.read().await;
+
+        let mut comparisons = Vec::new();
+        for id in run_ids {
+            let run = runs
+                .get(id)
+                .ok_or_else(|| RegistryError::RunNotFound(id.clone()))?
+                .clone();
+            let m = metrics.get(id).cloned();
+            comparisons.push((run, m));
+        }
+
+        Ok(RunComparison {
+            runs: comparisons,
+        })
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn make_run(id: &str, model_type: &str) -> TrainingRun {
+        TrainingRun {
+            run_id: id.to_string(),
+            experiment_name: format!("{}-exp", model_type),
+            model_type: model_type.to_string(),
+            hyperparameters: serde_json::json!({"lr": 0.001}),
+            git_commit: "abc123".into(),
+            data_hash: "sha256:deadbeef".into(),
+            started_at: SystemTime::now(),
+            finished_at: Some(SystemTime::now()),
+        }
+    }
+
+    fn make_version(id: &str, run_id: &str, model_type: &str) -> ModelVersion {
+        ModelVersion {
+            version_id: id.to_string(),
+            run_id: run_id.to_string(),
+            model_type: model_type.to_string(),
+            artifact_path: format!("models/{}/{}.safetensors", model_type, id),
+            stage: ModelStage::Candidate,
+            metrics: None,
+            registered_at: SystemTime::now(),
+            promoted_by: None,
+        }
+    }
+
+    #[test]
+    fn test_model_stage_display() {
+        assert_eq!(ModelStage::Candidate.to_string(), "candidate");
+        assert_eq!(ModelStage::Production.to_string(), "production");
+    }
+
+    #[tokio::test]
+    async fn test_log_and_register() {
+        let registry = InMemoryModelRegistry::new();
+        let run = make_run("run-1", "DQN");
+        registry.log_run(run).await.unwrap();
+
+        let version = make_version("v1", "run-1", "DQN");
+        registry.register_version(version).await.unwrap();
+
+        let prod = registry.get_production_model("DQN").await.unwrap();
+        assert!(prod.is_none()); // Not promoted yet
+    }
+
+    #[tokio::test]
+    async fn test_promote_lifecycle() {
+        let registry = InMemoryModelRegistry::new();
+        let run = make_run("run-1", "DQN");
+        registry.log_run(run).await.unwrap();
+
+        let version = make_version("v1", "run-1", "DQN");
+        registry.register_version(version).await.unwrap();
+
+        // Candidate → Staging
+        registry
+            .promote("v1", ModelStage::Staging, "ci")
+            .await
+            .unwrap();
+
+        // Staging → Production
+        registry
+            .promote("v1", ModelStage::Production, "ci")
+            .await
+            .unwrap();
+
+        let prod = registry.get_production_model("DQN").await.unwrap();
+        assert!(prod.is_some());
+        assert_eq!(prod.unwrap().version_id, "v1");
+    }
+
+    #[tokio::test]
+    async fn test_invalid_transition() {
+        let registry = InMemoryModelRegistry::new();
+        let version = make_version("v1", "run-1", "DQN");
+        registry.register_version(version).await.unwrap();
+
+        // Candidate → Production is invalid (must go through Staging)
+        let result = registry
+            .promote("v1", ModelStage::Production, "ci")
+            .await;
+        assert!(result.is_err());
+    }
+
+    #[tokio::test]
+    async fn test_promotion_archives_old() {
+        let registry = InMemoryModelRegistry::new();
+
+        // Register and promote v1 to production
+        registry
+            .register_version(make_version("v1", "run-1", "DQN"))
+            .await
+            .unwrap();
+        registry
+            .promote("v1", ModelStage::Staging, "ci")
+            .await
+            .unwrap();
+        registry
+            .promote("v1", ModelStage::Production, "ci")
+            .await
+            .unwrap();
+
+        // Register and promote v2 to production
+        registry
+            .register_version(make_version("v2", "run-2", "DQN"))
+            .await
+            .unwrap();
+        registry
+            .promote("v2", ModelStage::Staging, "ci")
+            .await
+            .unwrap();
+        registry
+            .promote("v2", ModelStage::Production, "ci")
+            .await
+            .unwrap();
+
+        // v1 should be archived, v2 should be production
+        let prod = registry.get_production_model("DQN").await.unwrap();
+        assert_eq!(prod.unwrap().version_id, "v2");
+    }
+
+    #[tokio::test]
+    async fn test_revert() {
+        let registry = InMemoryModelRegistry::new();
+
+        // v1 → production → archived (when v2 promoted)
+        registry
+            .register_version(make_version("v1", "run-1", "DQN"))
+            .await
+            .unwrap();
+        registry
+            .promote("v1", ModelStage::Staging, "ci")
+            .await
+            .unwrap();
+        registry
+            .promote("v1", ModelStage::Production, "ci")
+            .await
+            .unwrap();
+
+        registry
+            .register_version(make_version("v2", "run-2", "DQN"))
+            .await
+            .unwrap();
+        registry
+            .promote("v2", ModelStage::Staging, "ci")
+            .await
+            .unwrap();
+        registry
+            .promote("v2", ModelStage::Production, "ci")
+            .await
+            .unwrap();
+
+        // Revert should bring v1 back
+        let reverted = registry.revert("DQN").await.unwrap();
+        assert_eq!(reverted.version_id, "v1");
+        assert_eq!(reverted.stage, ModelStage::Production);
+    }
+
+    #[tokio::test]
+    async fn test_log_metrics() {
+        let registry = InMemoryModelRegistry::new();
+        registry
+            .register_version(make_version("v1", "run-1", "DQN"))
+            .await
+            .unwrap();
+
+        let metrics = ModelMetrics {
+            sharpe_ratio: 1.5,
+            accuracy: 0.62,
+            win_rate: 0.58,
+            max_drawdown: -0.05,
+            extra: HashMap::new(),
+        };
+        registry.log_metrics("v1", metrics).await.unwrap();
+
+        // Metrics should be attached to version
+        let versions = registry.versions.read().await;
+        let v = versions.get("v1").unwrap();
+        assert!(v.metrics.is_some());
+        assert!((v.metrics.as_ref().unwrap().sharpe_ratio - 1.5).abs() < 1e-10);
+    }
+
+    #[tokio::test]
+    async fn test_compare_runs() {
+        let registry = InMemoryModelRegistry::new();
+        registry
+            .log_run(make_run("run-1", "DQN"))
+            .await
+            .unwrap();
+        registry
+            .log_run(make_run("run-2", "DQN"))
+            .await
+            .unwrap();
+
+        let comparison = registry
+            .compare_runs(&["run-1".into(), "run-2".into()])
+            .await
+            .unwrap();
+        assert_eq!(comparison.runs.len(), 2);
+    }
+}
diff --git a/services/trading_service/src/attribution.rs b/services/trading_service/src/attribution.rs
new file mode 100644
index 000000000..d3972225d
--- /dev/null
+++ b/services/trading_service/src/attribution.rs
@@ -0,0 +1,262 @@
+//! P&L Attribution Calculator
+//!
+//! Decomposes realized trade P&L into per-model contributions.
+//! Each model gets credit proportional to:
+//! - Its ensemble weight at trade time
+//! - Whether its signal aligned with the realized direction
+
+use ml::ensemble::decision::EnsembleDecision;
+
+/// Per-model attribution for a single trade
+#[derive(Debug, Clone)]
+pub struct TradeAttribution {
+    /// Model identifier
+    pub model_id: String,
+    /// Model's ensemble weight at trade time
+    pub model_weight: f64,
+    /// Model's raw signal (-1.0 to 1.0)
+    pub model_signal: f64,
+    /// 1.0 if signal direction matched realized direction, -1.0 otherwise
+    pub signal_alignment: f64,
+    /// Attributed P&L contribution
+    pub pnl_contribution: f64,
+}
+
+/// Full attribution result for a closed trade
+#[derive(Debug, Clone)]
+pub struct AttributionResult {
+    /// Per-model attributions
+    pub attributions: Vec<TradeAttribution>,
+    /// Total realized P&L (should equal sum of contributions)
+    pub realized_pnl: f64,
+    /// Residual (rounding error, should be near zero)
+    pub residual: f64,
+}
+
+/// Calculate per-model P&L attribution from an ensemble decision and realized P&L.
+///
+/// For each model that voted:
+/// - `signal_alignment = 1.0` if sign(model_signal) == sign(realized_pnl), else `-1.0`
+/// - `pnl_contribution = model_weight × signal_alignment × |realized_pnl|`
+///
+/// Models with zero signal are treated as neutral (alignment = 0.0).
+pub fn attribute(decision: &EnsembleDecision, realized_pnl: f64) -> AttributionResult {
+    let votes = &decision.model_votes;
+
+    if votes.is_empty() || realized_pnl.abs() < 1e-12 {
+        return AttributionResult {
+            attributions: votes
+                .iter()
+                .map(|(id, v)| TradeAttribution {
+                    model_id: id.clone(),
+                    model_weight: v.weight,
+                    model_signal: v.signal,
+                    signal_alignment: 0.0,
+                    pnl_contribution: 0.0,
+                })
+                .collect(),
+            realized_pnl,
+            residual: realized_pnl,
+        };
+    }
+
+    let realized_direction = realized_pnl.signum();
+
+    let attributions: Vec<TradeAttribution> = votes
+        .iter()
+        .map(|(id, vote)| {
+            let alignment = compute_alignment(vote.signal, realized_direction);
+            let contribution = vote.weight * alignment * realized_pnl.abs();
+
+            TradeAttribution {
+                model_id: id.clone(),
+                model_weight: vote.weight,
+                model_signal: vote.signal,
+                signal_alignment: alignment,
+                pnl_contribution: contribution,
+            }
+        })
+        .collect();
+
+    let total_attributed: f64 = attributions.iter().map(|a| a.pnl_contribution).sum();
+    let residual = realized_pnl - total_attributed;
+
+    AttributionResult {
+        attributions,
+        realized_pnl,
+        residual,
+    }
+}
+
+/// Compute signal alignment: does the model's signal direction match the realized direction?
+///
+/// - Zero signal → neutral (0.0)
+/// - Same sign → aligned (1.0)
+/// - Opposite sign → misaligned (-1.0)
+fn compute_alignment(model_signal: f64, realized_direction: f64) -> f64 {
+    if model_signal.abs() < 1e-12 {
+        return 0.0;
+    }
+    if model_signal.signum() == realized_direction {
+        1.0
+    } else {
+        -1.0
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use ml::ensemble::decision::{ModelVote, TradingAction};
+    use std::collections::HashMap;
+
+    fn make_decision(votes: Vec<(&str, f64, f64)>) -> EnsembleDecision {
+        let mut model_votes = HashMap::new();
+        let mut total_signal = 0.0;
+
+        for (id, signal, weight) in &votes {
+            total_signal += signal * weight;
+            model_votes.insert(
+                id.to_string(),
+                ModelVote::new(id.to_string(), *signal, signal.abs(), *weight)
+                    .with_model_type("DQN".into()),
+            );
+        }
+
+        EnsembleDecision::new(
+            if total_signal > 0.0 {
+                TradingAction::Buy
+            } else if total_signal < 0.0 {
+                TradingAction::Sell
+            } else {
+                TradingAction::Hold
+            },
+            0.8,
+            total_signal,
+            0.0,
+            model_votes,
+        )
+    }
+
+    #[test]
+    fn test_correct_attribution_positive_pnl() {
+        // Two models, both bullish, trade was profitable
+        let decision = make_decision(vec![
+            ("dqn", 0.8, 0.6), // 60% weight, bullish
+            ("ppo", 0.5, 0.4), // 40% weight, bullish
+        ]);
+
+        let result = attribute(&decision, 100.0);
+        assert_eq!(result.attributions.len(), 2);
+
+        // DQN: 0.6 × 1.0 × 100.0 = 60.0
+        let dqn = result.attributions.iter().find(|a| a.model_id == "dqn");
+        assert!(dqn.is_some());
+        if let Some(dqn) = dqn {
+            assert!((dqn.pnl_contribution - 60.0).abs() < 1e-10);
+            assert!((dqn.signal_alignment - 1.0).abs() < 1e-10);
+        }
+
+        // PPO: 0.4 × 1.0 × 100.0 = 40.0
+        let ppo = result.attributions.iter().find(|a| a.model_id == "ppo");
+        assert!(ppo.is_some());
+        if let Some(ppo) = ppo {
+            assert!((ppo.pnl_contribution - 40.0).abs() < 1e-10);
+        }
+
+        // Sum should equal realized PnL
+        assert!(result.residual.abs() < 1e-10);
+    }
+
+    #[test]
+    fn test_attribution_with_disagreement() {
+        // DQN bullish, PPO bearish, trade was profitable (bullish correct)
+        let decision = make_decision(vec![
+            ("dqn", 0.8, 0.6),  // bullish, correct
+            ("ppo", -0.5, 0.4), // bearish, wrong
+        ]);
+
+        let result = attribute(&decision, 100.0);
+
+        // DQN: 0.6 × 1.0 × 100.0 = 60.0 (aligned)
+        if let Some(dqn) = result.attributions.iter().find(|a| a.model_id == "dqn") {
+            assert!((dqn.pnl_contribution - 60.0).abs() < 1e-10);
+        }
+
+        // PPO: 0.4 × (-1.0) × 100.0 = -40.0 (misaligned)
+        if let Some(ppo) = result.attributions.iter().find(|a| a.model_id == "ppo") {
+            assert!((ppo.pnl_contribution - (-40.0)).abs() < 1e-10);
+        }
+    }
+
+    #[test]
+    fn test_zero_pnl() {
+        let decision = make_decision(vec![("dqn", 0.8, 0.5), ("ppo", 0.5, 0.5)]);
+
+        let result = attribute(&decision, 0.0);
+
+        // All contributions should be zero
+        for attr in &result.attributions {
+            assert!(attr.pnl_contribution.abs() < 1e-12);
+            assert!(attr.signal_alignment.abs() < 1e-12);
+        }
+    }
+
+    #[test]
+    fn test_all_models_wrong() {
+        // All bearish, but market went up → they were wrong
+        let decision = make_decision(vec![("dqn", -0.7, 0.5), ("ppo", -0.6, 0.5)]);
+
+        let result = attribute(&decision, 50.0);
+
+        // Both bearish but pnl positive → misaligned
+        for attr in &result.attributions {
+            assert!((attr.signal_alignment - (-1.0)).abs() < 1e-10);
+            assert!(attr.pnl_contribution < 0.0);
+        }
+    }
+
+    #[test]
+    fn test_all_models_correct_bearish() {
+        // All bearish, PnL negative → bearish direction correct
+        let decision = make_decision(vec![("dqn", -0.7, 0.5), ("ppo", -0.6, 0.5)]);
+
+        let result = attribute(&decision, -100.0);
+
+        // Both bearish, PnL negative → aligned
+        for attr in &result.attributions {
+            assert!((attr.signal_alignment - 1.0).abs() < 1e-10);
+            assert!(attr.pnl_contribution > 0.0);
+        }
+    }
+
+    #[test]
+    fn test_neutral_model_gets_zero() {
+        let decision = make_decision(vec![
+            ("dqn", 0.8, 0.5),
+            ("neutral", 0.0, 0.5), // Zero signal = neutral
+        ]);
+
+        let result = attribute(&decision, 100.0);
+
+        if let Some(neutral) = result.attributions.iter().find(|a| a.model_id == "neutral") {
+            assert!(neutral.signal_alignment.abs() < 1e-12);
+            assert!(neutral.pnl_contribution.abs() < 1e-12);
+        }
+    }
+
+    #[test]
+    fn test_empty_votes() {
+        let decision = EnsembleDecision::new(
+            TradingAction::Hold,
+            0.0,
+            0.0,
+            0.0,
+            HashMap::new(),
+        );
+
+        let result = attribute(&decision, 100.0);
+        assert!(result.attributions.is_empty());
+        assert!((result.residual - 100.0).abs() < 1e-10);
+    }
+}
diff --git a/services/trading_service/src/feedback_loop.rs b/services/trading_service/src/feedback_loop.rs
new file mode 100644
index 000000000..a92b73dd6
--- /dev/null
+++ b/services/trading_service/src/feedback_loop.rs
@@ -0,0 +1,529 @@
+//! Autonomous Feedback Loop Orchestrator
+//!
+//! Runs as a background tokio task. Periodically:
+//! 1. Queries rolling performance metrics (from QuestDB or in-memory)
+//! 2. Runs weight optimizer → applies weight adjustments
+//! 3. Runs gate optimizer → applies threshold adjustments
+//! 4. Checks retraining triggers (Sharpe < threshold, accuracy < threshold)
+//! 5. Monitors ensemble Sharpe for kill switch activation
+
+use ml::ensemble::conviction_gates::ConvictionGateConfig;
+use ml::ensemble::gate_optimizer::{
+    GateBucketMetrics, GateOptimizationResult, GateOptimizer, GateOptimizerConfig,
+};
+use ml::ensemble::weight_optimizer::{
+    ModelRollingMetrics, OptimizationResult, WeightOptimizer, WeightOptimizerConfig,
+};
+use std::collections::HashMap;
+use std::sync::Arc;
+use std::time::Duration;
+use tokio::sync::Mutex;
+use tracing::{error, info, warn};
+
+/// Retraining trigger reasons
+#[derive(Debug, Clone)]
+pub enum RetrainingTrigger {
+    /// Sharpe ratio dropped below threshold
+    LowSharpe { model_id: String, sharpe: f64 },
+    /// Accuracy dropped below threshold
+    LowAccuracy { model_id: String, accuracy: f64 },
+    /// Models are too correlated
+    HighCorrelation { pair: (String, String), correlation: f64 },
+}
+
+/// Configuration for the feedback loop
+#[derive(Debug, Clone)]
+pub struct FeedbackLoopConfig {
+    /// How often to run the optimization cycle
+    pub cycle_interval: Duration,
+    /// Kill switch: if ensemble Sharpe drops below this, freeze all optimizers
+    pub kill_switch_threshold: f64,
+    /// Sharpe ratio threshold for retraining trigger
+    pub retrain_sharpe_threshold: f64,
+    /// Accuracy threshold for retraining trigger
+    pub retrain_accuracy_threshold: f64,
+    /// Correlation threshold for retraining trigger
+    pub retrain_correlation_threshold: f64,
+}
+
+impl Default for FeedbackLoopConfig {
+    fn default() -> Self {
+        Self {
+            cycle_interval: Duration::from_secs(24 * 3600), // 24h
+            kill_switch_threshold: -1.0,
+            retrain_sharpe_threshold: -0.5,
+            retrain_accuracy_threshold: 0.45,
+            retrain_correlation_threshold: 0.90,
+        }
+    }
+}
+
+/// Result of a single feedback loop cycle
+#[derive(Debug, Clone)]
+pub struct CycleResult {
+    /// Weight adjustments applied (if any)
+    pub weight_result: OptimizationResult,
+    /// Gate adjustments applied (if any)
+    pub gate_result: GateOptimizationResult,
+    /// Retraining triggers detected
+    pub retraining_triggers: Vec<RetrainingTrigger>,
+    /// Whether kill switch was activated this cycle
+    pub kill_switch_activated: bool,
+}
+
+/// Trait for providing rolling metrics to the feedback loop.
+/// In production, this queries QuestDB. In tests, it's mocked.
+pub trait MetricsProvider: Send + Sync {
+    /// Get rolling model-level metrics for weight optimization
+    fn get_model_metrics(&self) -> Vec<ModelRollingMetrics>;
+
+    /// Get confidence-bucketed metrics for gate optimization
+    fn get_gate_buckets(&self) -> Vec<GateBucketMetrics>;
+
+    /// Get current model weights
+    fn get_current_weights(&self) -> HashMap<String, f64>;
+
+    /// Get current conviction gate config
+    fn get_gate_config(&self) -> ConvictionGateConfig;
+
+    /// Get 7-day ensemble Sharpe ratio
+    fn get_ensemble_sharpe_7d(&self) -> f64;
+
+    /// Get model pairwise correlations (for retraining triggers)
+    fn get_model_correlations(&self) -> Vec<((String, String), f64)>;
+}
+
+/// Autonomous feedback loop orchestrator
+pub struct FeedbackLoop {
+    weight_optimizer: Arc<Mutex<WeightOptimizer>>,
+    gate_optimizer: Arc<Mutex<GateOptimizer>>,
+    config: FeedbackLoopConfig,
+}
+
+impl FeedbackLoop {
+    pub fn new(config: FeedbackLoopConfig) -> Self {
+        Self {
+            weight_optimizer: Arc::new(Mutex::new(WeightOptimizer::new(
+                WeightOptimizerConfig::default(),
+            ))),
+            gate_optimizer: Arc::new(Mutex::new(GateOptimizer::new(
+                GateOptimizerConfig::default(),
+            ))),
+            config,
+        }
+    }
+
+    /// Create with custom optimizer configs
+    pub fn with_optimizers(
+        config: FeedbackLoopConfig,
+        weight_config: WeightOptimizerConfig,
+        gate_config: GateOptimizerConfig,
+    ) -> Self {
+        Self {
+            weight_optimizer: Arc::new(Mutex::new(WeightOptimizer::new(weight_config))),
+            gate_optimizer: Arc::new(Mutex::new(GateOptimizer::new(gate_config))),
+            config,
+        }
+    }
+
+    /// Run one optimization cycle. Returns the results for observability.
+    pub async fn run_cycle(&self, metrics: &dyn MetricsProvider) -> CycleResult {
+        // 1. Check kill switch
+        let ensemble_sharpe = metrics.get_ensemble_sharpe_7d();
+        let kill_switch_activated = ensemble_sharpe < self.config.kill_switch_threshold;
+
+        if kill_switch_activated {
+            warn!(
+                ensemble_sharpe = ensemble_sharpe,
+                threshold = self.config.kill_switch_threshold,
+                "Kill switch activated — freezing all optimizers"
+            );
+            self.weight_optimizer.lock().await.freeze();
+            self.gate_optimizer.lock().await.freeze();
+
+            return CycleResult {
+                weight_result: OptimizationResult::KillSwitchActive,
+                gate_result: GateOptimizationResult::KillSwitchActive,
+                retraining_triggers: vec![],
+                kill_switch_activated: true,
+            };
+        }
+
+        // 2. Run weight optimization
+        let model_metrics = metrics.get_model_metrics();
+        let current_weights = metrics.get_current_weights();
+        let weight_result = self
+            .weight_optimizer
+            .lock()
+            .await
+            .optimize(&current_weights, &model_metrics);
+
+        if let OptimizationResult::Adjusted(ref adjustments) = weight_result {
+            info!(
+                count = adjustments.len(),
+                "Feedback loop: weight adjustments proposed"
+            );
+        }
+
+        // 3. Run gate optimization
+        let gate_config = metrics.get_gate_config();
+        let buckets = metrics.get_gate_buckets();
+        let gate_result = self
+            .gate_optimizer
+            .lock()
+            .await
+            .optimize(&gate_config, &buckets);
+
+        if let GateOptimizationResult::Adjusted(ref adjustments) = gate_result {
+            info!(
+                count = adjustments.len(),
+                "Feedback loop: gate adjustments proposed"
+            );
+        }
+
+        // 4. Check retraining triggers
+        let retraining_triggers = self.check_retraining_triggers(metrics);
+
+        if !retraining_triggers.is_empty() {
+            warn!(
+                count = retraining_triggers.len(),
+                "Feedback loop: retraining triggers detected"
+            );
+        }
+
+        CycleResult {
+            weight_result,
+            gate_result,
+            retraining_triggers,
+            kill_switch_activated: false,
+        }
+    }
+
+    /// Spawn the feedback loop as a background tokio task.
+    ///
+    /// Returns a handle that can be used to abort the loop.
+    pub fn spawn(
+        self: Arc<Self>,
+        metrics: Arc<dyn MetricsProvider>,
+    ) -> tokio::task::JoinHandle<()> {
+        let interval = self.config.cycle_interval;
+
+        tokio::spawn(async move {
+            let mut ticker = tokio::time::interval(interval);
+            // Skip the first immediate tick
+            ticker.tick().await;
+
+            loop {
+                ticker.tick().await;
+                info!("Feedback loop: starting optimization cycle");
+
+                let result = self.run_cycle(metrics.as_ref()).await;
+
+                if result.kill_switch_activated {
+                    error!("Feedback loop: kill switch active, pausing until manual intervention");
+                    // Sleep longer when kill switch is active to avoid log spam
+                    tokio::time::sleep(Duration::from_secs(3600)).await;
+                }
+            }
+        })
+    }
+
+    /// Freeze both optimizers (manual kill switch)
+    pub async fn freeze(&self) {
+        self.weight_optimizer.lock().await.freeze();
+        self.gate_optimizer.lock().await.freeze();
+    }
+
+    /// Unfreeze both optimizers (requires human acknowledgment)
+    pub async fn unfreeze(&self) {
+        self.weight_optimizer.lock().await.unfreeze();
+        self.gate_optimizer.lock().await.unfreeze();
+    }
+
+    /// Check if either optimizer is frozen
+    pub async fn is_frozen(&self) -> bool {
+        self.weight_optimizer.lock().await.is_frozen()
+            || self.gate_optimizer.lock().await.is_frozen()
+    }
+
+    fn check_retraining_triggers(&self, metrics: &dyn MetricsProvider) -> Vec<RetrainingTrigger> {
+        let mut triggers = Vec::new();
+
+        // Check per-model metrics
+        for m in &metrics.get_model_metrics() {
+            if m.sharpe_30d < self.config.retrain_sharpe_threshold {
+                triggers.push(RetrainingTrigger::LowSharpe {
+                    model_id: m.model_id.clone(),
+                    sharpe: m.sharpe_30d,
+                });
+            }
+
+            if m.prediction_accuracy < self.config.retrain_accuracy_threshold {
+                triggers.push(RetrainingTrigger::LowAccuracy {
+                    model_id: m.model_id.clone(),
+                    accuracy: m.prediction_accuracy,
+                });
+            }
+        }
+
+        // Check correlations
+        for ((a, b), corr) in &metrics.get_model_correlations() {
+            if *corr > self.config.retrain_correlation_threshold {
+                triggers.push(RetrainingTrigger::HighCorrelation {
+                    pair: (a.clone(), b.clone()),
+                    correlation: *corr,
+                });
+            }
+        }
+
+        triggers
+    }
+}
+
+impl std::fmt::Debug for FeedbackLoop {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        f.debug_struct("FeedbackLoop")
+            .field("config", &self.config)
+            .finish()
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use std::time::Instant;
+
+    /// Mock metrics provider for testing
+    struct MockMetrics {
+        model_metrics: Vec<ModelRollingMetrics>,
+        gate_buckets: Vec<GateBucketMetrics>,
+        weights: HashMap<String, f64>,
+        gate_config: ConvictionGateConfig,
+        ensemble_sharpe: f64,
+        correlations: Vec<((String, String), f64)>,
+    }
+
+    impl MockMetrics {
+        fn healthy() -> Self {
+            Self {
+                model_metrics: vec![
+                    ModelRollingMetrics {
+                        model_id: "dqn".into(),
+                        sharpe_30d: 1.5,
+                        win_rate_30d: 0.58,
+                        prediction_accuracy: 0.62,
+                        trade_count: 200,
+                        deployed_at: Instant::now() - Duration::from_secs(30 * 24 * 3600),
+                    },
+                    ModelRollingMetrics {
+                        model_id: "ppo".into(),
+                        sharpe_30d: 0.8,
+                        win_rate_30d: 0.55,
+                        prediction_accuracy: 0.58,
+                        trade_count: 200,
+                        deployed_at: Instant::now() - Duration::from_secs(30 * 24 * 3600),
+                    },
+                ],
+                gate_buckets: vec![
+                    GateBucketMetrics {
+                        confidence_lower: 0.50,
+                        confidence_upper: 0.60,
+                        win_rate: 0.56,
+                        trade_count: 100,
+                        avg_pnl: 0.002,
+                    },
+                    GateBucketMetrics {
+                        confidence_lower: 0.60,
+                        confidence_upper: 0.70,
+                        win_rate: 0.58,
+                        trade_count: 100,
+                        avg_pnl: 0.003,
+                    },
+                    GateBucketMetrics {
+                        confidence_lower: 0.70,
+                        confidence_upper: 0.80,
+                        win_rate: 0.62,
+                        trade_count: 100,
+                        avg_pnl: 0.005,
+                    },
+                ],
+                weights: {
+                    let mut w = HashMap::new();
+                    w.insert("dqn".into(), 0.5);
+                    w.insert("ppo".into(), 0.5);
+                    w
+                },
+                gate_config: ConvictionGateConfig::default(),
+                ensemble_sharpe: 1.2,
+                correlations: vec![],
+            }
+        }
+    }
+
+    impl MetricsProvider for MockMetrics {
+        fn get_model_metrics(&self) -> Vec<ModelRollingMetrics> {
+            self.model_metrics.clone()
+        }
+        fn get_gate_buckets(&self) -> Vec<GateBucketMetrics> {
+            self.gate_buckets.clone()
+        }
+        fn get_current_weights(&self) -> HashMap<String, f64> {
+            self.weights.clone()
+        }
+        fn get_gate_config(&self) -> ConvictionGateConfig {
+            self.gate_config.clone()
+        }
+        fn get_ensemble_sharpe_7d(&self) -> f64 {
+            self.ensemble_sharpe
+        }
+        fn get_model_correlations(&self) -> Vec<((String, String), f64)> {
+            self.correlations.clone()
+        }
+    }
+
+    #[tokio::test]
+    async fn test_healthy_cycle_adjusts_weights() {
+        let mut config = FeedbackLoopConfig::default();
+        config.cycle_interval = Duration::from_millis(100);
+
+        let feedback = FeedbackLoop::with_optimizers(
+            config,
+            WeightOptimizerConfig {
+                cooldown: Duration::ZERO,
+                ..Default::default()
+            },
+            GateOptimizerConfig {
+                cooldown: Duration::ZERO,
+                ..Default::default()
+            },
+        );
+
+        let metrics = MockMetrics::healthy();
+        let result = feedback.run_cycle(&metrics).await;
+
+        // Should produce weight adjustments (models have different Sharpe)
+        assert!(matches!(result.weight_result, OptimizationResult::Adjusted(_)));
+        assert!(!result.kill_switch_activated);
+        assert!(result.retraining_triggers.is_empty());
+    }
+
+    #[tokio::test]
+    async fn test_kill_switch_freezes_optimizers() {
+        let feedback = FeedbackLoop::new(FeedbackLoopConfig {
+            kill_switch_threshold: -1.0,
+            ..Default::default()
+        });
+
+        let mut metrics = MockMetrics::healthy();
+        metrics.ensemble_sharpe = -2.0; // Below kill switch
+
+        let result = feedback.run_cycle(&metrics).await;
+
+        assert!(result.kill_switch_activated);
+        assert!(matches!(
+            result.weight_result,
+            OptimizationResult::KillSwitchActive
+        ));
+        assert!(matches!(
+            result.gate_result,
+            GateOptimizationResult::KillSwitchActive
+        ));
+        assert!(feedback.is_frozen().await);
+    }
+
+    #[tokio::test]
+    async fn test_manual_freeze_unfreeze() {
+        let feedback = FeedbackLoop::new(FeedbackLoopConfig::default());
+
+        assert!(!feedback.is_frozen().await);
+        feedback.freeze().await;
+        assert!(feedback.is_frozen().await);
+
+        let metrics = MockMetrics::healthy();
+        let result = feedback.run_cycle(&metrics).await;
+        assert!(matches!(
+            result.weight_result,
+            OptimizationResult::KillSwitchActive
+        ));
+
+        feedback.unfreeze().await;
+        assert!(!feedback.is_frozen().await);
+    }
+
+    #[tokio::test]
+    async fn test_retraining_trigger_low_sharpe() {
+        let config = FeedbackLoopConfig {
+            retrain_sharpe_threshold: -0.5,
+            ..Default::default()
+        };
+        let feedback = FeedbackLoop::new(config);
+
+        let mut metrics = MockMetrics::healthy();
+        metrics.model_metrics[1].sharpe_30d = -1.0; // PPO tanking
+
+        let result = feedback.run_cycle(&metrics).await;
+
+        assert_eq!(result.retraining_triggers.len(), 1);
+        if let RetrainingTrigger::LowSharpe { ref model_id, .. } =
+            result.retraining_triggers[0]
+        {
+            assert_eq!(model_id, "ppo");
+        } else {
+            panic!("Expected LowSharpe trigger");
+        }
+    }
+
+    #[tokio::test]
+    async fn test_retraining_trigger_low_accuracy() {
+        let config = FeedbackLoopConfig {
+            retrain_accuracy_threshold: 0.45,
+            ..Default::default()
+        };
+        let feedback = FeedbackLoop::new(config);
+
+        let mut metrics = MockMetrics::healthy();
+        metrics.model_metrics[0].prediction_accuracy = 0.40; // DQN accuracy dropped
+
+        let result = feedback.run_cycle(&metrics).await;
+
+        let accuracy_triggers: Vec<_> = result
+            .retraining_triggers
+            .iter()
+            .filter(|t| matches!(t, RetrainingTrigger::LowAccuracy { .. }))
+            .collect();
+        assert_eq!(accuracy_triggers.len(), 1);
+    }
+
+    #[tokio::test]
+    async fn test_retraining_trigger_high_correlation() {
+        let config = FeedbackLoopConfig {
+            retrain_correlation_threshold: 0.90,
+            ..Default::default()
+        };
+        let feedback = FeedbackLoop::new(config);
+
+        let mut metrics = MockMetrics::healthy();
+        metrics.correlations = vec![
+            (("dqn".into(), "ppo".into()), 0.95), // Too correlated
+        ];
+
+        let result = feedback.run_cycle(&metrics).await;
+
+        let corr_triggers: Vec<_> = result
+            .retraining_triggers
+            .iter()
+            .filter(|t| matches!(t, RetrainingTrigger::HighCorrelation { .. }))
+            .collect();
+        assert_eq!(corr_triggers.len(), 1);
+    }
+
+    #[tokio::test]
+    async fn test_no_triggers_healthy_system() {
+        let feedback = FeedbackLoop::new(FeedbackLoopConfig::default());
+        let metrics = MockMetrics::healthy();
+
+        let result = feedback.run_cycle(&metrics).await;
+        assert!(result.retraining_triggers.is_empty());
+        assert!(!result.kill_switch_activated);
+    }
+}
diff --git a/services/trading_service/src/lib.rs b/services/trading_service/src/lib.rs
index 033d2086f..e1c521766 100644
--- a/services/trading_service/src/lib.rs
+++ b/services/trading_service/src/lib.rs
@@ -163,6 +163,15 @@ pub mod assets;
 /// Health check endpoints for Kubernetes probes
 pub mod health;
 
+/// P&L attribution: decomposes realized trade P&L into per-model contributions
+pub mod attribution;
+
+/// Autonomous feedback loop: weight/gate optimization with kill switch
+pub mod feedback_loop;
+
+/// QuestDB-backed metrics provider for the feedback loop
+pub mod questdb_metrics;
+
 // Re-export for tests
 pub use ensemble_coordinator::EnsembleCoordinator;
 pub use paper_trading_executor::PaperTradingExecutor;
diff --git a/services/trading_service/src/main.rs b/services/trading_service/src/main.rs
index 918150257..44850482b 100644
--- a/services/trading_service/src/main.rs
+++ b/services/trading_service/src/main.rs
@@ -37,6 +37,8 @@ use trading_service::services::ml_performance_monitor::MLPerformanceMonitor;
 use trading_service::services::monitoring::MonitoringServiceImpl;
 use trading_service::services::risk::RiskServiceImpl;
 use trading_service::services::trading::TradingServiceImpl;
+use trading_service::feedback_loop::{FeedbackLoop, FeedbackLoopConfig};
+use trading_service::questdb_metrics::QuestDBMetricsProvider;
 use trading_service::state::TradingServiceState;
 
 /// Default configuration values
@@ -528,6 +530,44 @@ async fn main() -> Result<()> {
         warn!("⚠️ Ensemble coordinator not available - prediction generation loop disabled");
     }
 
+    // Initialize autonomous feedback loop (weight + gate optimization with kill switch)
+    {
+        let feedback_config = FeedbackLoopConfig {
+            cycle_interval: Duration::from_secs(
+                std::env::var("FEEDBACK_LOOP_INTERVAL_SECS")
+                    .ok()
+                    .and_then(|s| s.parse().ok())
+                    .unwrap_or(24 * 3600),
+            ),
+            kill_switch_threshold: std::env::var("FEEDBACK_KILL_SWITCH_SHARPE")
+                .ok()
+                .and_then(|s| s.parse().ok())
+                .unwrap_or(-1.0),
+            ..Default::default()
+        };
+
+        let feedback_loop = Arc::new(FeedbackLoop::new(feedback_config));
+
+        // Connect to QuestDB for real metrics (pg wire protocol on port 8812)
+        let questdb_url = std::env::var("QUESTDB_PG_URL")
+            .unwrap_or_else(|_| "postgresql://admin:quest@localhost:8812/qdb".to_string());
+
+        match QuestDBMetricsProvider::try_new(&questdb_url).await {
+            Some(provider) => {
+                if let Err(e) = provider.ensure_tables().await {
+                    warn!("Failed to create QuestDB tables: {} — feedback loop will retry", e);
+                }
+                let metrics_provider: Arc<dyn trading_service::feedback_loop::MetricsProvider> =
+                    Arc::new(provider);
+                let _feedback_handle = Arc::clone(&feedback_loop).spawn(metrics_provider);
+                info!("Autonomous feedback loop spawned with QuestDB metrics provider");
+            }
+            None => {
+                warn!("QuestDB unavailable — feedback loop disabled until QUESTDB_PG_URL is reachable");
+            }
+        }
+    }
+
     // Subscribe to ML performance alerts
     let monitor_clone = Arc::clone(&ml_performance_monitor);
     tokio::spawn(async move {
diff --git a/services/trading_service/src/questdb_metrics.rs b/services/trading_service/src/questdb_metrics.rs
new file mode 100644
index 000000000..4aae8a2e0
--- /dev/null
+++ b/services/trading_service/src/questdb_metrics.rs
@@ -0,0 +1,422 @@
+//! QuestDB-backed MetricsProvider for the feedback loop
+//!
+//! Queries QuestDB via PostgreSQL wire protocol (port 8812) to provide
+//! rolling model metrics, confidence buckets, and ensemble Sharpe for
+//! the autonomous weight/gate optimizers.
+
+use crate::feedback_loop::MetricsProvider;
+use ml::ensemble::conviction_gates::ConvictionGateConfig;
+use ml::ensemble::gate_optimizer::GateBucketMetrics;
+use ml::ensemble::weight_optimizer::ModelRollingMetrics;
+use sqlx::postgres::PgPool;
+use std::collections::HashMap;
+use std::sync::Arc;
+use std::time::Instant;
+use tokio::sync::RwLock;
+use tracing::{debug, warn};
+
+/// QuestDB metrics provider that queries real time-series data.
+///
+/// Tables expected in QuestDB (created on first write via ILP):
+/// - `model_predictions`: model_id, signal, confidence, timestamp
+/// - `trade_outcomes`: model_id, realized_pnl, signal_alignment, timestamp
+/// - `ensemble_metrics`: sharpe_7d, sharpe_30d, timestamp
+pub struct QuestDBMetricsProvider {
+    pool: PgPool,
+    /// Cached gate config (updated from external source)
+    gate_config: Arc<RwLock<ConvictionGateConfig>>,
+    /// Cached weights (updated by the feedback loop itself)
+    current_weights: Arc<RwLock<HashMap<String, f64>>>,
+}
+
+impl QuestDBMetricsProvider {
+    pub async fn new(questdb_pg_url: &str) -> Result<Self, sqlx::Error> {
+        let pool = PgPool::connect(questdb_pg_url).await?;
+
+        Ok(Self {
+            pool,
+            gate_config: Arc::new(RwLock::new(ConvictionGateConfig::default())),
+            current_weights: Arc::new(RwLock::new(HashMap::new())),
+        })
+    }
+
+    /// Try to connect, returning None if QuestDB is unavailable
+    pub async fn try_new(questdb_pg_url: &str) -> Option<Self> {
+        match Self::new(questdb_pg_url).await {
+            Ok(provider) => Some(provider),
+            Err(e) => {
+                warn!("QuestDB unavailable at {}: {} — feedback loop will use defaults", questdb_pg_url, e);
+                None
+            }
+        }
+    }
+
+    /// Update the cached gate config (called when config changes)
+    pub async fn set_gate_config(&self, config: ConvictionGateConfig) {
+        *self.gate_config.write().await = config;
+    }
+
+    /// Update the cached weights (called after weight adjustments)
+    pub async fn set_weights(&self, weights: HashMap<String, f64>) {
+        *self.current_weights.write().await = weights;
+    }
+
+    /// Create the required tables if they don't exist.
+    /// QuestDB auto-creates tables on ILP write, but we create them
+    /// explicitly for querying so tests can verify structure.
+    pub async fn ensure_tables(&self) -> Result<(), sqlx::Error> {
+        // QuestDB uses CREATE TABLE IF NOT EXISTS with designated timestamp
+        sqlx::query(
+            "CREATE TABLE IF NOT EXISTS model_predictions (
+                model_id SYMBOL,
+                signal DOUBLE,
+                confidence DOUBLE,
+                prediction_accuracy DOUBLE,
+                win_rate DOUBLE,
+                timestamp TIMESTAMP
+            ) TIMESTAMP(timestamp) PARTITION BY DAY;"
+        )
+        .execute(&self.pool)
+        .await?;
+
+        sqlx::query(
+            "CREATE TABLE IF NOT EXISTS trade_outcomes (
+                model_id SYMBOL,
+                realized_pnl DOUBLE,
+                signal_alignment DOUBLE,
+                confidence_bucket_lower DOUBLE,
+                confidence_bucket_upper DOUBLE,
+                timestamp TIMESTAMP
+            ) TIMESTAMP(timestamp) PARTITION BY DAY;"
+        )
+        .execute(&self.pool)
+        .await?;
+
+        sqlx::query(
+            "CREATE TABLE IF NOT EXISTS ensemble_metrics (
+                sharpe_7d DOUBLE,
+                sharpe_30d DOUBLE,
+                total_pnl DOUBLE,
+                timestamp TIMESTAMP
+            ) TIMESTAMP(timestamp) PARTITION BY DAY;"
+        )
+        .execute(&self.pool)
+        .await?;
+
+        Ok(())
+    }
+
+    /// Query 30-day rolling Sharpe per model
+    async fn query_model_sharpe(&self) -> HashMap<String, f64> {
+        let result: Result<Vec<(String, f64)>, _> = sqlx::query_as(
+            "SELECT model_id, coalesce(avg(signal) / (stddev(signal) + 1e-10), 0.0) as sharpe
+             FROM model_predictions
+             WHERE timestamp > dateadd('d', -30, now())
+             GROUP BY model_id"
+        )
+        .fetch_all(&self.pool)
+        .await;
+
+        match result {
+            Ok(rows) => rows.into_iter().collect(),
+            Err(e) => {
+                debug!("QuestDB model Sharpe query failed: {}", e);
+                HashMap::new()
+            }
+        }
+    }
+
+    /// Query 30-day win rate per model
+    async fn query_model_win_rates(&self) -> HashMap<String, (f64, u64)> {
+        let result: Result<Vec<(String, f64, i64)>, _> = sqlx::query_as(
+            "SELECT model_id,
+                    sum(CASE WHEN signal_alignment > 0 THEN 1.0 ELSE 0.0 END) / count(1) as win_rate,
+                    count(*) as trade_count
+             FROM trade_outcomes
+             WHERE timestamp > dateadd('d', -30, now())
+             GROUP BY model_id"
+        )
+        .fetch_all(&self.pool)
+        .await;
+
+        match result {
+            Ok(rows) => rows
+                .into_iter()
+                .map(|(id, wr, cnt)| (id, (wr, cnt as u64)))
+                .collect(),
+            Err(e) => {
+                debug!("QuestDB win rate query failed: {}", e);
+                HashMap::new()
+            }
+        }
+    }
+
+    /// Query confidence-bucketed win rates for gate optimization
+    async fn query_confidence_buckets(&self) -> Vec<GateBucketMetrics> {
+        // 5 buckets: [0.5-0.6), [0.6-0.7), [0.7-0.8), [0.8-0.9), [0.9-1.0]
+        let result: Result<Vec<(f64, f64, f64, i64, f64)>, _> = sqlx::query_as(
+            "SELECT confidence_bucket_lower,
+                    confidence_bucket_upper,
+                    sum(CASE WHEN signal_alignment > 0 THEN 1.0 ELSE 0.0 END) / count(1) as win_rate,
+                    count(*) as trade_count,
+                    avg(realized_pnl) as avg_pnl
+             FROM trade_outcomes
+             WHERE timestamp > dateadd('d', -30, now())
+               AND confidence_bucket_lower IS NOT NULL
+             GROUP BY confidence_bucket_lower, confidence_bucket_upper
+             ORDER BY confidence_bucket_lower"
+        )
+        .fetch_all(&self.pool)
+        .await;
+
+        match result {
+            Ok(rows) => rows
+                .into_iter()
+                .map(|(lower, upper, win_rate, count, avg_pnl)| GateBucketMetrics {
+                    confidence_lower: lower,
+                    confidence_upper: upper,
+                    win_rate,
+                    trade_count: count as u64,
+                    avg_pnl,
+                })
+                .collect(),
+            Err(e) => {
+                debug!("QuestDB confidence bucket query failed: {}", e);
+                Vec::new()
+            }
+        }
+    }
+
+    /// Query latest ensemble Sharpe (7-day)
+    async fn query_ensemble_sharpe(&self) -> f64 {
+        let result: Result<Option<(f64,)>, _> = sqlx::query_as(
+            "SELECT sharpe_7d FROM ensemble_metrics ORDER BY timestamp DESC LIMIT 1"
+        )
+        .fetch_optional(&self.pool)
+        .await;
+
+        match result {
+            Ok(Some((sharpe,))) => sharpe,
+            Ok(None) => 0.0,
+            Err(e) => {
+                debug!("QuestDB ensemble Sharpe query failed: {}", e);
+                0.0
+            }
+        }
+    }
+}
+
+impl MetricsProvider for QuestDBMetricsProvider {
+    fn get_model_metrics(&self) -> Vec<ModelRollingMetrics> {
+        // block_in_place allows blocking inside a multi-threaded tokio runtime
+        // by moving other tasks off this thread. Safe because the feedback loop
+        // runs periodically (24h cycle) — not on the hot path.
+        let handle = tokio::runtime::Handle::current();
+        tokio::task::block_in_place(|| {
+            let sharpes = handle.block_on(self.query_model_sharpe());
+            let win_rates = handle.block_on(self.query_model_win_rates());
+
+            let mut metrics = Vec::new();
+            for (model_id, sharpe) in &sharpes {
+                let (win_rate, trade_count) = win_rates
+                    .get(model_id)
+                    .copied()
+                    .unwrap_or((0.5, 0));
+
+                metrics.push(ModelRollingMetrics {
+                    model_id: model_id.clone(),
+                    sharpe_30d: *sharpe,
+                    win_rate_30d: win_rate,
+                    prediction_accuracy: win_rate, // Using win_rate as proxy
+                    trade_count,
+                    deployed_at: Instant::now() - std::time::Duration::from_secs(30 * 24 * 3600),
+                });
+            }
+            metrics
+        })
+    }
+
+    fn get_gate_buckets(&self) -> Vec<GateBucketMetrics> {
+        let handle = tokio::runtime::Handle::current();
+        tokio::task::block_in_place(|| handle.block_on(self.query_confidence_buckets()))
+    }
+
+    fn get_current_weights(&self) -> HashMap<String, f64> {
+        let handle = tokio::runtime::Handle::current();
+        tokio::task::block_in_place(|| {
+            handle.block_on(async { self.current_weights.read().await.clone() })
+        })
+    }
+
+    fn get_gate_config(&self) -> ConvictionGateConfig {
+        let handle = tokio::runtime::Handle::current();
+        tokio::task::block_in_place(|| {
+            handle.block_on(async { self.gate_config.read().await.clone() })
+        })
+    }
+
+    fn get_ensemble_sharpe_7d(&self) -> f64 {
+        let handle = tokio::runtime::Handle::current();
+        tokio::task::block_in_place(|| handle.block_on(self.query_ensemble_sharpe()))
+    }
+
+    fn get_model_correlations(&self) -> Vec<((String, String), f64)> {
+        // Correlation requires cross-model signal comparison
+        // For now, query is deferred until we have enough data
+        Vec::new()
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    /// Drop and recreate all test tables for clean isolation between runs.
+    /// QuestDB TRUNCATE via PG wire can be flaky; DROP + CREATE is reliable.
+    async fn reset_tables(pool: &PgPool) {
+        for table in &["model_predictions", "trade_outcomes", "ensemble_metrics"] {
+            let drop_q = format!("DROP TABLE IF EXISTS {table};");
+            let _ = sqlx::query(&drop_q).execute(pool).await;
+        }
+    }
+
+    /// Poll QuestDB until a table has at least `min_rows` rows (up to 5s).
+    async fn wait_for_rows(pool: &PgPool, table: &str, min_rows: i64) {
+        for _ in 0..50 {
+            let query = format!("SELECT count(*) FROM {table}");
+            let row: Option<(i64,)> = sqlx::query_as(&query)
+                .fetch_optional(pool)
+                .await
+                .ok()
+                .flatten();
+            if row.map(|(c,)| c).unwrap_or(0) >= min_rows {
+                return;
+            }
+            tokio::time::sleep(std::time::Duration::from_millis(100)).await;
+        }
+    }
+
+    /// Comprehensive integration test: connect, write, read, verify metrics.
+    /// Single test avoids parallel interference on shared QuestDB tables.
+    ///
+    /// Run with: docker-compose up -d questdb
+    /// Then: SQLX_OFFLINE=true cargo test -p trading_service --lib -- questdb_metrics --include-ignored
+    #[tokio::test(flavor = "multi_thread")]
+    #[ignore = "Requires QuestDB on localhost:8812"]
+    async fn test_questdb_provider_full_lifecycle() {
+        let provider = QuestDBMetricsProvider::new("postgresql://admin:quest@localhost:8812/qdb")
+            .await
+            .expect("QuestDB should be running");
+
+        // Drop + recreate for clean slate (prior test runs may have left data)
+        reset_tables(&provider.pool).await;
+        provider.ensure_tables().await.expect("Tables should be created");
+
+        // Phase 1: Empty tables — metrics should return defaults
+        let metrics = provider.get_model_metrics();
+        assert!(metrics.is_empty(), "Empty tables should yield no metrics");
+
+        let sharpe = provider.get_ensemble_sharpe_7d();
+        assert!((sharpe - 0.0).abs() < 1e-10, "Empty ensemble Sharpe should be 0.0");
+
+        // Phase 2: Write test data (multiple rows per model for valid stddev)
+        sqlx::query(
+            "INSERT INTO model_predictions(model_id, signal, confidence, prediction_accuracy, win_rate, timestamp)
+             VALUES ('dqn', 0.8, 0.7, 0.62, 0.58, systimestamp()),
+                    ('dqn', 0.6, 0.8, 0.65, 0.60, systimestamp()),
+                    ('ppo', 0.3, 0.6, 0.55, 0.52, systimestamp()),
+                    ('ppo', 0.5, 0.7, 0.58, 0.54, systimestamp())"
+        )
+        .execute(&provider.pool)
+        .await
+        .expect("Insert predictions should succeed");
+
+        sqlx::query(
+            "INSERT INTO trade_outcomes(model_id, realized_pnl, signal_alignment, confidence_bucket_lower, confidence_bucket_upper, timestamp)
+             VALUES ('dqn', 100.0, 1.0, 0.60, 0.70, systimestamp()),
+                    ('dqn', -50.0, -1.0, 0.60, 0.70, systimestamp()),
+                    ('ppo', 75.0, 1.0, 0.50, 0.60, systimestamp())"
+        )
+        .execute(&provider.pool)
+        .await
+        .expect("Insert outcomes should succeed");
+
+        sqlx::query(
+            "INSERT INTO ensemble_metrics(sharpe_7d, sharpe_30d, total_pnl, timestamp)
+             VALUES (1.5, 1.2, 5000.0, systimestamp())"
+        )
+        .execute(&provider.pool)
+        .await
+        .expect("Insert ensemble metrics should succeed");
+
+        // Wait for QuestDB WAL to commit all tables
+        wait_for_rows(&provider.pool, "model_predictions", 4).await;
+        wait_for_rows(&provider.pool, "trade_outcomes", 3).await;
+        wait_for_rows(&provider.pool, "ensemble_metrics", 1).await;
+
+        // Phase 3: Read back and verify
+        let metrics = provider.get_model_metrics();
+        assert!(!metrics.is_empty(), "Should have model metrics after insert");
+        assert!(metrics.len() >= 2, "Should have at least DQN and PPO");
+
+        let sharpe = provider.get_ensemble_sharpe_7d();
+        assert!((sharpe - 1.5).abs() < 0.1, "Sharpe should be ~1.5, got {sharpe}");
+
+        let buckets = provider.get_gate_buckets();
+        assert!(!buckets.is_empty(), "Should have confidence buckets");
+    }
+
+    #[tokio::test(flavor = "multi_thread")]
+    #[ignore = "Requires QuestDB on localhost:8812"]
+    async fn test_questdb_provider_try_new_succeeds() {
+        let provider = QuestDBMetricsProvider::try_new("postgresql://admin:quest@localhost:8812/qdb").await;
+        assert!(provider.is_some(), "Should connect to QuestDB");
+    }
+
+    #[tokio::test(flavor = "multi_thread")]
+    async fn test_questdb_provider_try_new_fails_gracefully() {
+        // Connect to a non-existent QuestDB — should return None, not panic
+        let provider = QuestDBMetricsProvider::try_new("postgresql://admin:quest@localhost:19999/qdb").await;
+        assert!(provider.is_none(), "Should fail gracefully");
+    }
+
+    #[tokio::test(flavor = "multi_thread")]
+    #[ignore = "Requires QuestDB on localhost:8812"]
+    async fn test_feedback_loop_with_questdb() {
+        use crate::feedback_loop::{FeedbackLoop, FeedbackLoopConfig};
+        use ml::ensemble::gate_optimizer::GateOptimizerConfig;
+        use ml::ensemble::weight_optimizer::WeightOptimizerConfig;
+        use std::time::Duration;
+
+        let provider = QuestDBMetricsProvider::new("postgresql://admin:quest@localhost:8812/qdb")
+            .await
+            .expect("QuestDB should be running");
+
+        provider.ensure_tables().await.expect("Tables should be created");
+
+        // Set up some weights
+        let mut weights = HashMap::new();
+        weights.insert("dqn".to_string(), 0.5);
+        weights.insert("ppo".to_string(), 0.5);
+        provider.set_weights(weights).await;
+
+        let feedback = FeedbackLoop::with_optimizers(
+            FeedbackLoopConfig {
+                cycle_interval: Duration::from_millis(100),
+                ..Default::default()
+            },
+            WeightOptimizerConfig {
+                cooldown: Duration::ZERO,
+                ..Default::default()
+            },
+            GateOptimizerConfig {
+                cooldown: Duration::ZERO,
+                ..Default::default()
+            },
+        );
+
+        // Run a cycle — with empty QuestDB should report InsufficientData
+        let result = feedback.run_cycle(&provider).await;
+        assert!(!result.kill_switch_activated);
+    }
+}