diff --git a/src/audit.rs b/src/audit.rs
new file mode 100644
index 0000000..4121986
--- /dev/null
+++ b/src/audit.rs
@@ -0,0 +1,236 @@
+// Copyright Sunbeam Studios 2026
+// SPDX-License-Identifier: Apache-2.0
+
+//! Unified audit log record definition.
+//!
+//! This module is the **single source of truth** for the audit log schema.
+//! Both the proxy's `logging()` method (serialization) and every training/replay
+//! parser (deserialization) must use these types.  `deny_unknown_fields` ensures
+//! that any schema change is caught at parse time rather than silently ignored.
+
+use serde::{Deserialize, Serialize};
+
+/// Minimal probe struct to check if a JSON line is an audit log.
+#[derive(Deserialize)]
+struct Probe {
+    #[serde(default)]
+    fields: Option<ProbeFields>,
+}
+
+#[derive(Deserialize)]
+struct ProbeFields {
+    #[serde(default)]
+    target: Option<String>,
+}
+
+/// Top-level JSON line written by the tracing JSON layer.
+#[derive(Deserialize, Serialize, Debug, Clone)]
+pub struct AuditLogLine {
+    pub timestamp: String,
+    pub level: String,
+    pub fields: AuditFields,
+    /// Span information injected by tracing layers.
+    #[serde(default)]
+    pub span: Option<serde_json::Value>,
+    /// Span list injected by tracing layers.
+    #[serde(default)]
+    pub spans: Option<serde_json::Value>,
+}
+
+/// The request audit fields — canonical schema.
+///
+/// Every field the proxy emits in `tracing::info!(target = "audit", ...)`
+/// must appear here.  `deny_unknown_fields` will cause a hard parse error
+/// if the proxy starts emitting a field that isn't listed, forcing you to
+/// update this struct (and all downstream consumers) in one place.
+#[derive(Deserialize, Serialize, Debug, Clone)]
+#[serde(deny_unknown_fields)]
+pub struct AuditFields {
+    /// The literal "request" message from `tracing::info!("request")`.
+    #[serde(default = "default_dash")]
+    pub message: String,
+    /// The tracing target, always "audit" for request logs.
+    #[serde(default = "default_dash")]
+    pub target: String,
+
+    // --- request identity ---
+    #[serde(default)]
+    pub request_id: String,
+    pub method: String,
+    pub host: String,
+    pub path: String,
+    #[serde(default)]
+    pub query: String,
+    pub client_ip: String,
+
+    // --- response ---
+    #[serde(deserialize_with = "flexible_u16")]
+    pub status: u16,
+    #[serde(deserialize_with = "flexible_u64")]
+    pub duration_ms: u64,
+    #[serde(default, deserialize_with = "flexible_u64_default")]
+    pub content_length: u64,
+    #[serde(default, deserialize_with = "flexible_u64_default")]
+    pub response_bytes: u64,
+
+    // --- headers ---
+    #[serde(default = "default_dash")]
+    pub user_agent: String,
+    #[serde(default = "default_dash")]
+    pub referer: String,
+    #[serde(default = "default_dash")]
+    pub accept_language: String,
+    #[serde(default = "default_dash")]
+    pub accept: String,
+    #[serde(default = "default_dash")]
+    pub accept_encoding: String,
+    #[serde(default)]
+    pub has_cookies: bool,
+    #[serde(default = "default_dash")]
+    pub connection: String,
+
+    // --- infra ---
+    #[serde(default = "default_dash")]
+    pub cf_country: String,
+    #[serde(default)]
+    pub backend: String,
+    #[serde(default)]
+    pub error: String,
+    #[serde(default = "default_dash")]
+    pub http_version: String,
+    #[serde(default)]
+    pub header_count: u16,
+
+    // --- training only (not emitted by proxy, but present in external datasets) ---
+    /// Ground-truth label injected by external dataset parsers.
+    /// Values: "attack", "normal".
+    #[serde(default)]
+    pub label: Option<String>,
+}
+
+impl AuditLogLine {
+    /// Try to parse a JSON line as an audit log entry.
+    ///
+    /// - Returns `Ok(Some(entry))` for valid audit log lines.
+    /// - Returns `Ok(None)` for non-audit lines (TLS errors, etc.).
+    /// - Returns `Err` if the line IS an audit log but has unknown fields
+    ///   (schema drift that needs fixing).
+    pub fn try_parse(line: &str) -> Result<Option<Self>, String> {
+        // Quick probe: is this an audit-target line?
+        let probe: Probe = match serde_json::from_str(line) {
+            Ok(p) => p,
+            Err(_) => return Ok(None), // not valid JSON or missing fields
+        };
+        let is_audit = probe
+            .fields
+            .as_ref()
+            .and_then(|f| f.target.as_deref())
+            .map(|t| t == "audit")
+            .unwrap_or(false);
+
+        if !is_audit {
+            return Ok(None);
+        }
+
+        // Full parse with deny_unknown_fields — will error on schema drift.
+        match serde_json::from_str::<Self>(line) {
+            Ok(entry) => Ok(Some(entry)),
+            Err(e) => Err(format!(
+                "audit log schema mismatch (update src/audit.rs): {e}"
+            )),
+        }
+    }
+}
+
+impl Default for AuditFields {
+    fn default() -> Self {
+        Self {
+            message: "request".to_string(),
+            target: "audit".to_string(),
+            request_id: String::new(),
+            method: String::new(),
+            host: String::new(),
+            path: String::new(),
+            query: String::new(),
+            client_ip: String::new(),
+            status: 0,
+            duration_ms: 0,
+            content_length: 0,
+            response_bytes: 0,
+            user_agent: "-".to_string(),
+            referer: "-".to_string(),
+            accept_language: "-".to_string(),
+            accept: "-".to_string(),
+            accept_encoding: "-".to_string(),
+            has_cookies: false,
+            connection: "-".to_string(),
+            cf_country: "-".to_string(),
+            backend: String::new(),
+            error: String::new(),
+            http_version: "-".to_string(),
+            header_count: 0,
+            label: None,
+        }
+    }
+}
+
+fn default_dash() -> String {
+    "-".to_string()
+}
+
+pub fn flexible_u64<'de, D: serde::Deserializer<'de>>(
+    deserializer: D,
+) -> std::result::Result<u64, D::Error> {
+    #[derive(Deserialize)]
+    #[serde(untagged)]
+    enum StringOrNum {
+        Num(u64),
+        Str(String),
+    }
+    match StringOrNum::deserialize(deserializer)? {
+        StringOrNum::Num(n) => Ok(n),
+        StringOrNum::Str(s) => s.parse().map_err(serde::de::Error::custom),
+    }
+}
+
+fn flexible_u64_default<'de, D: serde::Deserializer<'de>>(
+    deserializer: D,
+) -> std::result::Result<u64, D::Error> {
+    #[derive(Deserialize)]
+    #[serde(untagged)]
+    enum Val {
+        Num(u64),
+        Str(String),
+    }
+    match Val::deserialize(deserializer) {
+        Ok(Val::Num(n)) => Ok(n),
+        Ok(Val::Str(s)) => Ok(s.parse().unwrap_or(0)),
+        Err(_) => Ok(0),
+    }
+}
+
+pub fn flexible_u16<'de, D: serde::Deserializer<'de>>(
+    deserializer: D,
+) -> std::result::Result<u16, D::Error> {
+    #[derive(Deserialize)]
+    #[serde(untagged)]
+    enum StringOrNum {
+        Num(u16),
+        Str(String),
+    }
+    match StringOrNum::deserialize(deserializer)? {
+        StringOrNum::Num(n) => Ok(n),
+        StringOrNum::Str(s) => s.parse().map_err(serde::de::Error::custom),
+    }
+}
+
+/// Strip the port suffix from a socket address string.
+pub fn strip_port(addr: &str) -> &str {
+    if addr.starts_with('[') {
+        addr.find(']').map(|i| &addr[1..i]).unwrap_or(addr)
+    } else if let Some(pos) = addr.rfind(':') {
+        &addr[..pos]
+    } else {
+        addr
+    }
+}
diff --git a/src/autotune/ddos.rs b/src/autotune/ddos.rs
index 68f74f6..fe538de 100644
--- a/src/autotune/ddos.rs
+++ b/src/autotune/ddos.rs
@@ -1,230 +1,6 @@
-use crate::autotune::optimizer::BayesianOptimizer;
-use crate::autotune::params::{ParamDef, ParamSpace, ParamType};
-use crate::ddos::replay::{ReplayArgs, replay_and_evaluate};
-use crate::ddos::train::{HeuristicThresholds, train_model_from_states, parse_logs};
-use anyhow::{Context, Result};
-use std::io::Write;
-use std::time::Instant;
+// Copyright Sunbeam Studios 2026
+// SPDX-License-Identifier: Apache-2.0
 
-pub struct AutotuneDdosArgs {
-    pub input: String,
-    pub output: String,
-    pub trials: usize,
-    pub beta: f64,
-    pub trial_log: Option<String>,
-}
-
-fn ddos_param_space() -> ParamSpace {
-    ParamSpace::new(vec![
-        ParamDef { name: "k".into(), param_type: ParamType::Integer { min: 1, max: 20 } },
-        ParamDef { name: "threshold".into(), param_type: ParamType::Continuous { min: 0.1, max: 0.95 } },
-        ParamDef { name: "window_secs".into(), param_type: ParamType::Integer { min: 10, max: 300 } },
-        ParamDef { name: "min_events".into(), param_type: ParamType::Integer { min: 3, max: 50 } },
-        ParamDef { name: "request_rate".into(), param_type: ParamType::Continuous { min: 1.0, max: 100.0 } },
-        ParamDef { name: "path_repetition".into(), param_type: ParamType::Continuous { min: 0.3, max: 0.99 } },
-        ParamDef { name: "error_rate".into(), param_type: ParamType::Continuous { min: 0.2, max: 0.95 } },
-        ParamDef { name: "suspicious_path_ratio".into(), param_type: ParamType::Continuous { min: 0.05, max: 0.8 } },
-        ParamDef { name: "no_cookies_threshold".into(), param_type: ParamType::Continuous { min: 0.01, max: 0.3 } },
-        ParamDef { name: "no_cookies_path_count".into(), param_type: ParamType::Continuous { min: 5.0, max: 100.0 } },
-    ])
-}
-
-pub fn run_autotune(args: AutotuneDdosArgs) -> Result<()> {
-    let space = ddos_param_space();
-    let mut optimizer = BayesianOptimizer::new(space);
-
-    let mut trial_log_file = if let Some(ref path) = args.trial_log {
-        Some(std::fs::File::create(path)?)
-    } else {
-        None
-    };
-
-    // Parse logs once upfront
-    eprintln!("Parsing logs from {}...", args.input);
-    let ip_states = parse_logs(&args.input)?;
-    eprintln!("  {} unique IPs", ip_states.len());
-
-    let mut best_objective = f64::NEG_INFINITY;
-    let mut best_model_bytes: Option<Vec<u8>> = None;
-
-    // Create a temporary directory for intermediate models
-    let tmp_dir = tempfile::tempdir().context("creating temp dir")?;
-
-    eprintln!("Starting DDoS autotune: {} trials, beta={}", args.trials, args.beta);
-
-    for trial_num in 1..=args.trials {
-        let params = optimizer.suggest();
-        let k = params[0] as usize;
-        let threshold = params[1];
-        let window_secs = params[2] as u64;
-        let min_events = params[3] as usize;
-        let request_rate = params[4];
-        let path_repetition = params[5];
-        let error_rate = params[6];
-        let suspicious_path_ratio = params[7];
-        let no_cookies_threshold = params[8];
-        let no_cookies_path_count = params[9];
-
-        let heuristics = HeuristicThresholds::new(
-            request_rate,
-            path_repetition,
-            error_rate,
-            suspicious_path_ratio,
-            no_cookies_threshold,
-            no_cookies_path_count,
-            min_events,
-        );
-
-        let start = Instant::now();
-
-        // Train model with these parameters
-        let train_result = match train_model_from_states(
-            &ip_states, &heuristics, k, threshold, window_secs, min_events,
-        ) {
-            Ok(r) => r,
-            Err(e) => {
-                eprintln!("  trial {trial_num}: TRAIN FAILED ({e})");
-                optimizer.observe(params, 0.0, start.elapsed());
-                continue;
-            }
-        };
-
-        // Save temporary model for replay
-        let tmp_model_path = tmp_dir.path().join(format!("trial_{trial_num}.bin"));
-        let encoded = match bincode::serialize(&train_result.model) {
-            Ok(e) => e,
-            Err(e) => {
-                eprintln!("  trial {trial_num}: SERIALIZE FAILED ({e})");
-                optimizer.observe(params, 0.0, start.elapsed());
-                continue;
-            }
-        };
-        if let Err(e) = std::fs::write(&tmp_model_path, &encoded) {
-            eprintln!("  trial {trial_num}: WRITE FAILED ({e})");
-            optimizer.observe(params, 0.0, start.elapsed());
-            continue;
-        }
-
-        // Replay to evaluate
-        let replay_args = ReplayArgs {
-            input: args.input.clone(),
-            model_path: tmp_model_path.to_string_lossy().into_owned(),
-            config_path: None,
-            k,
-            threshold,
-            window_secs,
-            min_events,
-            rate_limit: false,
-        };
-
-        let replay_result = match replay_and_evaluate(&replay_args) {
-            Ok(r) => r,
-            Err(e) => {
-                eprintln!("  trial {trial_num}: REPLAY FAILED ({e})");
-                optimizer.observe(params, 0.0, start.elapsed());
-                continue;
-            }
-        };
-        let duration = start.elapsed();
-
-        // Compute F-beta from replay false-positive analysis
-        let tp = replay_result.true_positive_ips as f64;
-        let fp = replay_result.false_positive_ips as f64;
-        let total_blocked = replay_result.ddos_blocked_ips.len() as f64;
-        let fn_ = if total_blocked > 0.0 { 0.0 } else { 1.0 }; // We don't know true FN without ground truth
-
-        let objective = if tp + fp > 0.0 {
-            let precision = tp / (tp + fp);
-            let recall = if tp + fn_ > 0.0 { tp / (tp + fn_) } else { 0.0 };
-            let b2 = args.beta * args.beta;
-            if precision + recall > 0.0 {
-                (1.0 + b2) * precision * recall / (b2 * precision + recall)
-            } else {
-                0.0
-            }
-        } else {
-            0.0
-        };
-
-        eprintln!(
-            "  trial {trial_num}/{}: fbeta={objective:.4} (k={k}, thr={threshold:.3}, win={window_secs}s, tp={}, fp={}) [{:.1}s]",
-            args.trials,
-            replay_result.true_positive_ips,
-            replay_result.false_positive_ips,
-            duration.as_secs_f64(),
-        );
-
-        // Log trial as JSONL
-        if let Some(ref mut f) = trial_log_file {
-            let trial_json = serde_json::json!({
-                "trial": trial_num,
-                "params": {
-                    "k": k,
-                    "threshold": threshold,
-                    "window_secs": window_secs,
-                    "min_events": min_events,
-                    "request_rate": request_rate,
-                    "path_repetition": path_repetition,
-                    "error_rate": error_rate,
-                    "suspicious_path_ratio": suspicious_path_ratio,
-                    "no_cookies_threshold": no_cookies_threshold,
-                    "no_cookies_path_count": no_cookies_path_count,
-                },
-                "objective": objective,
-                "duration_secs": duration.as_secs_f64(),
-                "true_positive_ips": replay_result.true_positive_ips,
-                "false_positive_ips": replay_result.false_positive_ips,
-                "ddos_blocked": replay_result.ddos_blocked,
-                "allowed": replay_result.allowed,
-                "attack_count": train_result.attack_count,
-                "normal_count": train_result.normal_count,
-            });
-            writeln!(f, "{}", trial_json)?;
-        }
-
-        if objective > best_objective {
-            best_objective = objective;
-            best_model_bytes = Some(encoded);
-        }
-
-        // Clean up temporary model
-        let _ = std::fs::remove_file(&tmp_model_path);
-
-        optimizer.observe(params, objective, duration);
-    }
-
-    // Save best model
-    if let Some(bytes) = best_model_bytes {
-        std::fs::write(&args.output, &bytes)?;
-        eprintln!("\nBest model saved to {}", args.output);
-    }
-
-    // Print summary
-    if let Some(best) = optimizer.best() {
-        eprintln!("\n═══ Autotune Results ═══════════════════════════════════════");
-        eprintln!("  Best trial:        #{}", best.trial_num);
-        eprintln!("  Best F-beta:       {:.4}", best.objective);
-        eprintln!("  Parameters:");
-        for (name, val) in best.param_names.iter().zip(best.params.iter()) {
-            eprintln!("    {:<30} = {:.6}", name, val);
-        }
-        eprintln!("\n  Heuristics TOML snippet:");
-        eprintln!("    request_rate = {:.2}", best.params[4]);
-        eprintln!("    path_repetition = {:.4}", best.params[5]);
-        eprintln!("    error_rate = {:.4}", best.params[6]);
-        eprintln!("    suspicious_path_ratio = {:.4}", best.params[7]);
-        eprintln!("    no_cookies_threshold = {:.4}", best.params[8]);
-        eprintln!("    no_cookies_path_count = {:.1}", best.params[9]);
-        eprintln!("    min_events = {}", best.params[3] as usize);
-        eprintln!("\n  Reproduce:");
-        eprintln!(
-            "    cargo run -- train-ddos --input {} --output {} --k {} --threshold {:.4} --window-secs {} --min-events {} --heuristics <toml>",
-            args.input, args.output,
-            best.params[0] as usize, best.params[1],
-            best.params[2] as u64, best.params[3] as usize,
-        );
-        eprintln!("══════════════════════════════════════════════════════════");
-    }
-
-    Ok(())
-}
+// Legacy KNN autotune removed — ensemble models are tuned via
+// `cargo run --features training -- sweep-cookie-weight` and the
+// training pipeline in src/training/.
diff --git a/src/autotune/scanner.rs b/src/autotune/scanner.rs
index 110c940..82bfb72 100644
--- a/src/autotune/scanner.rs
+++ b/src/autotune/scanner.rs
@@ -1,128 +1,6 @@
-use crate::autotune::optimizer::BayesianOptimizer;
-use crate::autotune::params::{ParamDef, ParamSpace, ParamType};
-use crate::scanner::train::{TrainScannerArgs, train_and_evaluate};
-use anyhow::Result;
-use std::io::Write;
-use std::time::Instant;
+// Copyright Sunbeam Studios 2026
+// SPDX-License-Identifier: Apache-2.0
 
-pub struct AutotuneScannerArgs {
-    pub input: String,
-    pub output: String,
-    pub wordlists: Option<String>,
-    pub csic: bool,
-    pub trials: usize,
-    pub beta: f64,
-    pub trial_log: Option<String>,
-}
-
-fn scanner_param_space() -> ParamSpace {
-    ParamSpace::new(vec![
-        ParamDef { name: "threshold".into(), param_type: ParamType::Continuous { min: 0.1, max: 0.95 } },
-        ParamDef { name: "learning_rate".into(), param_type: ParamType::LogScale { min: 0.001, max: 0.1 } },
-        ParamDef { name: "epochs".into(), param_type: ParamType::Integer { min: 100, max: 5000 } },
-        ParamDef { name: "class_weight_multiplier".into(), param_type: ParamType::Continuous { min: 0.5, max: 5.0 } },
-    ])
-}
-
-pub fn run_autotune(args: AutotuneScannerArgs) -> Result<()> {
-    let space = scanner_param_space();
-    let mut optimizer = BayesianOptimizer::new(space);
-
-    let mut trial_log_file = if let Some(ref path) = args.trial_log {
-        Some(std::fs::File::create(path)?)
-    } else {
-        None
-    };
-
-    let mut best_objective = f64::NEG_INFINITY;
-    let mut best_model_bytes: Option<Vec<u8>> = None;
-
-    eprintln!("Starting scanner autotune: {} trials, beta={}", args.trials, args.beta);
-
-    for trial_num in 1..=args.trials {
-        let params = optimizer.suggest();
-        let threshold = params[0];
-        let learning_rate = params[1];
-        let epochs = params[2] as usize;
-        let class_weight_multiplier = params[3];
-
-        let train_args = TrainScannerArgs {
-            input: args.input.clone(),
-            output: String::new(), // don't save intermediate models
-            wordlists: args.wordlists.clone(),
-            threshold,
-            csic: args.csic,
-        };
-
-        let start = Instant::now();
-        let result = match train_and_evaluate(&train_args, learning_rate, epochs, class_weight_multiplier) {
-            Ok(r) => r,
-            Err(e) => {
-                eprintln!("  trial {trial_num}: FAILED ({e})");
-                optimizer.observe(params, 0.0, start.elapsed());
-                continue;
-            }
-        };
-        let duration = start.elapsed();
-
-        let objective = result.test_metrics.fbeta(args.beta);
-
-        eprintln!(
-            "  trial {trial_num}/{}: fbeta={objective:.4} (threshold={threshold:.3}, lr={learning_rate:.5}, epochs={epochs}, cwm={class_weight_multiplier:.2}) [{:.1}s]",
-            args.trials,
-            duration.as_secs_f64(),
-        );
-
-        // Log trial as JSONL
-        if let Some(ref mut f) = trial_log_file {
-            let trial_json = serde_json::json!({
-                "trial": trial_num,
-                "params": {
-                    "threshold": threshold,
-                    "learning_rate": learning_rate,
-                    "epochs": epochs,
-                    "class_weight_multiplier": class_weight_multiplier,
-                },
-                "objective": objective,
-                "duration_secs": duration.as_secs_f64(),
-                "train_f1": result.train_metrics.f1(),
-                "test_precision": result.test_metrics.precision(),
-                "test_recall": result.test_metrics.recall(),
-            });
-            writeln!(f, "{}", trial_json)?;
-        }
-
-        if objective > best_objective {
-            best_objective = objective;
-            let encoded = bincode::serialize(&result.model)?;
-            best_model_bytes = Some(encoded);
-        }
-
-        optimizer.observe(params, objective, duration);
-    }
-
-    // Save best model
-    if let Some(bytes) = best_model_bytes {
-        std::fs::write(&args.output, &bytes)?;
-        eprintln!("\nBest model saved to {}", args.output);
-    }
-
-    // Print summary
-    if let Some(best) = optimizer.best() {
-        eprintln!("\n═══ Autotune Results ═══════════════════════════════════════");
-        eprintln!("  Best trial:        #{}", best.trial_num);
-        eprintln!("  Best F-beta:       {:.4}", best.objective);
-        eprintln!("  Parameters:");
-        for (name, val) in best.param_names.iter().zip(best.params.iter()) {
-            eprintln!("    {:<30} = {:.6}", name, val);
-        }
-        eprintln!("\n  Reproduce:");
-        eprintln!(
-            "    cargo run -- train-scanner --input {} --output {} --threshold {:.4}",
-            args.input, args.output, best.params[0],
-        );
-        eprintln!("══════════════════════════════════════════════════════════");
-    }
-
-    Ok(())
-}
+// Legacy linear-model autotune removed — ensemble models are tuned via
+// `cargo run --features training -- sweep-cookie-weight` and the
+// training pipeline in src/training/.
diff --git a/src/config.rs b/src/config.rs
index 8e139c6..6f6b0e7 100644
--- a/src/config.rs
+++ b/src/config.rs
@@ -1,3 +1,6 @@
+// Copyright Sunbeam Studios 2026
+// SPDX-License-Identifier: Apache-2.0
+
 use anyhow::{Context, Result};
 use serde::Deserialize;
 use std::fs;
@@ -18,7 +21,7 @@ pub struct Config {
     pub routes: Vec<RouteConfig>,
     /// Optional SSH TCP passthrough (port 22 → Gitea SSH).
     pub ssh: Option<SshConfig>,
-    /// Optional KNN-based DDoS detection.
+    /// Optional DDoS detection (ensemble: decision tree + MLP).
     pub ddos: Option<DDoSConfig>,
     /// Optional per-identity rate limiting.
     pub rate_limit: Option<RateLimitConfig>,
@@ -60,10 +63,6 @@ fn default_config_configmap() -> String { "pingora-config".to_string() }
 
 #[derive(Debug, Deserialize, Clone)]
 pub struct DDoSConfig {
-    #[serde(default)]
-    pub model_path: Option<String>,
-    #[serde(default = "default_k")]
-    pub k: usize,
     #[serde(default = "default_threshold")]
     pub threshold: f64,
     #[serde(default = "default_window_secs")]
@@ -74,8 +73,10 @@ pub struct DDoSConfig {
     pub min_events: usize,
     #[serde(default = "default_enabled")]
     pub enabled: bool,
-    #[serde(default = "default_use_ensemble")]
-    pub use_ensemble: bool,
+    /// When true, run the model and log decisions but never block traffic.
+    /// Useful for gathering data on model accuracy before enforcing.
+    #[serde(default)]
+    pub observe_only: bool,
 }
 
 #[derive(Debug, Deserialize, Clone)]
@@ -100,23 +101,20 @@ pub struct BucketConfig {
 
 #[derive(Debug, Deserialize, Clone)]
 pub struct ScannerConfig {
-    #[serde(default)]
-    pub model_path: Option<String>,
     #[serde(default = "default_scanner_threshold")]
     pub threshold: f64,
     #[serde(default = "default_scanner_enabled")]
     pub enabled: bool,
-    /// How often (seconds) to check the model file for changes. 0 = no hot-reload.
-    #[serde(default = "default_scanner_poll_interval")]
-    pub poll_interval_secs: u64,
     /// Bot allowlist rules. Verified bots bypass the scanner model.
     #[serde(default)]
     pub allowlist: Vec<BotAllowlistRule>,
     /// TTL (seconds) for verified bot IP cache entries.
     #[serde(default = "default_bot_cache_ttl")]
     pub bot_cache_ttl_secs: u64,
-    #[serde(default = "default_use_ensemble")]
-    pub use_ensemble: bool,
+    /// When true, run the model and log decisions but never block traffic.
+    /// Useful for gathering data on model accuracy before enforcing.
+    #[serde(default)]
+    pub observe_only: bool,
 }
 
 #[derive(Debug, Deserialize, Clone)]
@@ -136,17 +134,14 @@ pub struct BotAllowlistRule {
 }
 
 fn default_bot_cache_ttl() -> u64 { 86400 } // 24h
-fn default_use_ensemble() -> bool { true }
 
 fn default_scanner_threshold() -> f64 { 0.5 }
 fn default_scanner_enabled() -> bool { true }
-fn default_scanner_poll_interval() -> u64 { 30 }
 
 fn default_rl_enabled() -> bool { true }
 fn default_eviction_interval() -> u64 { 300 }
 fn default_stale_after() -> u64 { 600 }
 
-fn default_k() -> usize { 5 }
 fn default_threshold() -> f64 { 0.6 }
 fn default_window_secs() -> u64 { 60 }
 fn default_window_capacity() -> usize { 1000 }
diff --git a/src/dataset/cicids.rs b/src/dataset/cicids.rs
index 12abc47..d619e24 100644
--- a/src/dataset/cicids.rs
+++ b/src/dataset/cicids.rs
@@ -1,3 +1,6 @@
+// Copyright Sunbeam Studios 2026
+// SPDX-License-Identifier: Apache-2.0
+
 //! CIC-IDS2017 timing profile extractor.
 //!
 //! Parses CIC-IDS2017 CSV files and extracts statistical timing profiles
@@ -218,6 +221,262 @@ fn parse_csv_file(
     Ok(())
 }
 
+/// Convert CIC-IDS2017 flow records directly into DDoS training samples.
+///
+/// Maps network-layer flow features to our 14-dimensional HTTP-layer feature vector.
+/// Non-BENIGN labels → attack (1.0), BENIGN → normal (0.0).
+/// Uses a deterministic RNG seeded per-row to fill HTTP-only features (cookies, etc.)
+/// that don't exist in the network-layer data.
+pub fn extract_ddos_samples(csv_dir: &Path) -> Result<Vec<crate::dataset::sample::TrainingSample>> {
+    use crate::dataset::sample::TrainingSample;
+    use rand::prelude::*;
+    use rand::rngs::StdRng;
+
+    let entries: Vec<std::path::PathBuf> = if csv_dir.is_file() {
+        vec![csv_dir.to_path_buf()]
+    } else {
+        let mut files: Vec<std::path::PathBuf> = std::fs::read_dir(csv_dir)
+            .with_context(|| format!("reading directory {}", csv_dir.display()))?
+            .filter_map(|e| e.ok())
+            .map(|e| e.path())
+            .filter(|p| {
+                p.extension()
+                    .map(|e| e.to_ascii_lowercase() == "csv")
+                    .unwrap_or(false)
+            })
+            .collect();
+        files.sort();
+        files
+    };
+
+    if entries.is_empty() {
+        anyhow::bail!("no CSV files found in {}", csv_dir.display());
+    }
+
+    let mut samples = Vec::new();
+    let mut rng = StdRng::seed_from_u64(0xC1C1D5);
+
+    for csv_path in &entries {
+        let file_samples = extract_ddos_samples_from_csv(csv_path, &mut rng)
+            .with_context(|| format!("extracting DDoS samples from {}", csv_path.display()))?;
+        let filename = csv_path.file_name().unwrap_or_default().to_string_lossy();
+        let attack_count = file_samples.iter().filter(|s| s.label > 0.5).count();
+        let normal_count = file_samples.len() - attack_count;
+        eprintln!(
+            "  {}: {} samples ({} attack, {} normal)",
+            filename,
+            file_samples.len(),
+            attack_count,
+            normal_count
+        );
+        samples.extend(file_samples);
+    }
+
+    // Subsample if too large — cap at 500K to keep training tractable.
+    let max_samples = 500_000;
+    if samples.len() > max_samples {
+        // Stratified subsample: keep attack ratio balanced.
+        let mut attacks: Vec<TrainingSample> =
+            samples.iter().filter(|s| s.label > 0.5).cloned().collect();
+        let mut normals: Vec<TrainingSample> =
+            samples.iter().filter(|s| s.label <= 0.5).cloned().collect();
+
+        // Shuffle both
+        attacks.shuffle(&mut rng);
+        normals.shuffle(&mut rng);
+
+        // Take equal parts, favoring attacks if underrepresented
+        let attack_cap = max_samples / 2;
+        let normal_cap = max_samples - attack_cap.min(attacks.len());
+        attacks.truncate(attack_cap);
+        normals.truncate(normal_cap);
+
+        eprintln!(
+            "  subsampled to {} ({} attack, {} normal)",
+            attacks.len() + normals.len(),
+            attacks.len(),
+            normals.len()
+        );
+        samples = attacks;
+        samples.extend(normals);
+        samples.shuffle(&mut rng);
+    }
+
+    Ok(samples)
+}
+
+fn extract_ddos_samples_from_csv(
+    path: &Path,
+    rng: &mut rand::rngs::StdRng,
+) -> Result<Vec<crate::dataset::sample::TrainingSample>> {
+    use crate::dataset::sample::{DataSource, TrainingSample};
+    use crate::ddos::features::NUM_FEATURES;
+    use rand::prelude::*;
+
+    let mut rdr = csv::ReaderBuilder::new()
+        .flexible(true)
+        .trim(csv::Trim::All)
+        .from_path(path)?;
+
+    let headers: Vec<String> = rdr.headers()?.iter().map(|h| h.to_string()).collect();
+
+    let col_label = find_column(&headers, "Label")
+        .with_context(|| format!("missing 'Label' in {}", path.display()))?;
+    let col_flow_duration = find_column(&headers, "Flow Duration");
+    let col_total_fwd_pkts = find_column(&headers, "Total Fwd Packets");
+    let col_total_bwd_pkts = find_column(&headers, "Total Backward Packets");
+    let col_flow_pkts_s = find_column(&headers, "Flow Packets/s");
+    let col_flow_iat_mean = find_column(&headers, "Flow IAT Mean");
+    let col_avg_pkt_size = find_column(&headers, "Average Packet Size");
+    let col_pkt_len_std = find_column(&headers, "Packet Length Std");
+    let col_syn_flag = find_column(&headers, "SYN Flag Count");
+
+    let mut samples = Vec::new();
+
+    for result in rdr.records() {
+        let record = match result {
+            Ok(r) => r,
+            Err(_) => continue,
+        };
+
+        let label_str = match record.get(col_label) {
+            Some(l) => l.trim().to_string(),
+            None => continue,
+        };
+        if label_str.is_empty() {
+            continue;
+        }
+
+        let is_attack = label_str != "BENIGN";
+        let label_f32 = if is_attack { 1.0f32 } else { 0.0f32 };
+
+        // Parse numeric columns (0.0 fallback for missing/malformed).
+        let get_f64 = |col: Option<usize>| -> f64 {
+            col.and_then(|c| record.get(c))
+                .and_then(|v| v.trim().parse::<f64>().ok())
+                .unwrap_or(0.0)
+        };
+
+        let flow_duration_us = get_f64(col_flow_duration); // microseconds
+        let total_fwd_pkts = get_f64(col_total_fwd_pkts);
+        let total_bwd_pkts = get_f64(col_total_bwd_pkts);
+        let flow_pkts_s = get_f64(col_flow_pkts_s);
+        let flow_iat_mean = get_f64(col_flow_iat_mean); // microseconds
+        let avg_pkt_size = get_f64(col_avg_pkt_size);
+        let pkt_len_std = get_f64(col_pkt_len_std);
+        let syn_flag = get_f64(col_syn_flag);
+
+        let flow_duration_s = (flow_duration_us / 1_000_000.0).max(0.001);
+        let total_pkts = total_fwd_pkts + total_bwd_pkts;
+
+        // Map to our 14 HTTP-layer DDoS features:
+        let mut features = vec![0.0f32; NUM_FEATURES];
+
+        // 0: request_rate — packets/sec as proxy for requests/sec
+        features[0] = flow_pkts_s.max(0.0).min(10000.0) as f32;
+
+        // 1: unique_paths — approximate from packet diversity (std/mean ratio)
+        let diversity = if avg_pkt_size > 0.0 {
+            (pkt_len_std / avg_pkt_size).min(10.0)
+        } else {
+            0.0
+        };
+        features[1] = (diversity * 5.0 + 1.0) as f32;
+
+        // 2: unique_hosts — infer from port (attack traffic often targets one host)
+        features[2] = if is_attack { 1.0 } else { rng.random_range(1.0..5.0) as f32 };
+
+        // 3: error_rate — SYN-heavy flows suggest connection errors
+        let error_signal = if total_pkts > 0.0 {
+            (syn_flag / total_pkts.max(1.0)).min(1.0)
+        } else {
+            0.0
+        };
+        features[3] = if is_attack {
+            (error_signal + rng.random_range(0.1..0.5)).min(1.0) as f32
+        } else {
+            (error_signal * 0.3) as f32
+        };
+
+        // 4: avg_duration_ms — flow duration / total packets, in ms
+        features[4] = if total_pkts > 0.0 {
+            ((flow_duration_s * 1000.0) / total_pkts).min(5000.0) as f32
+        } else {
+            0.0
+        };
+
+        // 5: method_entropy — low for attacks (single method), moderate for normal
+        features[5] = if is_attack {
+            rng.random_range(0.0..0.3) as f32
+        } else {
+            rng.random_range(0.2..1.5) as f32
+        };
+
+        // 6: burst_score — inverse of inter-arrival time
+        let iat_s = (flow_iat_mean / 1_000_000.0).max(0.001);
+        features[6] = (1.0 / iat_s).min(500.0) as f32;
+
+        // 7: path_repetition — attacks repeat paths heavily
+        features[7] = if is_attack {
+            rng.random_range(0.6..1.0) as f32
+        } else {
+            rng.random_range(0.05..0.4) as f32
+        };
+
+        // 8: avg_content_length — from average packet size
+        features[8] = avg_pkt_size.max(0.0).min(10000.0) as f32;
+
+        // 9: unique_user_agents — low for attacks
+        features[9] = if is_attack {
+            rng.random_range(1.0..2.0) as f32
+        } else {
+            rng.random_range(1.0..4.0) as f32
+        };
+
+        // 10: cookie_ratio — bots don't send cookies
+        features[10] = if is_attack {
+            rng.random_range(0.0..0.1) as f32
+        } else {
+            rng.random_range(0.6..1.0) as f32
+        };
+
+        // 11: referer_ratio — bots rarely send referer
+        features[11] = if is_attack {
+            rng.random_range(0.0..0.1) as f32
+        } else {
+            rng.random_range(0.3..1.0) as f32
+        };
+
+        // 12: accept_language_ratio — bots don't send this
+        features[12] = if is_attack {
+            rng.random_range(0.0..0.1) as f32
+        } else {
+            rng.random_range(0.6..1.0) as f32
+        };
+
+        // 13: suspicious_path_ratio — attacks may probe paths
+        let is_web_attack = label_str.contains("Web Attack")
+            || label_str.contains("Bot")
+            || label_str.contains("Infiltration");
+        features[13] = if is_web_attack {
+            rng.random_range(0.2..0.7) as f32
+        } else if is_attack {
+            rng.random_range(0.0..0.3) as f32
+        } else {
+            rng.random_range(0.0..0.05) as f32
+        };
+
+        samples.push(TrainingSample {
+            features,
+            label: label_f32,
+            source: DataSource::SyntheticCicTiming,
+            weight: 0.7, // higher than pure synthetic (0.5), lower than prod (1.0)
+        });
+    }
+
+    Ok(samples)
+}
+
 /// Parse timing profiles from an in-memory CSV string (useful for tests).
 pub fn extract_timing_profiles_from_str(csv_content: &str) -> Result<Vec<TimingProfile>> {
     let dir = tempfile::tempdir()?;
diff --git a/src/dataset/download.rs b/src/dataset/download.rs
index c36b8cb..fa753e2 100644
--- a/src/dataset/download.rs
+++ b/src/dataset/download.rs
@@ -1,3 +1,6 @@
+// Copyright Sunbeam Studios 2026
+// SPDX-License-Identifier: Apache-2.0
+
 //! Download and cache upstream datasets for training.
 //!
 //! Cached under `~/.cache/sunbeam/<dataset>/`.  Files are only downloaded
@@ -19,8 +22,17 @@ fn cache_base() -> PathBuf {
 
 // --- CIC-IDS2017 ---
 
-/// Only the Friday DDoS file — contains DDoS Hulk, Slowloris, slowhttptest, GoldenEye.
-const CICIDS_FILE: &str = "Friday-WorkingHours-Afternoon-DDos.pcap_ISCX.csv";
+/// All CIC-IDS2017 CSV files — covers every attack day and normal baselines.
+const CICIDS_FILES: &[&str] = &[
+    "Monday-WorkingHours.pcap_ISCX.csv",
+    "Tuesday-WorkingHours.pcap_ISCX.csv",
+    "Wednesday-workingHours.pcap_ISCX.csv",
+    "Thursday-WorkingHours-Morning-WebAttacks.pcap_ISCX.csv",
+    "Thursday-WorkingHours-Afternoon-Infilteration.pcap_ISCX.csv",
+    "Friday-WorkingHours-Morning.pcap_ISCX.csv",
+    "Friday-WorkingHours-Afternoon-PortScan.pcap_ISCX.csv",
+    "Friday-WorkingHours-Afternoon-DDos.pcap_ISCX.csv",
+];
 
 /// Hugging Face mirror (public, no auth required).
 const CICIDS_BASE_URL: &str =
@@ -30,49 +42,56 @@ fn cicids_cache_dir() -> PathBuf {
     cache_base().join("cicids")
 }
 
-/// Return the path to the cached CIC-IDS2017 DDoS CSV, or `None` if not downloaded.
+/// Return the cache directory if ALL CIC-IDS2017 CSVs are downloaded, else `None`.
 pub fn cicids_cached_path() -> Option<PathBuf> {
-    let path = cicids_cache_dir().join(CICIDS_FILE);
-    if path.exists() {
-        Some(path)
+    let dir = cicids_cache_dir();
+    if CICIDS_FILES.iter().all(|f| dir.join(f).exists()) {
+        Some(dir)
     } else {
         None
     }
 }
 
-/// Download the CIC-IDS2017 Friday DDoS CSV to cache.  Returns the cached path.
+/// Download all CIC-IDS2017 CSV files to cache. Returns the cache directory.
 pub fn download_cicids() -> Result<PathBuf> {
     let dir = cicids_cache_dir();
-    let path = dir.join(CICIDS_FILE);
-
-    if path.exists() {
-        eprintln!("  cached: {}", path.display());
-        return Ok(path);
-    }
-
-    let url = format!("{CICIDS_BASE_URL}/{CICIDS_FILE}");
-    eprintln!("  downloading: {url}");
-    eprintln!("  (this is ~170 MB, may take a minute)");
-
     std::fs::create_dir_all(&dir)?;
 
-    // Stream to file to avoid holding 170MB in memory.
-    let resp = reqwest::blocking::Client::builder()
+    let client = reqwest::blocking::Client::builder()
         .timeout(std::time::Duration::from_secs(600))
-        .build()?
-        .get(&url)
-        .send()
-        .with_context(|| format!("fetching {url}"))?
-        .error_for_status()
-        .with_context(|| format!("HTTP error for {url}"))?;
+        .build()?;
 
-    let mut file = std::fs::File::create(&path)
-        .with_context(|| format!("creating {}", path.display()))?;
-    let bytes = resp.bytes().with_context(|| "reading response body")?;
-    std::io::Write::write_all(&mut file, &bytes)?;
+    for (i, filename) in CICIDS_FILES.iter().enumerate() {
+        let path = dir.join(filename);
+        if path.exists() {
+            eprintln!("  [{}/{}] cached: {}", i + 1, CICIDS_FILES.len(), filename);
+            continue;
+        }
 
-    eprintln!("  saved: {}", path.display());
-    Ok(path)
+        let url = format!("{CICIDS_BASE_URL}/{filename}");
+        eprintln!(
+            "  [{}/{}] downloading: {}",
+            i + 1,
+            CICIDS_FILES.len(),
+            filename
+        );
+
+        let resp = client
+            .get(&url)
+            .send()
+            .with_context(|| format!("fetching {url}"))?
+            .error_for_status()
+            .with_context(|| format!("HTTP error for {url}"))?;
+
+        let mut file = std::fs::File::create(&path)
+            .with_context(|| format!("creating {}", path.display()))?;
+        let bytes = resp.bytes().with_context(|| "reading response body")?;
+        std::io::Write::write_all(&mut file, &bytes)?;
+
+        eprintln!("    saved: {}", path.display());
+    }
+
+    Ok(dir)
 }
 
 // --- CSIC 2010 ---
@@ -80,7 +99,10 @@ pub fn download_cicids() -> Result<PathBuf> {
 /// Download CSIC 2010 dataset files to cache (delegates to scanner::csic).
 pub fn download_csic() -> Result<()> {
     if crate::scanner::csic::csic_is_cached() {
-        eprintln!("  cached: {}", crate::scanner::csic::csic_cache_path().display());
+        eprintln!(
+            "  cached: {}",
+            crate::scanner::csic::csic_cache_path().display()
+        );
         return Ok(());
     }
     // fetch_csic_dataset downloads, caches, and parses — we only need the download side-effect.
@@ -96,9 +118,9 @@ pub fn download_all() -> Result<()> {
     download_csic()?;
     eprintln!();
 
-    eprintln!("[2/2] CIC-IDS2017 DDoS timing profiles");
+    eprintln!("[2/2] CIC-IDS2017 (all attack days + normal baselines)");
     let path = download_cicids()?;
-    eprintln!("  ok: {}\n", path.display());
+    eprintln!("  ok: {} ({} files)\n", path.display(), CICIDS_FILES.len());
 
     eprintln!("all datasets cached.");
     Ok(())
@@ -116,4 +138,10 @@ mod tests {
         let cicids = cicids_cache_dir();
         assert!(cicids.to_str().unwrap().contains("cicids"));
     }
+
+    #[test]
+    fn test_all_files_listed() {
+        assert_eq!(CICIDS_FILES.len(), 8);
+        assert!(CICIDS_FILES.iter().all(|f| f.ends_with(".csv")));
+    }
 }
diff --git a/src/dataset/modsec.rs b/src/dataset/modsec.rs
index a35eb23..5b4443e 100644
--- a/src/dataset/modsec.rs
+++ b/src/dataset/modsec.rs
@@ -1,3 +1,6 @@
+// Copyright Sunbeam Studios 2026
+// SPDX-License-Identifier: Apache-2.0
+
 //! Parser for OWASP ModSecurity audit log files (Serial / concurrent format).
 //!
 //! ModSecurity audit logs consist of multi-section entries delimited by boundary
@@ -176,6 +179,7 @@ fn transaction_to_audit_fields(
     let content_length: u64 = get_header("content-length")
         .and_then(|v| v.parse().ok())
         .unwrap_or(0);
+    let accept = get_header("accept").filter(|a| a != "-" && !a.is_empty());
 
     // Section F: response status
     let status = sections
@@ -216,11 +220,13 @@ fn transaction_to_audit_fields(
         duration_ms: 0,
         content_length,
         user_agent,
-        has_cookies: Some(has_cookies),
-        referer,
-        accept_language,
+        has_cookies,
+        referer: referer.unwrap_or_else(|| "-".to_string()),
+        accept_language: accept_language.unwrap_or_else(|| "-".to_string()),
+        accept: accept.unwrap_or_else(|| "-".to_string()),
         backend: "-".to_string(),
         label: Some(label.clone()),
+        ..AuditFields::default()
     };
 
     Some((fields, label))
@@ -302,7 +308,7 @@ Content-Type: text/html
         assert_eq!(attack_fields.client_ip, "192.168.1.100");
         assert_eq!(attack_fields.user_agent, "curl/7.68.0");
         assert_eq!(attack_fields.status, 403);
-        assert!(!attack_fields.has_cookies.unwrap_or(true));
+        assert!(!attack_fields.has_cookies);
 
         // Second entry: normal (no rule match).
         let (normal_fields, normal_label) = &results[1];
@@ -311,9 +317,9 @@ Content-Type: text/html
         assert_eq!(normal_fields.path, "/index.html");
         assert_eq!(normal_fields.client_ip, "10.0.0.50");
         assert_eq!(normal_fields.status, 200);
-        assert!(normal_fields.has_cookies.unwrap_or(false));
-        assert!(normal_fields.referer.is_some());
-        assert!(normal_fields.accept_language.is_some());
+        assert!(normal_fields.has_cookies);
+        assert!(normal_fields.referer != "-");
+        assert!(normal_fields.accept_language != "-");
     }
 
     #[test]
diff --git a/src/dataset/prepare.rs b/src/dataset/prepare.rs
index c90d1cd..9ddbabf 100644
--- a/src/dataset/prepare.rs
+++ b/src/dataset/prepare.rs
@@ -1,3 +1,6 @@
+// Copyright Sunbeam Studios 2026
+// SPDX-License-Identifier: Apache-2.0
+
 //! Dataset preparation orchestrator.
 //!
 //! Combines production logs, external datasets (CSIC, OWASP ModSec), and
@@ -30,6 +33,10 @@ pub struct PrepareDatasetArgs {
     pub seed: u64,
     /// Path to heuristics.toml for auto-labeling production logs.
     pub heuristics: Option<String>,
+    /// Inject CSIC 2010 entries as labeled audit logs into production stream.
+    pub inject_csic: bool,
+    /// Inject OWASP ModSec entries as labeled audit logs (path to .log file).
+    pub inject_modsec: Option<String>,
 }
 
 impl Default for PrepareDatasetArgs {
@@ -41,6 +48,8 @@ impl Default for PrepareDatasetArgs {
             output: "dataset.bin".to_string(),
             seed: 42,
             heuristics: None,
+            inject_csic: false,
+            inject_modsec: None,
         }
     }
 }
@@ -71,18 +80,21 @@ pub fn run(args: PrepareDatasetArgs) -> Result<()> {
     scanner_samples.extend(prod_scanner);
     ddos_samples.extend(prod_ddos);
 
-    // --- 2. CSIC 2010 (scanner) ---
-    eprintln!("fetching CSIC 2010 dataset...");
-    let csic_entries = crate::scanner::csic::fetch_csic_dataset()?;
-    let csic_samples = entries_to_scanner_samples(&csic_entries, DataSource::Csic2010, 0.8)?;
-    eprintln!("  CSIC: {} scanner samples", csic_samples.len());
-    scanner_samples.extend(csic_samples);
+    // --- 2. Inject external datasets as labeled audit log entries ---
+    // These go through the same feature extraction as production logs,
+    // with ground-truth labels (no heuristic labeling needed).
+    if args.inject_csic {
+        eprintln!("injecting CSIC 2010 as labeled audit entries...");
+        let csic_entries = crate::scanner::csic::fetch_csic_dataset()?;
+        let csic_scanner = entries_to_scanner_samples(&csic_entries, DataSource::Csic2010, 0.8)?;
+        eprintln!("  CSIC injected: {} scanner samples", csic_scanner.len());
+        scanner_samples.extend(csic_scanner);
+    }
 
-    // --- 3. OWASP ModSec (scanner) ---
-    if let Some(owasp_path) = &args.owasp {
-        eprintln!("parsing OWASP ModSec audit log from {owasp_path}...");
+    if let Some(modsec_path) = &args.inject_modsec {
+        eprintln!("injecting ModSec audit log from {modsec_path}...");
         let modsec_entries =
-            crate::dataset::modsec::parse_modsec_audit_log(Path::new(owasp_path))?;
+            crate::dataset::modsec::parse_modsec_audit_log(Path::new(modsec_path))?;
         let entries_with_host: Vec<(AuditFields, String)> = modsec_entries
             .into_iter()
             .map(|(fields, _label)| {
@@ -90,16 +102,49 @@ pub fn run(args: PrepareDatasetArgs) -> Result<()> {
                 (fields, host_prefix)
             })
             .collect();
-        let modsec_samples =
+        let modsec_scanner =
             entries_to_scanner_samples(&entries_with_host, DataSource::OwaspModSec, 0.8)?;
-        eprintln!("  OWASP: {} scanner samples", modsec_samples.len());
-        scanner_samples.extend(modsec_samples);
+        eprintln!("  ModSec injected: {} scanner samples", modsec_scanner.len());
+        scanner_samples.extend(modsec_scanner);
     }
 
-    // --- 4. CIC-IDS2017 timing profiles (from cache if downloaded) ---
+    // --- 3. Legacy OWASP path (kept for backwards compat) ---
+    if let Some(owasp_path) = &args.owasp {
+        if args.inject_modsec.as_deref() != Some(owasp_path.as_str()) {
+            eprintln!("parsing OWASP ModSec audit log from {owasp_path}...");
+            let modsec_entries =
+                crate::dataset::modsec::parse_modsec_audit_log(Path::new(owasp_path))?;
+            let entries_with_host: Vec<(AuditFields, String)> = modsec_entries
+                .into_iter()
+                .map(|(fields, _label)| {
+                    let host_prefix = fields.host.split('.').next().unwrap_or("").to_string();
+                    (fields, host_prefix)
+                })
+                .collect();
+            let modsec_samples =
+                entries_to_scanner_samples(&entries_with_host, DataSource::OwaspModSec, 0.8)?;
+            eprintln!("  OWASP: {} scanner samples", modsec_samples.len());
+            scanner_samples.extend(modsec_samples);
+        }
+    }
+
+    // --- 4. CIC-IDS2017 (direct DDoS samples + timing profiles for synthetic) ---
     let cicids_profiles = if let Some(cached_path) = crate::dataset::download::cicids_cached_path()
     {
-        eprintln!("extracting CIC-IDS2017 timing profiles from cache...");
+        // Direct conversion: CIC-IDS2017 flows → DDoS training samples
+        eprintln!("extracting CIC-IDS2017 DDoS samples from cache...");
+        let cicids_ddos = crate::dataset::cicids::extract_ddos_samples(&cached_path)?;
+        let attack_count = cicids_ddos.iter().filter(|s| s.label > 0.5).count();
+        eprintln!(
+            "  CIC-IDS2017 direct: {} DDoS samples ({} attack, {} normal)",
+            cicids_ddos.len(),
+            attack_count,
+            cicids_ddos.len() - attack_count
+        );
+        ddos_samples.extend(cicids_ddos);
+
+        // Also extract timing profiles for synthetic generation
+        eprintln!("extracting CIC-IDS2017 timing profiles...");
         let profiles = crate::dataset::cicids::extract_timing_profiles(&cached_path)?;
         eprintln!("  extracted {} attack-type profiles", profiles.len());
         profiles
@@ -112,10 +157,10 @@ pub fn run(args: PrepareDatasetArgs) -> Result<()> {
     // --- 5. Synthetic data (both models, always generated) ---
     eprintln!("generating synthetic samples...");
     let config = crate::dataset::synthetic::SyntheticConfig {
-        num_ddos_attack: 10000,
-        num_ddos_normal: 10000,
-        num_scanner_attack: 5000,
-        num_scanner_normal: 5000,
+        num_ddos_attack: 50000,
+        num_ddos_normal: 50000,
+        num_scanner_attack: 25000,
+        num_scanner_normal: 25000,
         seed: args.seed,
     };
 
@@ -240,17 +285,9 @@ fn parse_production_logs(
 
     // --- Scanner samples from production logs ---
     for (fields, host_prefix) in &parsed_entries {
-        let has_cookies = fields.has_cookies.unwrap_or(false);
-        let has_referer = fields
-            .referer
-            .as_ref()
-            .map(|r| r != "-" && !r.is_empty())
-            .unwrap_or(false);
-        let has_accept_language = fields
-            .accept_language
-            .as_ref()
-            .map(|a| a != "-" && !a.is_empty())
-            .unwrap_or(false);
+        let has_cookies = fields.has_cookies;
+        let has_referer = !fields.referer.is_empty() && fields.referer != "-";
+        let has_accept_language = !fields.accept_language.is_empty() && fields.accept_language != "-";
 
         let feats = features::extract_features(
             &fields.method,
@@ -259,7 +296,7 @@ fn parse_production_logs(
             has_cookies,
             has_referer,
             has_accept_language,
-            "-",
+            &fields.accept,
             &fields.user_agent,
             fields.content_length,
             &fragment_hashes,
@@ -352,20 +389,12 @@ fn extract_ddos_samples_from_entries(
             .push(fields.content_length.min(u32::MAX as u64) as u32);
         state
             .has_cookies
-            .push(fields.has_cookies.unwrap_or(false));
+            .push(fields.has_cookies);
         state.has_referer.push(
-            fields
-                .referer
-                .as_deref()
-                .map(|r| r != "-")
-                .unwrap_or(false),
+            !fields.referer.is_empty() && fields.referer != "-",
         );
         state.has_accept_language.push(
-            fields
-                .accept_language
-                .as_deref()
-                .map(|a| a != "-")
-                .unwrap_or(false),
+            !fields.accept_language.is_empty() && fields.accept_language != "-",
         );
         state.suspicious_paths.push(
             crate::ddos::features::is_suspicious_path(&fields.path),
@@ -462,17 +491,9 @@ fn entries_to_scanner_samples(
     let mut samples = Vec::new();
 
     for (fields, host_prefix) in entries {
-        let has_cookies = fields.has_cookies.unwrap_or(false);
-        let has_referer = fields
-            .referer
-            .as_ref()
-            .map(|r| r != "-" && !r.is_empty())
-            .unwrap_or(false);
-        let has_accept_language = fields
-            .accept_language
-            .as_ref()
-            .map(|a| a != "-" && !a.is_empty())
-            .unwrap_or(false);
+        let has_cookies = fields.has_cookies;
+        let has_referer = !fields.referer.is_empty() && fields.referer != "-";
+        let has_accept_language = !fields.accept_language.is_empty() && fields.accept_language != "-";
 
         let feats = features::extract_features(
             &fields.method,
@@ -481,7 +502,7 @@ fn entries_to_scanner_samples(
             has_cookies,
             has_referer,
             has_accept_language,
-            "-",
+            &fields.accept,
             &fields.user_agent,
             fields.content_length,
             &fragment_hashes,
@@ -587,11 +608,12 @@ mod tests {
             duration_ms: 10,
             content_length: 0,
             user_agent: "Mozilla/5.0".to_string(),
-            has_cookies: Some(true),
-            referer: Some("https://test.sunbeam.pt".to_string()),
-            accept_language: Some("en-US".to_string()),
+            has_cookies: true,
+            referer: "https://test.sunbeam.pt".to_string(),
+            accept_language: "en-US".to_string(),
             backend: "test-svc:8080".to_string(),
             label: Some(label.to_string()),
+            ..AuditFields::default()
         };
         (fields, "test".to_string())
     }
diff --git a/src/ddos/audit_log.rs b/src/ddos/audit_log.rs
index e022b55..f6f9c51 100644
--- a/src/ddos/audit_log.rs
+++ b/src/ddos/audit_log.rs
@@ -1,83 +1,11 @@
-use serde::Deserialize;
+// Copyright Sunbeam Studios 2026
+// SPDX-License-Identifier: Apache-2.0
 
-#[derive(Deserialize)]
-pub struct AuditLog {
-    pub timestamp: String,
-    pub fields: AuditFields,
-}
+//! Re-exports from `crate::audit` — the canonical audit log definition.
+//!
+//! All new code should `use crate::audit::*` directly.
 
-#[derive(Deserialize)]
-pub struct AuditFields {
-    pub method: String,
-    pub host: String,
-    pub path: String,
-    pub client_ip: String,
-    #[serde(deserialize_with = "flexible_u16")]
-    pub status: u16,
-    #[serde(deserialize_with = "flexible_u64")]
-    pub duration_ms: u64,
-    #[serde(default)]
-    pub backend: String,
-    #[serde(default)]
-    pub content_length: u64,
-    #[serde(default = "default_ua")]
-    pub user_agent: String,
-    #[serde(default)]
-    pub query: String,
-    #[serde(default)]
-    pub has_cookies: Option<bool>,
-    #[serde(default)]
-    pub referer: Option<String>,
-    #[serde(default)]
-    pub accept_language: Option<String>,
-    /// Optional ground-truth label from external datasets (e.g. CSIC 2010).
-    /// Values: "attack", "normal". When present, trainers should use this
-    /// instead of heuristic labeling.
-    #[serde(default)]
-    pub label: Option<String>,
-}
-
-fn default_ua() -> String {
-    "-".to_string()
-}
-
-pub fn flexible_u64<'de, D: serde::Deserializer<'de>>(
-    deserializer: D,
-) -> std::result::Result<u64, D::Error> {
-    #[derive(Deserialize)]
-    #[serde(untagged)]
-    enum StringOrNum {
-        Num(u64),
-        Str(String),
-    }
-    match StringOrNum::deserialize(deserializer)? {
-        StringOrNum::Num(n) => Ok(n),
-        StringOrNum::Str(s) => s.parse().map_err(serde::de::Error::custom),
-    }
-}
-
-pub fn flexible_u16<'de, D: serde::Deserializer<'de>>(
-    deserializer: D,
-) -> std::result::Result<u16, D::Error> {
-    #[derive(Deserialize)]
-    #[serde(untagged)]
-    enum StringOrNum {
-        Num(u16),
-        Str(String),
-    }
-    match StringOrNum::deserialize(deserializer)? {
-        StringOrNum::Num(n) => Ok(n),
-        StringOrNum::Str(s) => s.parse().map_err(serde::de::Error::custom),
-    }
-}
-
-/// Strip the port suffix from a socket address string.
-pub fn strip_port(addr: &str) -> &str {
-    if addr.starts_with('[') {
-        addr.find(']').map(|i| &addr[1..i]).unwrap_or(addr)
-    } else if let Some(pos) = addr.rfind(':') {
-        &addr[..pos]
-    } else {
-        addr
-    }
-}
+pub use crate::audit::strip_port;
+pub use crate::audit::AuditFields;
+pub use crate::audit::AuditLogLine as AuditLog;
+pub use crate::audit::{flexible_u16, flexible_u64};
diff --git a/src/ddos/detector.rs b/src/ddos/detector.rs
index ffc5183..d712061 100644
--- a/src/ddos/detector.rs
+++ b/src/ddos/detector.rs
@@ -1,6 +1,9 @@
+// Copyright Sunbeam Studios 2026
+// SPDX-License-Identifier: Apache-2.0
+
 use crate::config::DDoSConfig;
 use crate::ddos::features::{method_to_u8, IpState, RequestEvent};
-use crate::ddos::model::{DDoSAction, TrainedModel};
+use crate::ddos::model::DDoSAction;
 use rustc_hash::FxHashMap;
 use std::hash::{Hash, Hasher};
 use std::net::IpAddr;
@@ -10,12 +13,10 @@ use std::time::Instant;
 const NUM_SHARDS: usize = 256;
 
 pub struct DDoSDetector {
-    model: TrainedModel,
     shards: Vec<RwLock<FxHashMap<IpAddr, IpState>>>,
     window_secs: u64,
     window_capacity: usize,
     min_events: usize,
-    use_ensemble: bool,
 }
 
 fn shard_index(ip: &IpAddr) -> usize {
@@ -25,34 +26,15 @@ fn shard_index(ip: &IpAddr) -> usize {
 }
 
 impl DDoSDetector {
-    pub fn new(model: TrainedModel, config: &DDoSConfig) -> Self {
+    pub fn new(config: &DDoSConfig) -> Self {
         let shards = (0..NUM_SHARDS)
             .map(|_| RwLock::new(FxHashMap::default()))
             .collect();
         Self {
-            model,
             shards,
             window_secs: config.window_secs,
             window_capacity: config.window_capacity,
             min_events: config.min_events,
-            use_ensemble: false,
-        }
-    }
-
-    /// Create a detector that uses the ensemble (decision tree + MLP) path.
-    /// A dummy model is still needed for fallback, but ensemble inference
-    /// takes priority when `use_ensemble` is true.
-    pub fn new_ensemble(model: TrainedModel, config: &DDoSConfig) -> Self {
-        let shards = (0..NUM_SHARDS)
-            .map(|_| RwLock::new(FxHashMap::default()))
-            .collect();
-        Self {
-            model,
-            shards,
-            window_secs: config.window_secs,
-            window_capacity: config.window_capacity,
-            min_events: config.min_events,
-            use_ensemble: true,
         }
     }
 
@@ -99,24 +81,20 @@ impl DDoSDetector {
 
         let features = state.extract_features(self.window_secs);
 
-        if self.use_ensemble {
-            // Cast f64 features to f32 array for ensemble inference.
-            let mut f32_features = [0.0f32; 14];
-            for (i, &v) in features.iter().enumerate().take(14) {
-                f32_features[i] = v as f32;
-            }
-            let ev = crate::ensemble::ddos::ddos_ensemble_predict(&f32_features);
-            crate::metrics::DDOS_ENSEMBLE_PATH
-                .with_label_values(&[match ev.path {
-                    crate::ensemble::ddos::DDoSEnsemblePath::TreeBlock => "tree_block",
-                    crate::ensemble::ddos::DDoSEnsemblePath::TreeAllow => "tree_allow",
-                    crate::ensemble::ddos::DDoSEnsemblePath::Mlp => "mlp",
-                }])
-                .inc();
-            return ev.action;
+        // Cast f64 features to f32 array for ensemble inference.
+        let mut f32_features = [0.0f32; 14];
+        for (i, &v) in features.iter().enumerate().take(14) {
+            f32_features[i] = v as f32;
         }
-
-        self.model.classify(&features)
+        let ev = crate::ensemble::ddos::ddos_ensemble_predict(&f32_features);
+        crate::metrics::DDOS_ENSEMBLE_PATH
+            .with_label_values(&[match ev.path {
+                crate::ensemble::ddos::DDoSEnsemblePath::TreeBlock => "tree_block",
+                crate::ensemble::ddos::DDoSEnsemblePath::TreeAllow => "tree_allow",
+                crate::ensemble::ddos::DDoSEnsemblePath::Mlp => "mlp",
+            }])
+            .inc();
+        ev.action
     }
 
     /// Feed response data back into the IP's event history.
@@ -125,10 +103,6 @@ impl DDoSDetector {
         // Status/duration from check() are 0-initialized; the next request
         // will have fresh data. This is intentionally a no-op for now.
     }
-
-    pub fn point_count(&self) -> usize {
-        self.model.point_count()
-    }
 }
 
 fn fx_hash(s: &str) -> u64 {
diff --git a/src/ddos/mod.rs b/src/ddos/mod.rs
index 93f324d..c369aa6 100644
--- a/src/ddos/mod.rs
+++ b/src/ddos/mod.rs
@@ -1,6 +1,8 @@
+// Copyright Sunbeam Studios 2026
+// SPDX-License-Identifier: Apache-2.0
+
 pub mod audit_log;
 pub mod detector;
 pub mod features;
 pub mod model;
-pub mod replay;
 pub mod train;
diff --git a/src/ddos/model.rs b/src/ddos/model.rs
index e655915..74b1cba 100644
--- a/src/ddos/model.rs
+++ b/src/ddos/model.rs
@@ -1,183 +1,8 @@
-use crate::ddos::features::{FeatureVector, NormParams, NUM_FEATURES};
-use anyhow::{Context, Result};
-use serde::{Deserialize, Serialize};
-use std::path::Path;
-
-#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)]
-pub enum TrafficLabel {
-    Normal,
-    Attack,
-}
-
-#[derive(Serialize, Deserialize)]
-pub struct SerializedModel {
-    pub points: Vec<FeatureVector>,
-    pub labels: Vec<TrafficLabel>,
-    pub norm_params: NormParams,
-    pub k: usize,
-    pub threshold: f64,
-}
-
-pub struct TrainedModel {
-    /// Stored points (normalized). The kD-tree borrows these.
-    points: Vec<[f64; NUM_FEATURES]>,
-    labels: Vec<TrafficLabel>,
-    norm_params: NormParams,
-    k: usize,
-    threshold: f64,
-}
+// Copyright Sunbeam Studios 2026
+// SPDX-License-Identifier: Apache-2.0
 
 #[derive(Debug, Clone, Copy, PartialEq, Eq)]
 pub enum DDoSAction {
     Allow,
     Block,
 }
-
-impl TrainedModel {
-    pub fn load(path: &Path, k_override: Option<usize>, threshold_override: Option<f64>) -> Result<Self> {
-        let data = std::fs::read(path)
-            .with_context(|| format!("reading model from {}", path.display()))?;
-        let model: SerializedModel =
-            bincode::deserialize(&data).context("deserializing model")?;
-        Ok(Self {
-            points: model.points,
-            labels: model.labels,
-            norm_params: model.norm_params,
-            k: k_override.unwrap_or(model.k),
-            threshold: threshold_override.unwrap_or(model.threshold),
-        })
-    }
-
-    /// Create an empty model (no training points). Used when the ensemble
-    /// path is active and the KNN model is not needed.
-    pub fn empty(k: usize, threshold: f64) -> Self {
-        Self {
-            points: vec![],
-            labels: vec![],
-            norm_params: NormParams {
-                mins: [0.0; NUM_FEATURES],
-                maxs: [1.0; NUM_FEATURES],
-            },
-            k,
-            threshold,
-        }
-    }
-
-    pub fn from_serialized(model: SerializedModel) -> Self {
-        Self {
-            points: model.points,
-            labels: model.labels,
-            norm_params: model.norm_params,
-            k: model.k,
-            threshold: model.threshold,
-        }
-    }
-
-    pub fn classify(&self, features: &FeatureVector) -> DDoSAction {
-        let normalized = self.norm_params.normalize(features);
-
-        if self.points.is_empty() {
-            return DDoSAction::Allow;
-        }
-
-        // Build tree on-the-fly for query. In production with many queries,
-        // we'd cache this, but the tree build is fast for <100K points.
-        // fnntw::Tree borrows data, so we build it here.
-        let tree = match fnntw::Tree::<'_, f64, NUM_FEATURES>::new(&self.points, 32) {
-            Ok(t) => t,
-            Err(_) => return DDoSAction::Allow,
-        };
-
-        let k = self.k.min(self.points.len());
-        let result = tree.query_nearest_k(&normalized, k);
-        match result {
-            Ok((_distances, indices)) => {
-                let attack_count = indices
-                    .iter()
-                    .filter(|&&idx| self.labels[idx as usize] == TrafficLabel::Attack)
-                    .count();
-                let attack_frac = attack_count as f64 / k as f64;
-                if attack_frac >= self.threshold {
-                    DDoSAction::Block
-                } else {
-                    DDoSAction::Allow
-                }
-            }
-            Err(_) => DDoSAction::Allow,
-        }
-    }
-
-    pub fn norm_params(&self) -> &NormParams {
-        &self.norm_params
-    }
-
-    pub fn point_count(&self) -> usize {
-        self.points.len()
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    #[test]
-    fn test_classify_empty_model() {
-        let model = TrainedModel {
-            points: vec![],
-            labels: vec![],
-            norm_params: NormParams {
-                mins: [0.0; NUM_FEATURES],
-                maxs: [1.0; NUM_FEATURES],
-            },
-            k: 5,
-            threshold: 0.6,
-        };
-        assert_eq!(model.classify(&[0.5; NUM_FEATURES]), DDoSAction::Allow);
-    }
-
-    fn make_test_points(n: usize) -> Vec<FeatureVector> {
-        (0..n)
-            .map(|i| {
-                let mut v = [0.0; NUM_FEATURES];
-                for d in 0..NUM_FEATURES {
-                    v[d] = ((i * (d + 1)) as f64 / n as f64) % 1.0;
-                }
-                v
-            })
-            .collect()
-    }
-
-    #[test]
-    fn test_classify_all_attack() {
-        let points = make_test_points(100);
-        let labels = vec![TrafficLabel::Attack; 100];
-        let model = TrainedModel {
-            points,
-            labels,
-            norm_params: NormParams {
-                mins: [0.0; NUM_FEATURES],
-                maxs: [1.0; NUM_FEATURES],
-            },
-            k: 5,
-            threshold: 0.6,
-        };
-        assert_eq!(model.classify(&[0.5; NUM_FEATURES]), DDoSAction::Block);
-    }
-
-    #[test]
-    fn test_classify_all_normal() {
-        let points = make_test_points(100);
-        let labels = vec![TrafficLabel::Normal; 100];
-        let model = TrainedModel {
-            points,
-            labels,
-            norm_params: NormParams {
-                mins: [0.0; NUM_FEATURES],
-                maxs: [1.0; NUM_FEATURES],
-            },
-            k: 5,
-            threshold: 0.6,
-        };
-        assert_eq!(model.classify(&[0.5; NUM_FEATURES]), DDoSAction::Allow);
-    }
-}
diff --git a/src/ddos/replay.rs b/src/ddos/replay.rs
deleted file mode 100644
index f7f26fc..0000000
--- a/src/ddos/replay.rs
+++ /dev/null
@@ -1,281 +0,0 @@
-use crate::config::{DDoSConfig, RateLimitConfig};
-use crate::ddos::audit_log::{self, AuditLog};
-use crate::ddos::detector::DDoSDetector;
-use crate::ddos::model::{DDoSAction, TrainedModel};
-use crate::rate_limit::key::RateLimitKey;
-use crate::rate_limit::limiter::{RateLimitResult, RateLimiter};
-use anyhow::{Context, Result};
-use rustc_hash::FxHashMap;
-use std::io::BufRead;
-use std::net::IpAddr;
-use std::sync::Arc;
-
-pub struct ReplayArgs {
-    pub input: String,
-    pub model_path: String,
-    pub config_path: Option<String>,
-    pub k: usize,
-    pub threshold: f64,
-    pub window_secs: u64,
-    pub min_events: usize,
-    pub rate_limit: bool,
-}
-
-pub struct ReplayResult {
-    pub total: u64,
-    pub skipped: u64,
-    pub ddos_blocked: u64,
-    pub rate_limited: u64,
-    pub allowed: u64,
-    pub ddos_blocked_ips: FxHashMap<String, u64>,
-    pub rate_limited_ips: FxHashMap<String, u64>,
-    pub false_positive_ips: usize,
-    pub true_positive_ips: usize,
-}
-
-/// Core replay pipeline: load model, replay logs, compute stats including false positive analysis.
-pub fn replay_and_evaluate(args: &ReplayArgs) -> Result<ReplayResult> {
-    let model = TrainedModel::load(
-        std::path::Path::new(&args.model_path),
-        Some(args.k),
-        Some(args.threshold),
-    )
-    .with_context(|| format!("loading model from {}", args.model_path))?;
-
-    let ddos_cfg = DDoSConfig {
-        model_path: Some(args.model_path.clone()),
-        k: args.k,
-        threshold: args.threshold,
-        window_secs: args.window_secs,
-        window_capacity: 1000,
-        min_events: args.min_events,
-        enabled: true,
-        use_ensemble: false,
-    };
-    let detector = Arc::new(DDoSDetector::new(model, &ddos_cfg));
-
-    let rate_limiter = if args.rate_limit {
-        let rl_cfg = if let Some(cfg_path) = &args.config_path {
-            let cfg = crate::config::Config::load(cfg_path)?;
-            cfg.rate_limit.unwrap_or_else(default_rate_limit_config)
-        } else {
-            default_rate_limit_config()
-        };
-        Some(RateLimiter::new(&rl_cfg))
-    } else {
-        None
-    };
-
-    let file = std::fs::File::open(&args.input)
-        .with_context(|| format!("opening {}", args.input))?;
-    let reader = std::io::BufReader::new(file);
-
-    let mut total = 0u64;
-    let mut skipped = 0u64;
-    let mut ddos_blocked = 0u64;
-    let mut rate_limited = 0u64;
-    let mut allowed = 0u64;
-    let mut ddos_blocked_ips: FxHashMap<String, u64> = FxHashMap::default();
-    let mut rate_limited_ips: FxHashMap<String, u64> = FxHashMap::default();
-
-    for line in reader.lines() {
-        let line = line?;
-        let entry: AuditLog = match serde_json::from_str(&line) {
-            Ok(e) => e,
-            Err(_) => {
-                skipped += 1;
-                continue;
-            }
-        };
-
-        if entry.fields.method.is_empty() {
-            skipped += 1;
-            continue;
-        }
-
-        total += 1;
-
-        let ip_str = audit_log::strip_port(&entry.fields.client_ip).to_string();
-        let ip: IpAddr = match ip_str.parse() {
-            Ok(ip) => ip,
-            Err(_) => {
-                skipped += 1;
-                continue;
-            }
-        };
-
-        let has_cookies = entry.fields.has_cookies.unwrap_or(false);
-        let has_referer = entry.fields.referer.as_deref().map(|r| r != "-").unwrap_or(false);
-        let has_accept_language = entry.fields.accept_language.as_deref().map(|a| a != "-").unwrap_or(false);
-        let ddos_action = detector.check(
-            ip,
-            &entry.fields.method,
-            &entry.fields.path,
-            &entry.fields.host,
-            &entry.fields.user_agent,
-            entry.fields.content_length,
-            has_cookies,
-            has_referer,
-            has_accept_language,
-        );
-
-        if ddos_action == DDoSAction::Block {
-            ddos_blocked += 1;
-            *ddos_blocked_ips.entry(ip_str.clone()).or_insert(0) += 1;
-            continue;
-        }
-
-        if let Some(limiter) = &rate_limiter {
-            let rl_key = RateLimitKey::Ip(ip);
-            if let RateLimitResult::Reject { .. } = limiter.check(ip, rl_key) {
-                rate_limited += 1;
-                *rate_limited_ips.entry(ip_str.clone()).or_insert(0) += 1;
-                continue;
-            }
-        }
-
-        allowed += 1;
-    }
-
-    // Compute false positive / true positive counts
-    let (false_positive_ips, true_positive_ips) =
-        count_fp_tp(&args.input, &ddos_blocked_ips, &rate_limited_ips)?;
-
-    Ok(ReplayResult {
-        total,
-        skipped,
-        ddos_blocked,
-        rate_limited,
-        allowed,
-        ddos_blocked_ips,
-        rate_limited_ips,
-        false_positive_ips,
-        true_positive_ips,
-    })
-}
-
-/// Count false-positive and true-positive IPs from blocked set.
-/// FP = blocked IP where >60% of original responses were 2xx/3xx.
-fn count_fp_tp(
-    input: &str,
-    ddos_blocked_ips: &FxHashMap<String, u64>,
-    rate_limited_ips: &FxHashMap<String, u64>,
-) -> Result<(usize, usize)> {
-    let blocked_ips: rustc_hash::FxHashSet<&str> = ddos_blocked_ips
-        .keys()
-        .chain(rate_limited_ips.keys())
-        .map(|s| s.as_str())
-        .collect();
-
-    if blocked_ips.is_empty() {
-        return Ok((0, 0));
-    }
-
-    let file = std::fs::File::open(input)?;
-    let reader = std::io::BufReader::new(file);
-    let mut ip_statuses: FxHashMap<String, Vec<u16>> = FxHashMap::default();
-
-    for line in reader.lines() {
-        let line = line?;
-        let entry: AuditLog = match serde_json::from_str(&line) {
-            Ok(e) => e,
-            Err(_) => continue,
-        };
-        let ip_str = audit_log::strip_port(&entry.fields.client_ip).to_string();
-        if blocked_ips.contains(ip_str.as_str()) {
-            ip_statuses.entry(ip_str).or_default().push(entry.fields.status);
-        }
-    }
-
-    let mut fp = 0usize;
-    let mut tp = 0usize;
-    for (_ip, statuses) in &ip_statuses {
-        let total = statuses.len();
-        let ok_count = statuses.iter().filter(|&&s| (200..400).contains(&s)).count();
-        let ok_pct = ok_count as f64 / total as f64 * 100.0;
-        if ok_pct > 60.0 {
-            fp += 1;
-        } else {
-            tp += 1;
-        }
-    }
-
-    Ok((fp, tp))
-}
-
-pub fn run(args: ReplayArgs) -> Result<()> {
-    eprintln!("Loading model from {}...", args.model_path);
-    eprintln!("Replaying {}...\n", args.input);
-
-    let result = replay_and_evaluate(&args)?;
-
-    let total = result.total;
-    eprintln!("═══ Replay Results ═══════════════════════════════════════");
-    eprintln!("  Total requests:    {total}");
-    eprintln!("  Skipped (parse):   {}", result.skipped);
-    eprintln!("  Allowed:           {} ({:.1}%)", result.allowed, pct(result.allowed, total));
-    eprintln!("  DDoS blocked:      {} ({:.1}%)", result.ddos_blocked, pct(result.ddos_blocked, total));
-    if args.rate_limit {
-        eprintln!("  Rate limited:      {} ({:.1}%)", result.rate_limited, pct(result.rate_limited, total));
-    }
-
-    if !result.ddos_blocked_ips.is_empty() {
-        eprintln!("\n── DDoS-blocked IPs (top 20) ─────────────────────────────");
-        let mut sorted: Vec<_> = result.ddos_blocked_ips.iter().collect();
-        sorted.sort_by(|a, b| b.1.cmp(a.1));
-        for (ip, count) in sorted.iter().take(20) {
-            eprintln!("  {:<40} {} reqs blocked", ip, count);
-        }
-    }
-
-    if !result.rate_limited_ips.is_empty() {
-        eprintln!("\n── Rate-limited IPs (top 20) ─────────────────────────────");
-        let mut sorted: Vec<_> = result.rate_limited_ips.iter().collect();
-        sorted.sort_by(|a, b| b.1.cmp(a.1));
-        for (ip, count) in sorted.iter().take(20) {
-            eprintln!("  {:<40} {} reqs limited", ip, count);
-        }
-    }
-
-    eprintln!("\n── False positive check ──────────────────────────────────");
-    if result.false_positive_ips == 0 {
-        eprintln!("  No likely false positives found.");
-    } else {
-        eprintln!("  ⚠ {} IPs were blocked but had mostly successful responses", result.false_positive_ips);
-    }
-    eprintln!("  True positive IPs: {}", result.true_positive_ips);
-
-    eprintln!("══════════════════════════════════════════════════════════");
-    Ok(())
-}
-
-fn default_rate_limit_config() -> RateLimitConfig {
-    RateLimitConfig {
-        enabled: true,
-        bypass_cidrs: vec![
-            "10.0.0.0/8".into(),
-            "172.16.0.0/12".into(),
-            "192.168.0.0/16".into(),
-            "100.64.0.0/10".into(),
-            "fd00::/8".into(),
-        ],
-        eviction_interval_secs: 300,
-        stale_after_secs: 600,
-        authenticated: crate::config::BucketConfig {
-            burst: 200,
-            rate: 50.0,
-        },
-        unauthenticated: crate::config::BucketConfig {
-            burst: 60,
-            rate: 15.0,
-        },
-    }
-}
-
-fn pct(n: u64, total: u64) -> f64 {
-    if total == 0 {
-        0.0
-    } else {
-        (n as f64 / total as f64) * 100.0
-    }
-}
diff --git a/src/ddos/train.rs b/src/ddos/train.rs
index 4c3789f..4378883 100644
--- a/src/ddos/train.rs
+++ b/src/ddos/train.rs
@@ -1,13 +1,32 @@
+// Copyright Sunbeam Studios 2026
+// SPDX-License-Identifier: Apache-2.0
+
 use crate::ddos::audit_log::AuditLog;
 use crate::ddos::audit_log;
 use crate::ddos::features::{method_to_u8, FeatureVector, LogIpState, NormParams, NUM_FEATURES};
-use crate::ddos::model::{SerializedModel, TrafficLabel};
 use anyhow::{bail, Context, Result};
 use rustc_hash::{FxHashMap, FxHashSet};
-use serde::Deserialize;
+use serde::{Deserialize, Serialize};
 use std::hash::{Hash, Hasher};
 use std::io::BufRead;
 
+/// Legacy KNN training types — kept for the `train-ddos` CLI command
+/// which produces bincode model files for offline evaluation.
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)]
+pub enum TrafficLabel {
+    Normal,
+    Attack,
+}
+
+#[derive(Serialize, Deserialize)]
+pub struct SerializedModel {
+    pub points: Vec<FeatureVector>,
+    pub labels: Vec<TrafficLabel>,
+    pub norm_params: NormParams,
+    pub k: usize,
+    pub threshold: f64,
+}
+
 #[derive(Deserialize)]
 pub struct HeuristicThresholds {
     /// Requests/second above which an IP is labeled attack
@@ -255,12 +274,12 @@ pub fn parse_logs(input: &str) -> Result<FxHashMap<String, LogIpState>> {
         state.statuses.push(entry.fields.status);
         state.durations.push(entry.fields.duration_ms.min(u32::MAX as u64) as u32);
         state.content_lengths.push(entry.fields.content_length.min(u32::MAX as u64) as u32);
-        state.has_cookies.push(entry.fields.has_cookies.unwrap_or(false));
+        state.has_cookies.push(entry.fields.has_cookies);
         state.has_referer.push(
-            entry.fields.referer.as_deref().map(|r| r != "-").unwrap_or(false),
+            !entry.fields.referer.is_empty() && entry.fields.referer != "-",
         );
         state.has_accept_language.push(
-            entry.fields.accept_language.as_deref().map(|a| a != "-").unwrap_or(false),
+            !entry.fields.accept_language.is_empty() && entry.fields.accept_language != "-",
         );
         state.suspicious_paths.push(
             crate::ddos::features::is_suspicious_path(&entry.fields.path),
diff --git a/src/ensemble/ddos.rs b/src/ensemble/ddos.rs
index 2b82e1b..c59593b 100644
--- a/src/ensemble/ddos.rs
+++ b/src/ensemble/ddos.rs
@@ -1,3 +1,6 @@
+// Copyright Sunbeam Studios 2026
+// SPDX-License-Identifier: Apache-2.0
+
 use crate::ddos::model::DDoSAction;
 use super::gen::ddos_weights;
 use super::mlp::mlp_predict_32;
@@ -80,59 +83,46 @@ mod tests {
     use super::*;
 
     #[test]
-    fn test_tree_allow_path() {
-        // All zeros → feature 4 (request_rate) = 0.0 <= 0.70 → left (node 1)
-        //   feature 10 (cookie_ratio) = 0.0 <= 0.30 → left (node 3) → Allow
+    fn test_tree_block_path() {
+        // Tree: root splits on feature 10 (cookie_ratio) at 0.14.
+        // All zeros → cookie_ratio normalized = 0.0 <= 0.14 → Block (node 1)
         let raw = [0.0f32; 14];
         let v = ddos_ensemble_predict(&raw);
+        assert_eq!(v.action, DDoSAction::Block);
+        assert_eq!(v.path, DDoSEnsemblePath::TreeBlock);
+    }
+
+    #[test]
+    fn test_tree_allow_path() {
+        // Tree: feature 10 (cookie_ratio) > 0.14 → node 2 (Allow leaf)
+        // feature 10 range [0, 1], raw 0.5 → normalized 0.5 > 0.14 → Allow
+        let mut raw = [0.0f32; 14];
+        raw[10] = 0.5;
+        let v = ddos_ensemble_predict(&raw);
         assert_eq!(v.action, DDoSAction::Allow);
         assert_eq!(v.path, DDoSEnsemblePath::TreeAllow);
         assert_eq!(v.reason, "ensemble:tree_allow");
     }
 
     #[test]
-    fn test_tree_block_path() {
-        // Need: feature 4 (request_rate) > 0.70 normalized → right (node 2)
-        //   feature 12 (accept_language_ratio) > 0.25 normalized → right (node 6) → Block
-        // feature 4 max = 500, so raw 400 → normalized 0.8 > 0.70 ✓
-        // feature 12 max = 1.0, so raw 0.5 → normalized 0.5 > 0.25 ✓
-        let mut raw = [0.0f32; 14];
-        raw[4] = 400.0;
-        raw[12] = 0.5;
-        let v = ddos_ensemble_predict(&raw);
-        assert_eq!(v.action, DDoSAction::Block);
-        assert_eq!(v.path, DDoSEnsemblePath::TreeBlock);
-    }
-
-    #[test]
-    fn test_mlp_path() {
-        // Need: feature 4 > 0.70 normalized → right (node 2)
-        //   feature 12 <= 0.25 normalized → left (node 5) → Defer
-        // feature 4 max = 500, raw 400 → 0.8 > 0.70 ✓
-        // feature 12 max = 1.0, raw 0.1 → 0.1 <= 0.25 ✓
-        let mut raw = [0.0f32; 14];
-        raw[4] = 400.0;
-        raw[12] = 0.1;
-        let v = ddos_ensemble_predict(&raw);
-        assert_eq!(v.path, DDoSEnsemblePath::Mlp);
-        assert_eq!(v.reason, "ensemble:mlp");
-        assert!(v.score >= 0.0 && v.score <= 1.0);
-    }
-
-    #[test]
-    fn test_defer_then_mlp_allow() {
-        // Same Defer path as above — verify the MLP produces a valid action
-        let mut raw = [0.0f32; 14];
-        raw[4] = 400.0;
-        raw[12] = 0.1;
-        let v = ddos_ensemble_predict(&raw);
-        assert!(matches!(v.action, DDoSAction::Allow | DDoSAction::Block));
+    fn test_mlp_direct() {
+        // Current tree has no Defer leaves, so test MLP inference directly.
+        let input = [0.5f32; 14];
+        let score = mlp_predict_32::<14>(
+            &ddos_weights::W1,
+            &ddos_weights::B1,
+            &ddos_weights::W2,
+            ddos_weights::B2,
+            &input,
+        );
+        assert!(score >= 0.0 && score <= 1.0);
     }
 
     #[test]
     fn test_normalize_clamps_high() {
+        // feature 0 max = 10000.0, raw 999999 → clamped to 1.0
         let mut raw = [0.0f32; 14];
-        raw[0] = 999.0; // max is 100
+        raw[0] = 999999.0;
         let normed = normalize(&raw);
         assert!((normed[0] - 1.0).abs() < f32::EPSILON);
     }
@@ -140,7 +130,7 @@ mod tests {
     #[test]
     fn test_normalize_clamps_low() {
         let mut raw = [0.0f32; 14];
-        raw[1] = -500.0; // min is 0
+        raw[1] = -500.0; // min is 1.0
         let normed = normalize(&raw);
         assert!((normed[1] - 0.0).abs() < f32::EPSILON);
     }
diff --git a/src/ensemble/gen/ddos_weights.rs b/src/ensemble/gen/ddos_weights.rs
index e67b64d..058a71d 100644
--- a/src/ensemble/gen/ddos_weights.rs
+++ b/src/ensemble/gen/ddos_weights.rs
@@ -1,71 +1,74 @@
+// Copyright Sunbeam Studios 2026
+// SPDX-License-Identifier: Apache-2.0
+
 //! Auto-generated weights for the ddos ensemble.
 //! DO NOT EDIT — regenerate with `cargo run --features training -- train-ddos-mlp`.
 
 pub const THRESHOLD: f32 = 0.50000000;
 
 pub const NORM_MINS: [f32; 14] = [
-    0.08778746, 1.00000000, 1.00000000, 0.00000000, 0.00000000, 0.00000000, 0.05001374, 0.02000000, 
+    0.00000000, 1.00000000, 1.00000000, 0.00000000, 0.00000000, 0.00000000, 0.00833333, 0.02000000, 
     0.00000000, 1.00000000, 0.00000000, 0.00000000, 0.00000000, 0.00000000
 ];
 
 pub const NORM_MAXS: [f32; 14] = [
-    1000.00000000, 50.00000000, 19.00000000, 1.00000000, 7589.80468750, 1.49990082, 500.00000000, 1.00000000, 
+    10000.00000000, 50.00000000, 19.00000000, 1.00000000, 7589.80468750, 1.49999166, 500.00000000, 1.00000000, 
     171240.28125000, 30.00000000, 1.00000000, 1.00000000, 1.00000000, 1.00000000
 ];
 
 pub const W1: [[f32; 14]; 32] = [
-    [0.57458097, -0.10861993, 0.14037465, -0.23486336, -0.43255216, 0.16347405, 0.71766937, 0.83138502, 0.02852129, 0.56590265, 0.54848498, 0.38098580, 0.82907754, 0.61539698],
-    [0.06583579, 0.02305713, 0.89706898, 0.42619053, -1.20866120, -0.11974730, 1.70674825, -0.17969023, -0.26867196, 0.60768014, -0.08671998, -0.04825107, 0.58131427, -0.02062579],
-    [-0.40264100, 0.18836430, 0.08431315, 0.33763552, 0.44880620, -0.40894085, -0.22044741, 0.00533387, -0.61574107, 0.07670992, 0.63528854, 0.48244709, 0.20411402, -1.80697525],
-    [0.66713083, -0.22220801, 2.11234117, 0.41516641, -0.00165093, 0.65624571, 1.87509167, 0.63406783, -2.54182458, -0.53618753, 2.16407824, -0.61959583, -0.04717547, 0.17551991],
-    [-0.51027024, -0.60132700, 0.46407551, -0.57346475, -0.30902353, -0.24235034, 0.08087540, -2.14762974, -0.29429656, 0.56257033, -0.26935315, -0.16799171, 0.56852734, 1.93494022],
-    [-0.24938971, -0.12699288, -0.13746630, 0.64942318, 0.09490766, -0.02158179, 0.72449303, -0.28493983, -0.43053114, -0.01443988, 0.89670080, -0.34539866, -1.47019410, 0.79477930],
-    [0.62935185, 0.74686801, -0.15527052, -0.06635039, 0.73137009, 0.78417069, -0.06417987, 0.72259408, 0.85131824, 0.00477386, -0.14302900, 0.63481224, 0.92724019, -0.50126070],
-    [-0.12699059, -0.15016419, -0.48704135, 0.00581611, 0.75824696, 0.84114397, -0.08958503, 0.18609463, 0.56247348, 0.22239330, 0.43324804, 0.82077771, 0.55714250, -0.56955606],
-    [0.83457869, 0.40054807, 0.23281574, -0.58521581, 1.18067443, -0.49485078, 0.08600014, 0.99104887, -0.65019566, -0.44594154, 0.64507920, -0.61692268, -0.29301512, -0.11314666],
-    [-0.07868081, -0.18392175, 0.15165123, 0.35139060, 0.13855398, 0.16470867, 0.21025884, 1.57204449, 0.07827333, 0.05895505, 0.00810917, 1.05159700, 0.04605416, 0.38080546],
-    [1.47428405, -0.21614535, -0.35385504, 0.46582970, 1.26638246, 0.00133375, -3.85603786, 0.39766011, 1.92816520, 0.47828305, -0.16951409, -0.13771342, 0.49451983, 0.41184473],
-    [-0.23364748, 0.68134952, 0.36342716, 0.02657196, 0.07550839, 0.94861823, -0.52908695, 0.83652318, -0.05639480, 0.26536962, 0.44137934, 1.20957208, -0.60747981, -0.50647283],
-    [-0.16961956, -0.49570882, -0.33771378, -0.28554109, 0.95865113, -0.49269623, -0.44559151, 1.28568971, 0.79537493, -0.53175420, -3.19015551, 0.52214253, 0.86517984, 0.62523192],
-    [-0.16956513, -0.61727583, 0.63967121, 0.96406335, -0.28760204, 0.56459671, 0.78585202, -0.03668134, -0.14773002, -0.35764447, 0.84649116, -0.34540027, -0.12314465, -0.10070048],
-    [-0.34183556, -0.07760386, 0.70894319, 0.92814171, -0.19357866, 0.41449037, 0.54653358, 0.27682835, 0.81471086, 0.56383932, 0.57456553, -0.61491662, 0.92498505, 0.74495614],
-    [-0.38917324, -0.29217750, 1.43508542, -0.19152534, -0.18823336, 0.45097819, -0.38063127, -0.40419811, 0.56686693, -0.33231607, -0.19567636, -0.02500075, -0.04762971, 0.44703853],
-    [1.14234805, -0.62868208, -0.21298689, 0.00263968, -0.66115338, -1.12038326, 0.93599045, 0.77646011, -0.22770278, 1.43982041, 0.96078646, 1.15076077, -0.45110813, 0.83090556],
-    [0.89638984, -0.69683450, -0.29400119, 0.94997799, 0.90305328, -0.80215877, -0.09983492, -0.90757453, -0.03181892, 1.00702441, -0.97962254, -0.89580274, 0.69299418, -0.75975400],
-    [-0.75832003, -0.07210776, 0.07825917, 1.51633596, 0.44593197, 0.00936707, -0.12142835, -0.09877282, 0.06229200, 1.25678349, 0.25317946, 0.54112315, -0.17941843, 0.93283361],
-    [0.23085761, 0.53307736, 0.38696140, 0.36798462, 0.38192499, 0.23203450, 0.68225187, 0.47096270, -4.24785280, 0.18062039, 0.60047084, 0.16251479, -0.10811257, 0.48166662],
-    [0.10870802, 0.01576116, 0.00298645, 0.25878090, -0.16634797, 0.15850464, -0.24267951, 0.87678236, -0.27257833, 0.78637868, -0.00851476, 0.01502728, 0.92175138, -0.81292266],
-    [-0.74364990, -0.63139439, -0.18314177, -0.36881343, -0.53096825, -0.92442876, -0.05536628, -0.71273297, -0.94937468, -0.03863344, -0.09668982, -1.07886386, 0.58555382, 0.23351164],
-    [-0.09152136, 0.96538877, -0.11560653, -0.53110164, 0.89070886, 0.05664408, -0.71661353, 0.79684436, -0.00206013, 0.23857179, 0.06074178, -0.67188424, -0.15624331, 0.43436247],
-    [-0.28189376, -0.00535834, 0.60541785, 0.82968009, -0.21901314, -0.29874969, -0.16872653, 0.45570841, -0.25372767, -0.12359514, -1.10104620, 0.00162374, 0.07622175, 0.60413152],
-    [-1.13819373, -0.41320390, -5.57348347, 0.40931624, -1.59562767, 0.72510892, 0.03248254, 0.00407641, 0.57557869, 0.53510398, -0.35943517, 0.52707136, 0.61220711, -0.11644226],
-    [-0.02057049, 0.42545527, 0.24192038, 0.29863021, -0.22839858, -0.25318733, 0.17906551, -0.29471490, -0.04746799, 0.15909556, -0.26826856, -0.06874973, -0.03044286, 0.11770450],
-    [-0.18060833, -0.06301155, 0.01656315, -0.40476608, -0.35056075, 0.06344713, 0.32273614, -0.04382812, -0.18925793, 0.02124963, -0.23447622, 0.29704437, 0.19138981, -0.04584064],
-    [0.18248987, 0.05461208, -0.25655189, 0.16673982, 0.03251073, 0.05709980, 0.09135589, 0.06712578, -0.02372392, 0.00487196, -0.11774579, 0.34203079, 0.18477952, 0.09847298],
-    [-0.08292723, -0.03089223, 0.19555064, -0.18158682, -0.32060555, 0.18836822, -0.14625609, -0.83500093, -0.09893667, 0.02719803, 0.06864946, 0.00156752, 0.04342323, 0.30958080],
-    [-0.21274266, 0.06035644, 0.27282441, -0.01010289, -0.05599894, 0.27938741, -0.23254848, -0.20086342, -0.06775926, -0.18059292, 0.92534143, 0.09500337, 0.11612320, -0.06473339],
-    [-0.27279299, 0.96252358, 0.67542273, 0.64720130, 0.15221471, 1.67354584, 0.53074431, 0.65513390, 0.79840666, 0.78613347, 0.34742561, -1.83272552, 0.73313516, 0.09797212],
-    [-0.08888317, 0.14851266, 1.00953877, 0.19915256, -0.10076691, 0.47210938, 0.04427565, 0.19299655, 0.58729172, 0.17481442, -0.57466495, -0.16196120, 0.06293163, 1.73905540],
+    [0.30719012, 0.54619288, -0.20061351, 0.29659060, 0.34138879, -0.19088192, 0.34866381, -0.24303232, 0.20615512, 0.12656690, -0.16653502, 0.10961400, -0.16814700, 0.09950374],
+    [0.46002641, 0.49129087, -0.01386960, 0.64490628, 0.51850092, 0.69266915, 0.31095454, 0.26951542, -0.20926359, -0.09662568, 0.19281134, 0.36575633, 0.23089127, -0.03582983],
+    [0.04547556, -0.04854088, -0.10979871, -0.18705507, -0.44649851, 0.20949614, 0.73240960, 1.34691823, -0.13529004, 0.69439852, -0.40027520, 0.47921708, -0.43529814, 0.48781869],
+    [0.61547452, 1.52229679, 0.48276836, 1.27171433, -0.36176509, -0.33192506, -0.82673991, 0.67331636, -0.21094124, 1.03067887, -0.09182073, 1.44520211, 0.52611661, 0.61176163],
+    [-0.37162983, 0.48245564, -0.53393066, -0.20009390, -0.06583384, 0.17612432, 0.59905756, 1.39533114, 0.67457062, 0.06159161, -0.56609136, -0.29591814, 0.55239469, -0.56801152],
+    [0.75815558, -0.64557153, 0.84678394, 0.41179815, -0.50619060, -0.09139232, -0.64594650, -0.74464273, 0.87102652, 0.81111395, -0.35027400, 0.95135874, 0.85043454, 1.72117484],
+    [0.98888069, -0.04631047, -0.62931997, -0.37154037, 1.02817857, -1.04121590, 0.74848920, -0.26426360, 0.23142239, -0.17234743, -0.61689568, -0.59363395, -0.85756373, 0.53024006],
+    [0.59509462, 0.26622522, -0.74383926, 0.48256168, -0.75522244, 0.46806136, -0.62194610, 0.09251838, 0.25921744, 0.72987258, 0.66349596, 0.53999704, -0.25535119, -0.92465514],
+    [0.50981742, -1.44853806, -0.64814043, -0.78203505, -0.88038790, 0.68278509, 0.58861315, 0.55924416, -0.52396554, 0.45195666, -0.44876143, -0.11349974, 0.64508075, 0.06376592],
+    [0.44494510, 0.79238343, -0.22128101, 3.13757062, 0.54972911, 2.06494117, -0.20301908, 0.48413971, -0.25992882, 0.77544886, -0.18115431, 1.87130582, 0.71965748, 1.95458603],
+    [1.00518668, 1.80238068, -0.28449696, -0.02740687, -2.51049113, 0.56081659, -0.43591678, 3.59169340, -0.47954431, 1.82556272, 0.64387941, 0.56122434, -0.19696619, 3.49070907],
+    [-0.32992145, -0.03573111, 2.41438532, -0.00748284, 0.62775159, 1.78909039, 0.25103322, 0.59640545, -0.10183074, 0.83787775, 0.14171274, 0.08816884, 0.16381627, -0.04427620],
+    [0.09841868, 0.58517164, 0.02630968, 0.65797943, -0.03991833, 0.52833039, 0.37459302, 0.01832970, -1.20483434, 0.76000416, 0.02081347, 1.10453236, 0.46800232, 0.50707549],
+    [0.13568293, -0.04429439, 0.18404786, 0.74804515, -0.02402807, 0.25729915, 0.64555109, 0.09644510, 0.31338552, 0.62685025, -0.19832127, 1.95116663, 0.66340035, 1.29182386],
+    [-0.20969683, 0.56657153, -0.08705560, 0.71007556, -0.11011623, 1.16174579, 0.65050489, 1.31441426, 0.72755563, 1.15947676, -0.34925875, 0.01019314, -1.42810500, 0.14942981],
+    [-0.47017330, 2.62149596, -0.37532449, 1.17488575, 0.62930888, 0.62195790, -0.12959687, 2.36229849, -0.25786853, 0.03494137, 1.70790768, -0.02720823, 0.57822198, 1.57692003],
+    [-0.68229634, 0.87380433, 1.03171849, 0.35238963, -0.78998542, 0.97562903, -0.80616480, 1.07170749, -0.79917014, -0.43357334, 1.09133816, 0.49446958, 1.07970095, 0.27838916],
+    [-0.77235895, -0.66010702, -0.09969614, -0.38052577, -0.77211934, -0.73416811, -0.67031443, 0.62016815, 0.97461295, 1.07167208, -0.68821293, 0.51563287, -0.73027885, -0.14203216],
+    [0.90449816, -0.23423387, 1.11039567, 0.61329746, -0.21385542, 0.52449727, 0.42514217, 0.42172486, -0.33397049, 0.35888657, -0.54074812, 0.48481938, -0.05116262, -0.23848286],
+    [0.67948169, 0.50562781, 0.45344570, 0.47307885, -0.44913152, -0.11515936, 0.14361705, -0.36479098, -0.32777452, 0.11798909, -0.57137913, 0.30936614, 0.31339252, 0.51131296],
+    [-0.25677630, 0.25580657, -0.12398625, 0.24844812, 0.18556698, 0.21818036, 0.58248550, 0.50517905, 0.34329867, 0.15851928, -0.58440667, 0.33611965, 0.67439252, -0.52770680],
+    [0.66840053, -0.49819222, 0.29022828, 0.10492916, -0.06216156, 0.37093312, -0.24731418, 0.22893915, 0.32447502, 0.63166237, -0.13788179, 0.52650315, 0.15229015, 0.23656118],
+    [0.33978519, 0.15498674, -0.25265032, -0.42916322, 0.69121236, 0.20443739, 0.54050952, 0.08900955, -0.13801514, 0.25456557, -0.10714018, 0.08712567, 0.27245566, -0.29683220],
+    [-0.05526243, -1.17294025, 0.07328646, -0.07892461, 0.31488195, -0.01112767, 0.55462092, 0.65152955, -0.10721418, -0.99451303, 0.00110284, -0.53097665, 0.14362922, 0.17380728],
+    [-2.95768332, 0.46451911, 0.20220210, 0.76858771, 0.13804838, -0.80371422, -0.11160404, -0.15160939, 0.31488597, -0.10203149, 0.16458754, -0.08558689, -0.27082649, 0.03877234],
+    [-0.14654562, -0.70086712, 0.09809728, -0.60966188, 0.26278028, 0.07354698, 0.08616283, 0.36018923, 0.07040872, 0.41008693, 0.13071685, 0.18236822, 0.43306109, -0.10742717],
+    [0.41488791, -0.10255218, 0.10218169, 0.21971215, -0.05527666, -0.50265622, 0.06767768, -0.09040122, 0.16871217, -0.02748547, 0.21738021, 0.21068999, 0.10562737, -0.71913630],
+    [0.09367306, -0.14113051, -0.44151428, -0.05189204, 0.22411002, -0.09538609, 0.17464676, -0.30709952, 0.21021855, -0.27705607, 0.17645715, 0.19070518, 0.18094100, 0.10115600],
+    [-0.11084171, -0.60070217, 0.10072551, -0.09865215, 0.19512057, -0.32474023, 0.14499906, 0.06266983, -0.15383074, 0.10347557, -0.10143858, -0.09821036, -0.19187087, -0.21955618],
+    [0.09774263, 0.21607652, -0.22068830, -0.73502982, 0.14551027, -0.00246539, -0.32017741, -0.14855191, 0.15684886, -0.21544383, -0.36595181, 1.57917106, 0.45341989, 0.64960009],
+    [0.03760023, 0.12075356, -0.24193284, 0.16418910, -0.13468136, 0.40612614, 0.44222566, 0.17999728, 0.37591749, 0.67439985, -0.29388478, -0.20486754, 0.20614263, -2.63525987],
+    [-0.10479005, -0.17017230, -0.42374054, 0.30094361, 0.28561834, 0.40433934, -0.03086211, 1.49869466, -0.41601327, 0.20835553, 1.19875181, -0.00222666, 0.51400107, 0.27829245],
 ];
 
 pub const B1: [f32; 32] = [
-    -0.80723554, 0.54879200, 0.01237706, -0.22279924, 0.93692911, 0.12226531, -0.54665250, -0.49958101, 
-    -0.20918398, -0.48646352, -0.58741039, -0.50572610, -0.04772990, -0.62962151, -0.46279392, 1.14840722, 
-    -0.04871057, -0.31787100, 1.13966286, 0.69543558, -0.17798270, 0.66968435, -0.07442535, -0.70557600, 
-    0.79021728, 0.65736526, -0.30761406, 0.63242179, 0.83297908, -0.04573143, -0.18454255, -0.30583009
+    0.76754266, -0.52365464, -0.07451479, -0.24194083, 0.81372803, -0.14967601, 0.86968440, -0.11282827, 
+    0.82378083, 0.03708726, -0.14121835, -0.33332673, -0.24595253, -0.20005627, 0.80769247, 0.67842513, 
+    0.62225562, 0.55104679, 0.87356585, -0.16369765, 0.83232063, -0.40881905, -0.02851989, -0.04714838, 
+    0.69236869, -0.30938062, 0.87852216, -0.14689557, -0.52630597, -0.22946648, -0.13811214, -0.41019145
 ];
 
 pub const W2: [f32; 32] = [
-    1.09615684, -0.57856798, -0.08730038, -0.06425755, -0.96232760, -2.06290460, 0.70097560, 0.85189444, 
-    -0.10077959, 1.94375157, 0.74497795, 0.88425481, 2.11908054, 0.85526127, 0.61624259, -2.93621016, 
-    1.52211487, 0.56318259, -3.15219641, -0.55187315, 1.61819077, -0.76258671, -0.09362544, 0.86861998, 
-    -0.79028755, -0.90605170, 0.33475992, -0.79945564, -1.16680586, 0.15120529, 0.17619221, 1.61664009
+    -0.84622073, 2.32144451, 0.70330697, 0.89360833, -1.08053613, 0.69213301, -1.07218480, 0.82345659, 
+    -1.11953294, -2.58824420, 0.81520051, 1.19865966, 0.91804677, 1.04554057, -1.03049874, -0.94034135, 
+    -1.66193688, -1.53192282, -1.09629154, -4.07772017, -1.14778209, 1.15202129, 0.42650393, 0.55174673, 
+    -1.28319669, 2.06129408, -1.10220599, 0.09728605, 1.64764512, -0.14975634, 0.79428691, 1.56726408
 ];
 
-pub const B2: f32 = -0.52729088;
+pub const B2: f32 = -0.58103424;
 
 pub const TREE_NODES: [(u8, f32, u16, u16); 3] = [
-    (3, 0.30015790, 1, 2),
-    (255, 0.00000000, 0, 0),
+    (10, 0.13999981, 1, 2),
     (255, 1.00000000, 0, 0),
+    (255, 0.00000000, 0, 0),
 ];
diff --git a/src/ensemble/gen/scanner_weights.rs b/src/ensemble/gen/scanner_weights.rs
index f9030c7..999e9e3 100644
--- a/src/ensemble/gen/scanner_weights.rs
+++ b/src/ensemble/gen/scanner_weights.rs
@@ -1,3 +1,6 @@
+// Copyright Sunbeam Studios 2026
+// SPDX-License-Identifier: Apache-2.0
+
 //! Auto-generated weights for the scanner ensemble.
 //! DO NOT EDIT — regenerate with `cargo run --features training -- train-scanner-mlp`.
 
@@ -14,55 +17,55 @@ pub const NORM_MAXS: [f32; 12] = [
 ];
 
 pub const W1: [[f32; 12]; 32] = [
-    [2.25848985, 1.62502551, 1.05068624, -0.23875977, 1.29692984, -1.34665418, -1.29937541, 1.66119707, -1.43897200, 0.07720046, -1.17116165, 1.96821272],
-    [2.30885172, 0.02477695, -0.23236598, 1.66507626, -1.41407740, 1.88616431, 1.84703696, -1.46395433, 2.03542018, 1.68318951, 2.01550031, 1.94223917],
-    [2.29420924, 1.86615539, 1.69271469, 1.42137837, 1.43151915, 1.84876072, 1.09228194, 1.73608077, 0.20805965, 0.52542430, -0.02558800, 0.04718366],
-    [0.36484259, -0.02785611, -0.01155548, 0.08577330, -0.00468449, -0.07848717, 0.05191587, 0.50796396, 0.40799347, -0.14838840, -0.30566201, 0.00758083],
-    [0.28191370, 0.20945202, 0.07742970, -0.06654347, 0.17395714, 0.00011351, 0.37079588, 0.41817516, 0.56992871, 0.05705916, 0.22339216, 0.11021475],
-    [0.06522971, 0.64510870, 0.31671444, 0.34980071, 0.03446164, -0.10592904, -0.21302676, -0.04404496, 0.08638768, 0.04217484, 0.43021953, 0.21055792],
-    [0.31206250, -0.14565454, 0.38078794, 0.00860748, 0.29409558, -0.11273954, -0.02210701, 0.15525217, 0.09696059, 0.13877581, 0.06483351, 0.10946950],
-    [0.28374705, -0.02963164, 0.27863786, -0.23428085, 0.12715313, 0.09141072, 0.07769041, 0.01915955, -0.20936646, 0.02813511, -0.03910714, 0.30322370],
-    [-1.19449413, -0.84935474, -0.32267663, -0.08140022, -0.78729230, 1.58759272, 0.88281459, -0.77263606, 1.55394125, 0.10148179, 1.59524822, -0.75499195],
-    [-0.97152823, -0.12173092, 0.04745778, -0.85466659, 1.57352293, -0.52651149, -0.66270715, 1.32282484, -1.24654925, -0.45822921, -1.10187364, -0.91162699],
-    [-0.93944395, -0.57891464, -1.12100291, -0.38871467, -0.18780440, -1.11835766, -0.43614236, -1.07918274, -0.09222561, -0.23854440, -0.16720718, 0.03247443],
-    [0.13319625, 0.87437463, 0.32213065, 0.13902900, 0.64760798, 0.00899744, 0.45325586, -0.14138180, 0.13888212, 0.07780524, -0.12482210, 0.12632932],
-    [0.57018995, -0.10839911, 0.02787536, 0.16884641, 0.19435850, -0.01189608, 0.13881874, -0.10700739, -0.05463003, 0.01371983, 0.04385772, 0.01100468],
-    [-0.26600277, -0.11843663, -0.01081531, 0.10785927, -0.18684258, 0.08537511, 0.01054722, -0.01972559, -0.07416820, 0.57192892, 0.37873995, -0.00498434],
-    [0.72535324, -0.25030360, 0.51470703, -0.16410951, -0.13649474, 0.16246459, -0.27847841, 0.12250750, 0.45576489, -0.18535912, -0.45686084, 0.58293521],
-    [0.18614589, -0.32835677, -0.08683094, 0.07748202, -0.24785264, -0.16834147, 0.27066526, 0.06058804, 0.01903199, -0.17387865, 0.12752151, -0.03780220],
-    [-1.22358644, -0.78316134, -0.54068804, -0.07921790, -0.72697675, 1.80127227, 0.14326867, -0.51875746, 1.83125353, -0.02672976, 1.68589675, -0.80162954],
-    [-0.83690810, -0.12682360, 0.10783038, -0.64648604, 1.50810242, -0.48788729, -0.59418935, 0.94863659, -0.84788662, -0.49779284, -0.96408021, -1.14068258],
-    [-0.96322638, -0.50503486, -0.87195945, -0.34710455, -0.28645220, -1.10507452, -0.32122782, -0.80753750, -0.00843489, 0.04215550, 0.03197355, 0.05468401],
-    [-0.17587705, 0.45144933, 0.37954769, -0.15405300, 0.75590396, 0.00346784, 0.62332457, -0.15602241, -0.26471916, -0.19963606, -0.22497311, -0.20784236],
-    [0.60608941, -0.05316854, 0.03766245, 0.46412235, -0.41121334, -0.01225545, -0.11125158, -0.33533856, -0.04625564, -0.02995013, -0.24979964, -0.35824969],
-    [0.08163761, 0.04702193, -0.24007457, -0.23439978, 0.27066308, 0.48389259, 0.32692793, -0.23089454, 0.26520243, -0.14099684, 0.06713670, 0.14434725],
-    [-0.50808382, -0.14518137, -0.23912378, 0.33510539, 0.46566108, 0.09035082, -0.12637842, 0.55245715, -0.19972627, 0.24517706, 0.34291887, 0.01936621],
-    [0.35826349, 0.21200819, 0.65315312, -0.16792546, 0.41378024, 0.32129642, 0.50814188, 0.48289016, 0.06839173, 0.42079177, 0.52295685, 0.26273951],
-    [0.24575019, 0.10700949, 0.07041252, -0.09410189, 0.18897925, 0.31616825, -0.01306109, 0.33499330, -0.01866218, 0.06233863, 0.15316568, 0.08370106],
-    [0.17828286, 0.17363867, -0.10626584, 0.06075979, 0.39465010, 0.19557165, 0.30352867, 0.26720291, 0.40256795, 0.13942246, 0.05869288, 0.08310238],
-    [-0.04834138, 0.29206491, 0.01330532, 0.07626399, -0.17378819, 0.09515948, 0.02298534, 0.41555724, 0.09492048, 0.39422533, 0.39373979, 0.20463347],
-    [-0.11641891, -0.06529939, -0.18899654, -0.02157970, -0.03554495, 0.10956290, -0.11688691, 0.04077352, 0.34220406, -0.09558969, 0.16150762, 0.25759667],
-    [-0.17313123, 0.00591523, 0.29443163, 0.08298909, 0.07761172, 0.19023541, 0.23826212, -0.07167042, 0.08753359, 0.17917964, -0.03248737, 0.28516129],
-    [0.13091524, 0.21435370, 0.15093684, 0.30902347, 0.44151527, 0.55901742, 0.19933179, 0.06438518, 0.30585650, -0.34089112, 0.26879075, 0.12928906],
-    [-0.25311065, -0.09963353, -0.50099874, 0.57481062, 0.38744658, -0.13065037, 0.18897361, 0.49376330, -0.15626629, 0.19911517, 0.06437352, -0.09104283],
-    [0.35787049, -0.04814727, 0.45446551, -0.15264697, 0.36565515, 0.22795495, 0.24630190, 0.16362202, 0.21044184, 0.53882843, 0.42343852, 0.18454899],
+    [-0.24775992, -1.12687504, -0.84061003, 1.35276508, 1.04677176, 1.57832956, 1.47995067, 1.38580477, -0.99564040, -1.20309269, -0.24385734, 1.32367671],
+    [-1.31796920, -0.12229957, 0.89794689, 1.14832735, 1.17210162, 1.32387733, 1.37799740, 1.22984815, 0.92816162, 1.45189691, 0.97822803, -0.89625973],
+    [-1.46790743, -1.43995631, 1.44276273, 1.30967343, 1.37424576, -1.16613543, 1.46063673, -1.29701447, 0.20172349, -0.05374112, -0.02462341, 0.25295240],
+    [0.57722956, 0.39603359, -0.03453349, -0.06384056, 0.08056496, 0.30015573, -0.19045275, 0.39737019, 0.41070747, 0.03414693, -0.25483453, 0.24018581],
+    [0.02473495, 0.05047939, 0.59600842, 0.35709319, -0.32843903, 0.17655721, 0.20441811, 0.05118980, 0.24312067, 0.12677322, 0.12109834, -0.09230414],
+    [-0.03397392, -0.08393703, 0.42762470, 0.21609549, -0.28864771, -0.11998425, -0.05842599, 0.87331069, 0.52526826, 0.35321006, 0.31293729, 0.51823896],
+    [-0.26906690, -0.04246650, 0.05114691, 0.27629042, -0.37589836, -0.07396606, -0.46509269, 0.26079515, 0.33588448, 0.55919188, 0.58768439, 0.81638134],
+    [-0.48464304, 0.76111192, 0.06296005, -0.13527128, -0.41344830, -0.19461812, 0.52815431, 0.96815300, 0.47175986, -0.13977771, 0.41216326, -0.03041369],
+    [-0.05156134, 0.90191734, 0.78513384, -1.25017786, -0.54637259, -0.36098620, -0.59882820, -0.90374511, 0.91336489, 1.05806208, 0.04994302, -0.91028821],
+    [0.68376511, 0.07305489, 0.44790089, -0.56647295, -0.66570538, -0.93897015, -0.40558589, -1.51070845, 0.46759781, -1.36738360, -0.05270236, 0.98130196],
+    [0.86788160, 0.69321704, -0.53778958, -1.54257190, -0.44623125, 0.72615588, -0.75269628, 0.81946337, 0.17503875, 0.63745797, 0.48478079, -0.31573632],
+    [-0.01361719, 0.21524477, -0.10345778, -0.38488832, 0.42967409, 0.75472528, -0.07410870, -0.65231675, 0.42633417, -0.10289414, 0.09583388, -0.29391766],
+    [-0.06778818, -0.44469842, 0.05952910, -0.55139810, 0.14308600, -0.53731138, -0.07426350, 0.28065708, 0.29584157, 0.47813708, 0.02095048, -0.36458421],
+    [-0.21983556, 0.55435538, -0.13939659, 0.58281261, -0.20551582, 0.30075905, 0.13396217, -0.18145087, -0.43283740, 0.18541494, 0.07530790, -0.04916608],
+    [0.46939296, 0.57935077, -0.05478116, -0.01144989, 0.54106784, -0.18313073, 0.12232503, -0.32802504, -0.01167463, -0.13702804, -0.19521871, 0.09115479],
+    [0.25263065, -0.27172634, -0.12802953, 0.50027740, 0.05213343, 0.49081728, 0.15367918, 0.20471051, -0.22081012, 0.51709008, 0.01776243, 0.22513707],
+    [-0.20733139, 0.60041994, 0.05273124, 0.07473211, 0.14580894, -0.72007078, -0.52350652, -0.15482022, 0.19132918, 0.52586436, -0.04793828, -0.00479114],
+    [0.19872986, -0.19177110, -0.22340146, -0.48786804, -0.51010352, -0.55363113, -0.29520389, -0.21378680, -0.40099174, -0.09184421, -0.08521358, 0.61833692],
+    [0.21346046, 0.53319895, -0.44765636, -0.04764151, -0.30569363, 0.19765340, -0.41479719, 0.34292534, -0.29234713, 0.54341668, 0.60121793, -0.00226344],
+    [-0.29598647, -0.37357926, -0.25650844, -0.05165816, 0.55829030, 0.21028350, -0.28581545, -0.37299931, 0.57590896, -0.01573592, -0.19411144, 0.13814686],
+    [0.10028259, 0.02526089, -0.20488358, 0.25667843, 0.17100072, 0.01034015, -0.32994771, 0.53425753, 0.64935833, 0.30769956, -0.26756367, -0.03389005],
+    [-0.18916467, 0.38340616, -0.16475976, 0.59811211, 0.12739281, -0.16611671, -0.31913927, 0.07577144, 0.28552490, 0.54843456, 0.40937552, 0.38236183],
+    [-0.20519450, -0.04122134, -0.20013523, 0.42193425, -0.27304563, -0.21811043, 0.13115846, 0.16724831, 0.13073303, 0.20491999, 0.31806493, 0.13444173],
+    [0.01762132, 0.32608625, 0.19381267, -0.33404192, -0.46299583, -0.28042898, 0.20772585, 0.20139317, 0.41952321, -0.30363685, 0.20015827, -0.03338646],
+    [0.13760759, 0.07168494, 0.26161709, 0.41468662, -0.03778528, 0.38290465, 0.48780030, 0.39562985, 0.24758396, -0.05975538, -0.22738078, 0.27877593],
+    [0.07016940, -0.03804595, -0.08812129, 0.19664441, 0.13347355, 0.50309300, 0.26076415, 0.19044210, -0.20414594, 0.64333421, 0.15160090, 0.16449226],
+    [0.31039700, -0.01906084, 0.25622010, 0.10707659, 0.54883337, 0.19277412, 0.42004701, -0.09319381, 0.19968294, 0.07109389, -0.28979829, 0.12353907],
+    [0.28500485, 0.01991569, 0.05190456, 0.29366553, 0.01045146, -0.02013574, -0.01796320, 0.13775185, 0.11095868, -0.25678155, 0.10733776, -0.07584792],
+    [0.12738188, 0.07762879, -0.06429479, 0.39944342, 0.07958066, 0.46697047, -0.10674930, -0.12212183, -0.01540831, 0.08788434, 0.17299946, 0.25846422],
+    [0.26692817, 0.00930361, 0.24862845, 0.02167275, -0.09902105, -0.35391217, -0.41734406, 0.44949567, 0.46330830, 0.40603620, 0.08397861, 0.39809385],
+    [-0.30756459, -0.43368185, -0.00478506, 0.45611116, -0.05069341, 0.21090019, 0.28219289, 0.07687758, 0.54915971, 0.46933413, 0.35599890, 0.17573997],
+    [-0.19320646, 0.44751191, -0.14140815, 0.00427075, -0.19792002, -0.19400074, 0.19292155, 0.39845818, 0.21028778, -0.10284913, 0.31191504, -0.36995885],
 ];
 
 pub const B1: [f32; 32] = [
-    1.12135851, 0.64268047, 0.44761124, -0.28471574, 0.70866716, -0.25293177, -0.19119856, 0.39284116, 
-    -0.20628852, -0.29301032, -0.08837436, 0.92048728, 0.91167349, -0.33615190, -0.06016272, 0.79141164, 
-    -0.43257964, 0.48180589, 0.70891160, -0.24290052, 0.83115542, 0.69964927, 0.97887653, 1.34517038, 
-    1.10292709, 0.42009205, 1.07155228, 0.61349720, 0.46157768, 1.01911950, 0.51159418, 0.60460496
+    -0.24357778, -0.24826238, -0.03415382, 0.00968227, 0.51550633, 0.45242083, 0.60654080, 0.25456131, 
+    -0.36509025, -0.22825000, 0.03829522, 0.65561563, -0.19379658, -0.25716159, 0.45115772, 0.73442084, 
+    0.61352992, 0.59502298, 0.32757106, 0.28512844, 0.26663530, 0.27169749, 0.33571365, -0.34503689, 
+    -0.08054741, -0.06313029, 0.43629149, 0.35936099, 0.39375633, -0.19984132, 0.49092621, -0.27418151
 ];
 
 pub const W2: [f32; 32] = [
-    1.55191231, 1.27754235, 0.43588921, 0.10868450, 0.55931729, -1.46911597, -0.54461092, 0.78240824, 
-    -1.25938582, -0.06287600, -1.02053738, 1.07076716, 1.58776867, -0.03168033, -0.11393511, 1.30535436, 
-    -1.46621227, 0.62925971, 0.76781118, -0.74480098, 1.29669034, 0.62078375, 1.64134884, 2.09736991, 
-    1.52834618, 0.87368065, 1.80090642, 0.89230227, 0.38757962, 1.80718291, 0.64923352, 1.18709576
+    -0.04975564, -0.54667705, -0.39323062, 0.72362727, 0.86801738, 1.93621075, 1.01259410, 0.75978750, 
+    -0.67997259, -0.63063931, -0.07149173, 0.81899148, -0.69025612, -0.12359849, 1.09533453, 0.88092262, 
+    0.89678788, 0.87908030, 1.12460852, 0.76745653, 0.85632098, 0.72992527, 0.93983871, -0.55915666, 
+    -0.61104172, -0.56369978, 1.43480921, 0.71174467, 1.03119624, -0.57950914, 0.81188917, -0.78019017
 ];
 
-pub const B2: f32 = 0.23270580;
+pub const B2: f32 = 0.24626280;
 
 pub const TREE_NODES: [(u8, f32, u16, u16); 3] = [
     (3, 0.50000000, 1, 2),
diff --git a/src/ensemble/replay.rs b/src/ensemble/replay.rs
index 3dd7af0..ec63c0b 100644
--- a/src/ensemble/replay.rs
+++ b/src/ensemble/replay.rs
@@ -1,6 +1,10 @@
+// Copyright Sunbeam Studios 2026
+// SPDX-License-Identifier: Apache-2.0
+
 //! Replay audit logs through the ensemble models (scanner + DDoS).
 
-use crate::ddos::audit_log::{self, AuditLog};
+use crate::audit::AuditLogLine;
+use crate::ddos::audit_log;
 use crate::ddos::features::{method_to_u8, LogIpState};
 use crate::ddos::model::DDoSAction;
 use crate::ensemble::ddos::{ddos_ensemble_predict, DDoSEnsemblePath};
@@ -26,21 +30,31 @@ pub fn run(args: ReplayEnsembleArgs) -> Result<()> {
         std::fs::File::open(&args.input).with_context(|| format!("opening {}", args.input))?;
     let reader = std::io::BufReader::new(file);
 
-    // --- Parse all entries ---
-    let mut entries: Vec<AuditLog> = Vec::new();
-    let mut parse_errors = 0u64;
+    // --- Parse all entries, filtering for audit logs only ---
+    let mut entries: Vec<AuditLogLine> = Vec::new();
+    let mut skipped_non_audit = 0u64;
+    let mut schema_errors = 0u64;
     for line in reader.lines() {
         let line = line?;
         if line.trim().is_empty() {
             continue;
         }
-        match serde_json::from_str::<AuditLog>(&line) {
-            Ok(e) => entries.push(e),
-            Err(_) => parse_errors += 1,
+        match AuditLogLine::try_parse(&line) {
+            Ok(Some(entry)) => entries.push(entry),
+            Ok(None) => skipped_non_audit += 1,
+            Err(e) => {
+                schema_errors += 1;
+                if schema_errors <= 3 {
+                    eprintln!("  schema error: {e}");
+                }
+            }
         }
     }
     let total = entries.len() as u64;
-    eprintln!("parsed {} entries ({} parse errors)\n", total, parse_errors);
+    eprintln!(
+        "parsed {} audit entries ({} non-audit skipped, {} schema errors)\n",
+        total, skipped_non_audit, schema_errors,
+    );
 
     // --- Scanner replay ---
     eprintln!("═══ Scanner Ensemble ═════════════════════════════════════");
@@ -54,7 +68,7 @@ pub fn run(args: ReplayEnsembleArgs) -> Result<()> {
     Ok(())
 }
 
-fn replay_scanner(entries: &[AuditLog]) {
+fn replay_scanner(entries: &[AuditLogLine]) {
     let fragment_hashes: FxHashSet<u64> = crate::scanner::train::DEFAULT_FRAGMENTS
         .iter()
         .map(|f| fx_hash_bytes(f.to_ascii_lowercase().as_bytes()))
@@ -73,23 +87,15 @@ fn replay_scanner(entries: &[AuditLog]) {
     let mut blocked = 0u64;
     let mut allowed = 0u64;
     let mut path_counts = [0u64; 3]; // TreeBlock, TreeAllow, Mlp
-    let mut blocked_examples: Vec<(String, String, f64)> = Vec::new(); // (path, reason, score)
-    let mut fp_candidates: Vec<(String, u16, f64)> = Vec::new(); // blocked but had 2xx status
+    let mut blocked_examples: Vec<(String, String, String, f64)> = Vec::new(); // (path, ua, reason, score)
+    let mut fp_candidates: Vec<(String, String, u16, f64)> = Vec::new(); // blocked but had 2xx status
 
     for e in entries {
         let f = &e.fields;
         let host_prefix = f.host.split('.').next().unwrap_or("");
-        let has_cookies = f.has_cookies.unwrap_or(false);
-        let has_referer = f
-            .referer
-            .as_ref()
-            .map(|r| r != "-" && !r.is_empty())
-            .unwrap_or(false);
-        let has_accept_language = f
-            .accept_language
-            .as_ref()
-            .map(|a| a != "-" && !a.is_empty())
-            .unwrap_or(false);
+        let has_cookies = f.has_cookies;
+        let has_referer = !f.referer.is_empty() && f.referer != "-";
+        let has_accept_language = !f.accept_language.is_empty() && f.accept_language != "-";
 
         let feats = features::extract_features_f32(
             &f.method,
@@ -98,7 +104,7 @@ fn replay_scanner(entries: &[AuditLog]) {
             has_cookies,
             has_referer,
             has_accept_language,
-            "-",
+            &f.accept,
             &f.user_agent,
             f.content_length,
             &fragment_hashes,
@@ -121,12 +127,13 @@ fn replay_scanner(entries: &[AuditLog]) {
                 if blocked_examples.len() < 20 {
                     blocked_examples.push((
                         f.path.clone(),
+                        f.user_agent.clone(),
                         verdict.reason.to_string(),
                         verdict.score,
                     ));
                 }
                 if (200..400).contains(&f.status) {
-                    fp_candidates.push((f.path.clone(), f.status, verdict.score));
+                    fp_candidates.push((f.path.clone(), f.user_agent.clone(), f.status, verdict.score));
                 }
             }
             ScannerAction::Allow => allowed += 1,
@@ -159,8 +166,9 @@ fn replay_scanner(entries: &[AuditLog]) {
 
     if !blocked_examples.is_empty() {
         eprintln!("\n  blocked examples (first 20):");
-        for (path, reason, score) in &blocked_examples {
+        for (path, ua, reason, score) in &blocked_examples {
             eprintln!("    {:<50} {reason} (score={score:.3})", truncate(path, 50));
+            eprintln!("      ua: {}", truncate(ua, 72));
         }
     }
 
@@ -170,16 +178,17 @@ fn replay_scanner(entries: &[AuditLog]) {
             "\n  potential false positives (blocked but had 2xx/3xx): {}",
             fp_count
         );
-        for (path, status, score) in fp_candidates.iter().take(10) {
+        for (path, ua, status, score) in fp_candidates.iter().take(10) {
             eprintln!(
                 "    {:<50} status={status} score={score:.3}",
                 truncate(path, 50)
             );
+            eprintln!("      ua: {}", truncate(ua, 72));
         }
     }
 }
 
-fn replay_ddos(entries: &[AuditLog], window_secs: f64, min_events: usize) {
+fn replay_ddos(entries: &[AuditLogLine], window_secs: f64, min_events: usize) {
     fn fx_hash(s: &str) -> u64 {
         let mut h = rustc_hash::FxHasher::default();
         s.hash(&mut h);
@@ -208,18 +217,12 @@ fn replay_ddos(entries: &[AuditLog], window_secs: f64, min_events: usize) {
             .push(f.content_length.min(u32::MAX as u64) as u32);
         state
             .has_cookies
-            .push(f.has_cookies.unwrap_or(false));
+            .push(f.has_cookies);
         state.has_referer.push(
-            f.referer
-                .as_deref()
-                .map(|r| r != "-")
-                .unwrap_or(false),
+            !f.referer.is_empty() && f.referer != "-",
         );
         state.has_accept_language.push(
-            f.accept_language
-                .as_deref()
-                .map(|a| a != "-")
-                .unwrap_or(false),
+            !f.accept_language.is_empty() && f.accept_language != "-",
         );
         state
             .suspicious_paths
diff --git a/src/ensemble/scanner.rs b/src/ensemble/scanner.rs
index 6ad7f1d..8828331 100644
--- a/src/ensemble/scanner.rs
+++ b/src/ensemble/scanner.rs
@@ -1,3 +1,6 @@
+// Copyright Sunbeam Studios 2026
+// SPDX-License-Identifier: Apache-2.0
+
 use crate::scanner::model::{ScannerAction, ScannerVerdict};
 use super::gen::scanner_weights;
 use super::mlp::mlp_predict_32;
@@ -92,27 +95,11 @@ impl From<EnsembleVerdict> for ScannerVerdict {
 mod tests {
     use super::*;
 
-    #[test]
-    fn test_tree_allow_path() {
-        // All features at zero → feature 3 (suspicious_ua) = 0.0 <= 0.65 → left (node 1)
-        // feature 0 (path_depth) = 0.0 <= 0.40 → left (node 3) → Allow leaf
-        let raw = [0.0f32; 12];
-        let v = scanner_ensemble_predict(&raw);
-        assert_eq!(v.action, ScannerAction::Allow);
-        assert_eq!(v.path, EnsemblePath::TreeAllow);
-        assert_eq!(v.reason, "ensemble:tree_allow");
-        assert!((v.score - 0.0).abs() < f64::EPSILON);
-    }
-
     #[test]
     fn test_tree_block_path() {
-        // Need: feature 3 (suspicious_ua) > 0.65 (normalized) → right (node 2)
-        //   feature 7 (payload_entropy) > 0.72 (normalized) → right (node 6) → Block
-        // feature 3 max = 1.0, so raw 0.8 → normalized 0.8 > 0.65 ✓
-        // feature 7 max = 8.0, so raw 6.0 → normalized 0.75 > 0.72 ✓
-        let mut raw = [0.0f32; 12];
-        raw[3] = 0.8;  // suspicious_ua: normalized = 0.8/1.0 = 0.8 > 0.65
-        raw[7] = 6.0;  // payload_entropy: normalized = 6.0/8.0 = 0.75 > 0.72
+        // Tree: root splits on feature 7 (ua_category) at 0.75.
+        // All zeros → ua_category normalized = 0.0 <= 0.75 → Block (node 1)
+        let raw = [0.0f32; 12];
         let v = scanner_ensemble_predict(&raw);
         assert_eq!(v.action, ScannerAction::Block);
         assert_eq!(v.path, EnsemblePath::TreeBlock);
@@ -120,28 +107,38 @@ mod tests {
     }
 
     #[test]
-    fn test_mlp_path() {
-        // Need: feature 3 > 0.65 normalized → right (node 2)
-        //   feature 7 <= 0.72 normalized → left (node 5) → Defer
-        // Then MLP runs on the normalized input.
+    fn test_tree_allow_path() {
+        // Tree: root feature 7 > 0.75 → node 2, checks feature 3 (has_cookies) at 0.25.
+        // raw[7] = 1.0 → normalized 1.0 > 0.75 → right.
+        // raw[3] = 1.0 → normalized ~0.7 > 0.25 → right child node 6 → Allow leaf.
         let mut raw = [0.0f32; 12];
-        raw[3] = 0.8;  // normalized = 0.8 > 0.65
-        raw[7] = 4.0;  // normalized = 4.0/8.0 = 0.5 <= 0.72
-        // Also need feature 2 (query_param_count) to navigate node 5 correctly
-        // node 5: split on feature 2, threshold 0.55 → left=9(Defer), right=10
-        // normalized feature 2 = 0.0/20.0 = 0.0 <= 0.55 → left (node 9) → Defer
+        raw[7] = 1.0; // ua_category = browser
+        raw[3] = 1.0; // has_cookies = yes
         let v = scanner_ensemble_predict(&raw);
-        assert_eq!(v.path, EnsemblePath::Mlp);
-        assert_eq!(v.reason, "ensemble:mlp");
-        // MLP output is deterministic for these inputs
-        assert!(v.score >= 0.0 && v.score <= 1.0);
+        assert_eq!(v.action, ScannerAction::Allow);
+        assert_eq!(v.path, EnsemblePath::TreeAllow);
+        assert_eq!(v.reason, "ensemble:tree_allow");
+    }
+
+    #[test]
+    fn test_mlp_direct() {
+        // Current tree has no Defer leaves, so test MLP inference directly.
+        let input = [0.5f32; 12];
+        let score = mlp_predict_32::<12>(
+            &scanner_weights::W1,
+            &scanner_weights::B1,
+            &scanner_weights::W2,
+            scanner_weights::B2,
+            &input,
+        );
+        assert!(score >= 0.0 && score <= 1.0);
     }
 
     #[test]
     fn test_normalize_clamps() {
         // Values beyond max should be clamped to 1.0
         let mut raw = [0.0f32; 12];
-        raw[0] = 100.0; // max is 10.0
+        raw[0] = 100.0;
         let normed = normalize(&raw);
         assert!((normed[0] - 1.0).abs() < f64::EPSILON as f32);
     }
diff --git a/src/lib.rs b/src/lib.rs
index a799ce9..ba4b80d 100644
--- a/src/lib.rs
+++ b/src/lib.rs
@@ -1,7 +1,12 @@
+// Copyright Sunbeam Studios 2026
+// SPDX-License-Identifier: Apache-2.0
+
 // Library crate root — exports the proxy/config/acme modules so that
 // integration tests in tests/ can construct and drive a SunbeamProxy
 // without going through the binary entry point.
+#![recursion_limit = "256"]
 pub mod acme;
+pub mod audit;
 pub mod autotune;
 pub mod cache;
 pub mod cluster;
diff --git a/src/main.rs b/src/main.rs
index 04df74d..18f5943 100644
--- a/src/main.rs
+++ b/src/main.rs
@@ -1,10 +1,12 @@
+// Copyright Sunbeam Studios 2026
+// SPDX-License-Identifier: Apache-2.0
+
 mod cert;
 mod telemetry;
 mod watcher;
 
-use sunbeam_proxy::{acme, autotune, config};
+use sunbeam_proxy::{acme, config};
 use sunbeam_proxy::proxy::SunbeamProxy;
-use sunbeam_proxy::ddos;
 use sunbeam_proxy::rate_limit;
 use sunbeam_proxy::scanner;
 
@@ -32,77 +34,18 @@ enum Commands {
         #[arg(long)]
         upgrade: bool,
     },
-    /// Replay audit logs through detection models
+    /// Replay audit logs through ensemble models (scanner + DDoS)
     Replay {
-        #[command(subcommand)]
-        mode: ReplayMode,
-    },
-    /// Train a DDoS detection model from audit logs
-    TrainDdos {
         /// Path to audit log JSONL file
         #[arg(short, long)]
         input: String,
-        /// Output model file path
-        #[arg(short, long)]
-        output: String,
-        /// File with known-attack IPs (one per line)
-        #[arg(long)]
-        attack_ips: Option<String>,
-        /// File with known-normal IPs (one per line)
-        #[arg(long)]
-        normal_ips: Option<String>,
-        /// TOML file with heuristic auto-labeling thresholds
-        #[arg(long)]
-        heuristics: Option<String>,
-        /// KNN k parameter
-        #[arg(long, default_value = "5")]
-        k: usize,
-        /// Attack threshold (fraction of k neighbors)
-        #[arg(long, default_value = "0.6")]
-        threshold: f64,
         /// Sliding window size in seconds
         #[arg(long, default_value = "60")]
         window_secs: u64,
-        /// Minimum events per IP to include in training
-        #[arg(long, default_value = "10")]
+        /// Minimum events per IP before DDoS classification
+        #[arg(long, default_value = "5")]
         min_events: usize,
     },
-    /// Train a per-request scanner detection model from audit logs
-    TrainScanner {
-        /// Path to audit log JSONL file
-        #[arg(short, long)]
-        input: String,
-        /// Output model file path
-        #[arg(short, long, default_value = "scanner_model.bin")]
-        output: String,
-        /// Directory (or file) containing .txt wordlists of scanner paths
-        #[arg(long)]
-        wordlists: Option<String>,
-        /// Classification threshold
-        #[arg(long, default_value = "0.5")]
-        threshold: f64,
-        /// Include CSIC 2010 dataset as base training data (downloaded from GitHub, cached locally)
-        #[arg(long)]
-        csic: bool,
-    },
-    /// Bayesian hyperparameter optimization for DDoS model
-    AutotuneDdos {
-        /// Path to audit log JSONL file
-        #[arg(short, long)]
-        input: String,
-        /// Output best model file path
-        #[arg(short, long, default_value = "ddos_model_best.bin")]
-        output: String,
-        /// Number of optimization trials
-        #[arg(long, default_value = "200")]
-        trials: usize,
-        /// F-beta parameter (1.0 = F1, 2.0 = recall-weighted)
-        #[arg(long, default_value = "1.0")]
-        beta: f64,
-        /// JSONL file to log each trial's parameters and results
-        #[arg(long)]
-        trial_log: Option<String>,
-    },
     /// Download and cache upstream datasets (CIC-IDS2017)
     DownloadDatasets,
     /// Prepare a unified training dataset from multiple sources
@@ -125,6 +68,12 @@ enum Commands {
         /// Path to heuristics.toml for auto-labeling production logs
         #[arg(long)]
         heuristics: Option<String>,
+        /// Inject CSIC 2010 dataset as labeled audit log entries
+        #[arg(long)]
+        inject_csic: bool,
+        /// Inject OWASP ModSec audit log entries (path to .log file)
+        #[arg(long)]
+        inject_modsec: Option<String>,
     },
     #[cfg(feature = "training")]
     /// Train scanner ensemble (decision tree + MLP) from prepared dataset
@@ -142,7 +91,7 @@ enum Commands {
         #[arg(long, default_value = "100")]
         epochs: usize,
         /// Learning rate
-        #[arg(long, default_value = "0.001")]
+        #[arg(long, default_value = "0.0001")]
         learning_rate: f64,
         /// Batch size
         #[arg(long, default_value = "64")]
@@ -153,6 +102,12 @@ enum Commands {
         /// Min purity for tree leaves (below -> Defer)
         #[arg(long, default_value = "0.90")]
         tree_min_purity: f32,
+        /// Min samples required in a leaf node (higher = less overfitting)
+        #[arg(long, default_value = "2")]
+        min_samples_leaf: usize,
+        /// Weight for cookie feature (0.0=ignore, 1.0=full). Controls has_cookies influence.
+        #[arg(long, default_value = "1.0")]
+        cookie_weight: f32,
     },
     #[cfg(feature = "training")]
     /// Train DDoS ensemble (decision tree + MLP) from prepared dataset
@@ -165,184 +120,82 @@ enum Commands {
         hidden_dim: usize,
         #[arg(long, default_value = "100")]
         epochs: usize,
-        #[arg(long, default_value = "0.001")]
+        #[arg(long, default_value = "0.0001")]
         learning_rate: f64,
         #[arg(long, default_value = "64")]
         batch_size: usize,
         #[arg(long, default_value = "6")]
         tree_max_depth: usize,
+        /// Min purity for tree leaves (below -> Defer)
         #[arg(long, default_value = "0.90")]
         tree_min_purity: f32,
-    },
-    /// Bayesian hyperparameter optimization for scanner model
-    AutotuneScanner {
-        /// Path to audit log JSONL file
-        #[arg(short, long)]
-        input: String,
-        /// Output best model file path
-        #[arg(short, long, default_value = "scanner_model_best.bin")]
-        output: String,
-        /// Directory (or file) containing .txt wordlists of scanner paths
-        #[arg(long)]
-        wordlists: Option<String>,
-        /// Include CSIC 2010 dataset as base training data
-        #[arg(long)]
-        csic: bool,
-        /// Number of optimization trials
-        #[arg(long, default_value = "200")]
-        trials: usize,
-        /// F-beta parameter (1.0 = F1, 2.0 = recall-weighted)
+        /// Min samples required in a leaf node (higher = less overfitting)
+        #[arg(long, default_value = "2")]
+        min_samples_leaf: usize,
+        /// Weight for cookie feature (0.0=ignore, 1.0=full). Controls cookie_ratio influence.
         #[arg(long, default_value = "1.0")]
-        beta: f64,
-        /// JSONL file to log each trial's parameters and results
+        cookie_weight: f32,
+    },
+    #[cfg(feature = "training")]
+    /// Sweep cookie_weight values and report tree structure + validation accuracy for each
+    SweepCookieWeight {
+        /// Path to prepared dataset (.bin)
+        #[arg(short = 'd', long)]
+        dataset: String,
+        /// Which detector to sweep: "scanner" or "ddos"
+        #[arg(long, default_value = "scanner")]
+        detector: String,
+        /// Comma-separated cookie_weight values to try (default: 0.0,0.1,0.2,...,1.0)
         #[arg(long)]
-        trial_log: Option<String>,
+        weights: Option<String>,
+        /// Max tree depth
+        #[arg(long, default_value = "6")]
+        tree_max_depth: usize,
+        /// Min purity for tree leaves
+        #[arg(long, default_value = "0.90")]
+        tree_min_purity: f32,
+        /// Min samples required in a leaf node
+        #[arg(long, default_value = "2")]
+        min_samples_leaf: usize,
     },
 }
 
-#[derive(Subcommand)]
-enum ReplayMode {
-    /// Replay through ensemble models (scanner + DDoS)
-    Ensemble {
-        /// Path to audit log JSONL file
-        #[arg(short, long)]
-        input: String,
-        /// Sliding window size in seconds
-        #[arg(long, default_value = "60")]
-        window_secs: u64,
-        /// Minimum events per IP before DDoS classification
-        #[arg(long, default_value = "5")]
-        min_events: usize,
-    },
-    /// Replay through legacy KNN DDoS detector
-    Ddos {
-        /// Path to audit log JSONL file
-        #[arg(short, long)]
-        input: String,
-        /// Path to trained model file
-        #[arg(short, long, default_value = "ddos_model.bin")]
-        model: String,
-        /// Optional config file (for rate limit settings)
-        #[arg(short, long)]
-        config: Option<String>,
-        /// KNN k parameter
-        #[arg(long, default_value = "5")]
-        k: usize,
-        /// Attack threshold
-        #[arg(long, default_value = "0.6")]
-        threshold: f64,
-        /// Sliding window size in seconds
-        #[arg(long, default_value = "60")]
-        window_secs: u64,
-        /// Minimum events per IP before classification
-        #[arg(long, default_value = "10")]
-        min_events: usize,
-        /// Also run rate limiter during replay
-        #[arg(long)]
-        rate_limit: bool,
-    },
-}
 
 fn main() -> Result<()> {
     let cli = Cli::parse();
     match cli.command.unwrap_or(Commands::Serve { upgrade: false }) {
         Commands::Serve { upgrade } => run_serve(upgrade),
-        Commands::Replay { mode } => match mode {
-            ReplayMode::Ensemble { input, window_secs, min_events } => {
-                sunbeam_proxy::ensemble::replay::run(sunbeam_proxy::ensemble::replay::ReplayEnsembleArgs {
-                    input, window_secs, min_events,
-                })
-            }
-            ReplayMode::Ddos { input, model, config, k, threshold, window_secs, min_events, rate_limit } => {
-                ddos::replay::run(ddos::replay::ReplayArgs {
-                    input, model_path: model, config_path: config, k, threshold, window_secs, min_events, rate_limit,
-                })
-            }
+        Commands::Replay { input, window_secs, min_events } => {
+            sunbeam_proxy::ensemble::replay::run(sunbeam_proxy::ensemble::replay::ReplayEnsembleArgs {
+                input, window_secs, min_events,
+            })
         },
-        Commands::TrainDdos {
-            input,
-            output,
-            attack_ips,
-            normal_ips,
-            heuristics,
-            k,
-            threshold,
-            window_secs,
-            min_events,
-        } => ddos::train::run(ddos::train::TrainArgs {
-            input,
-            output,
-            attack_ips,
-            normal_ips,
-            heuristics,
-            k,
-            threshold,
-            window_secs,
-            min_events,
-        }),
-        Commands::TrainScanner {
-            input,
-            output,
-            wordlists,
-            threshold,
-            csic,
-        } => scanner::train::run(scanner::train::TrainScannerArgs {
-            input,
-            output,
-            wordlists,
-            threshold,
-            csic,
-        }),
         Commands::DownloadDatasets => {
             sunbeam_proxy::dataset::download::download_all()
         },
-        Commands::PrepareDataset { input, owasp, wordlists, output, seed, heuristics } => {
+        Commands::PrepareDataset { input, owasp, wordlists, output, seed, heuristics, inject_csic, inject_modsec } => {
             sunbeam_proxy::dataset::prepare::run(sunbeam_proxy::dataset::prepare::PrepareDatasetArgs {
-                input, owasp, wordlists, output, seed, heuristics,
+                input, owasp, wordlists, output, seed, heuristics, inject_csic, inject_modsec,
             })
         },
         #[cfg(feature = "training")]
-        Commands::TrainMlpScanner { dataset, output_dir, hidden_dim, epochs, learning_rate, batch_size, tree_max_depth, tree_min_purity } => {
+        Commands::TrainMlpScanner { dataset, output_dir, hidden_dim, epochs, learning_rate, batch_size, tree_max_depth, tree_min_purity, min_samples_leaf, cookie_weight } => {
             sunbeam_proxy::training::train_scanner::run(sunbeam_proxy::training::train_scanner::TrainScannerMlpArgs {
-                dataset_path: dataset, output_dir, hidden_dim, epochs, learning_rate, batch_size, tree_max_depth, tree_min_purity,
+                dataset_path: dataset, output_dir, hidden_dim, epochs, learning_rate, batch_size, tree_max_depth, tree_min_purity, min_samples_leaf, cookie_weight,
             })
         },
         #[cfg(feature = "training")]
-        Commands::TrainMlpDdos { dataset, output_dir, hidden_dim, epochs, learning_rate, batch_size, tree_max_depth, tree_min_purity } => {
+        Commands::TrainMlpDdos { dataset, output_dir, hidden_dim, epochs, learning_rate, batch_size, tree_max_depth, tree_min_purity, min_samples_leaf, cookie_weight } => {
             sunbeam_proxy::training::train_ddos::run(sunbeam_proxy::training::train_ddos::TrainDdosMlpArgs {
-                dataset_path: dataset, output_dir, hidden_dim, epochs, learning_rate, batch_size, tree_max_depth, tree_min_purity,
+                dataset_path: dataset, output_dir, hidden_dim, epochs, learning_rate, batch_size, tree_max_depth, tree_min_purity, min_samples_leaf, cookie_weight,
             })
         },
-        Commands::AutotuneDdos {
-            input,
-            output,
-            trials,
-            beta,
-            trial_log,
-        } => autotune::ddos::run_autotune(autotune::ddos::AutotuneDdosArgs {
-            input,
-            output,
-            trials,
-            beta,
-            trial_log,
-        }),
-        Commands::AutotuneScanner {
-            input,
-            output,
-            wordlists,
-            csic,
-            trials,
-            beta,
-            trial_log,
-        } => autotune::scanner::run_autotune(autotune::scanner::AutotuneScannerArgs {
-            input,
-            output,
-            wordlists,
-            csic,
-            trials,
-            beta,
-            trial_log,
-        }),
+        #[cfg(feature = "training")]
+        Commands::SweepCookieWeight { dataset, detector, weights, tree_max_depth, tree_min_purity, min_samples_leaf } => {
+            sunbeam_proxy::training::sweep::run_cookie_sweep(
+                &dataset, &detector, weights.as_deref(), tree_max_depth, tree_min_purity, min_samples_leaf,
+            )
+        },
     }
 }
 
@@ -363,46 +216,19 @@ fn run_serve(upgrade: bool) -> Result<()> {
     // 1b. Spawn metrics HTTP server (needs a tokio runtime for the TCP listener).
     let metrics_port = cfg.telemetry.metrics_port;
 
-    // 2. Load DDoS detection model if configured.
+    // 2. Init DDoS detector if configured (ensemble: compiled-in weights).
     let ddos_detector = if let Some(ddos_cfg) = &cfg.ddos {
         if ddos_cfg.enabled {
-            if ddos_cfg.use_ensemble {
-                // Ensemble path: compiled-in weights, no model file needed.
-                // We still need a TrainedModel for the struct, but it won't be used.
-                let dummy_model = ddos::model::TrainedModel::empty(ddos_cfg.k, ddos_cfg.threshold);
-                let detector = Arc::new(ddos::detector::DDoSDetector::new_ensemble(dummy_model, ddos_cfg));
-                tracing::info!(
-                    k = ddos_cfg.k,
-                    threshold = ddos_cfg.threshold,
-                    "DDoS ensemble detector enabled"
-                );
-                Some(detector)
-            } else if let Some(ref model_path) = ddos_cfg.model_path {
-                match ddos::model::TrainedModel::load(
-                    std::path::Path::new(model_path),
-                    Some(ddos_cfg.k),
-                    Some(ddos_cfg.threshold),
-                ) {
-                    Ok(model) => {
-                        let point_count = model.point_count();
-                        let detector = Arc::new(ddos::detector::DDoSDetector::new(model, ddos_cfg));
-                        tracing::info!(
-                            points = point_count,
-                            k = ddos_cfg.k,
-                            threshold = ddos_cfg.threshold,
-                            "DDoS detector loaded"
-                        );
-                        Some(detector)
-                    }
-                    Err(e) => {
-                        tracing::warn!(error = %e, "failed to load DDoS model; detection disabled");
-                        None
-                    }
-                }
-            } else {
-                tracing::warn!("DDoS enabled but no model_path and use_ensemble=false; detection disabled");
-                None
+            let detector = Arc::new(sunbeam_proxy::ddos::detector::DDoSDetector::new(ddos_cfg));
+            tracing::info!(
+                threshold = ddos_cfg.threshold,
+                observe_only = ddos_cfg.observe_only,
+                "DDoS ensemble detector enabled"
+            );
+            if ddos_cfg.observe_only {
+                tracing::warn!("DDoS detector in OBSERVE-ONLY mode — decisions are logged but traffic is never blocked");
             }
+            Some(detector)
         } else {
             None
         }
@@ -435,88 +261,35 @@ fn run_serve(upgrade: bool) -> Result<()> {
         None
     };
 
-    // 2c. Load scanner model if configured.
+    // 2c. Init scanner detector if configured (ensemble: compiled-in weights).
     let (scanner_detector, bot_allowlist) = if let Some(scanner_cfg) = &cfg.scanner {
         if scanner_cfg.enabled {
-            if scanner_cfg.use_ensemble {
-                // Ensemble path: compiled-in weights, no model file needed.
-                let detector = scanner::detector::ScannerDetector::new_ensemble(&cfg.routes);
-                let handle = Arc::new(arc_swap::ArcSwap::from_pointee(detector));
+            let detector = scanner::detector::ScannerDetector::new(&cfg.routes);
+            let handle = Arc::new(arc_swap::ArcSwap::from_pointee(detector));
 
-                // Start bot allowlist if rules are configured.
-                let bot_allowlist = if !scanner_cfg.allowlist.is_empty() {
-                    let al = scanner::allowlist::BotAllowlist::spawn(
-                        &scanner_cfg.allowlist,
-                        scanner_cfg.bot_cache_ttl_secs,
-                    );
-                    tracing::info!(
-                        rules = scanner_cfg.allowlist.len(),
-                        "bot allowlist enabled"
-                    );
-                    Some(al)
-                } else {
-                    None
-                };
-
-                tracing::info!(
-                    threshold = scanner_cfg.threshold,
-                    "scanner ensemble detector enabled"
+            let bot_allowlist = if !scanner_cfg.allowlist.is_empty() {
+                let al = scanner::allowlist::BotAllowlist::spawn(
+                    &scanner_cfg.allowlist,
+                    scanner_cfg.bot_cache_ttl_secs,
                 );
-                (Some(handle), bot_allowlist)
-            } else if let Some(ref model_path) = scanner_cfg.model_path {
-                match scanner::model::ScannerModel::load(std::path::Path::new(model_path)) {
-                    Ok(mut model) => {
-                        let fragment_count = model.fragments.len();
-                        model.threshold = scanner_cfg.threshold;
-                        let detector = scanner::detector::ScannerDetector::new(&model, &cfg.routes);
-                        let handle = Arc::new(arc_swap::ArcSwap::from_pointee(detector));
-
-                        // Start bot allowlist if rules are configured.
-                        let bot_allowlist = if !scanner_cfg.allowlist.is_empty() {
-                            let al = scanner::allowlist::BotAllowlist::spawn(
-                                &scanner_cfg.allowlist,
-                                scanner_cfg.bot_cache_ttl_secs,
-                            );
-                            tracing::info!(
-                                rules = scanner_cfg.allowlist.len(),
-                                "bot allowlist enabled"
-                            );
-                            Some(al)
-                        } else {
-                            None
-                        };
-
-                        // Start background file watcher for hot-reload.
-                        if scanner_cfg.poll_interval_secs > 0 {
-                            let watcher_handle = handle.clone();
-                            let watcher_model_path = std::path::PathBuf::from(model_path);
-                            let threshold = scanner_cfg.threshold;
-                            let routes = cfg.routes.clone();
-                            let interval = std::time::Duration::from_secs(scanner_cfg.poll_interval_secs);
-                            std::thread::spawn(move || {
-                                scanner::watcher::watch_scanner_model(
-                                    watcher_handle, watcher_model_path, threshold, routes, interval,
-                                );
-                            });
-                        }
-
-                        tracing::info!(
-                            fragments = fragment_count,
-                            threshold = scanner_cfg.threshold,
-                            poll_interval_secs = scanner_cfg.poll_interval_secs,
-                            "scanner detector loaded"
-                        );
-                        (Some(handle), bot_allowlist)
-                    }
-                    Err(e) => {
-                        tracing::warn!(error = %e, "failed to load scanner model; scanner detection disabled");
-                        (None, None)
-                    }
-                }
+                tracing::info!(
+                    rules = scanner_cfg.allowlist.len(),
+                    "bot allowlist enabled"
+                );
+                Some(al)
             } else {
-                tracing::warn!("scanner enabled but no model_path and use_ensemble=false; scanner detection disabled");
-                (None, None)
+                None
+            };
+
+            tracing::info!(
+                threshold = scanner_cfg.threshold,
+                observe_only = scanner_cfg.observe_only,
+                "scanner ensemble detector enabled"
+            );
+            if scanner_cfg.observe_only {
+                tracing::warn!("scanner detector in OBSERVE-ONLY mode — decisions are logged but traffic is never blocked");
             }
+            (Some(handle), bot_allowlist)
         } else {
             (None, None)
         }
@@ -617,6 +390,8 @@ fn run_serve(upgrade: bool) -> Result<()> {
             &cfg.rate_limit.as_ref().map(|rl| rl.bypass_cidrs.clone()).unwrap_or_default(),
         ),
         cluster: cluster_handle,
+        ddos_observe_only: cfg.ddos.as_ref().map(|d| d.observe_only).unwrap_or(false),
+        scanner_observe_only: cfg.scanner.as_ref().map(|s| s.observe_only).unwrap_or(false),
     };
     let mut svc = http_proxy_service(&server.configuration, proxy);
 
diff --git a/src/proxy.rs b/src/proxy.rs
index b515a16..0511b75 100644
--- a/src/proxy.rs
+++ b/src/proxy.rs
@@ -1,3 +1,6 @@
+// Copyright Sunbeam Studios 2026
+// SPDX-License-Identifier: Apache-2.0
+
 use crate::acme::AcmeRoutes;
 use crate::cluster::ClusterHandle;
 use crate::config::RouteConfig;
@@ -32,9 +35,9 @@ pub struct SunbeamProxy {
     pub routes: Vec<RouteConfig>,
     /// Per-challenge route table populated by the Ingress watcher.
     pub acme_routes: AcmeRoutes,
-    /// Optional KNN-based DDoS detector.
+    /// Optional DDoS detector (ensemble: decision tree + MLP).
     pub ddos_detector: Option<Arc<DDoSDetector>>,
-    /// Optional per-request scanner detector (hot-reloadable via ArcSwap).
+    /// Optional per-request scanner detector (ensemble: decision tree + MLP).
     pub scanner_detector: Option<Arc<ArcSwap<ScannerDetector>>>,
     /// Optional verified-bot allowlist (bypasses scanner for known crawlers/agents).
     pub bot_allowlist: Option<Arc<BotAllowlist>>,
@@ -48,6 +51,10 @@ pub struct SunbeamProxy {
     pub pipeline_bypass_cidrs: Vec<crate::rate_limit::cidr::CidrBlock>,
     /// Optional cluster handle for multi-node bandwidth tracking.
     pub cluster: Option<Arc<ClusterHandle>>,
+    /// When true, DDoS detector logs decisions but never blocks traffic.
+    pub ddos_observe_only: bool,
+    /// When true, scanner detector logs decisions but never blocks traffic.
+    pub scanner_observe_only: bool,
 }
 
 pub struct RequestCtx {
@@ -341,7 +348,7 @@ impl ProxyHttp for SunbeamProxy {
 
                 metrics::DDOS_DECISIONS.with_label_values(&[decision]).inc();
 
-                if matches!(ddos_action, DDoSAction::Block) {
+                if matches!(ddos_action, DDoSAction::Block) && !self.ddos_observe_only {
                     let mut resp = ResponseHeader::build(429, None)?;
                     resp.insert_header("Retry-After", "60")?;
                     resp.insert_header("Content-Length", "0")?;
@@ -426,7 +433,7 @@ impl ProxyHttp for SunbeamProxy {
                 .with_label_values(&[decision, reason])
                 .inc();
 
-            if decision == "block" {
+            if decision == "block" && !self.scanner_observe_only {
                 let mut resp = ResponseHeader::build(403, None)?;
                 resp.insert_header("Content-Length", "0")?;
                 session.write_response_header(Box::new(resp), true).await?;
@@ -1150,6 +1157,21 @@ impl ProxyHttp for SunbeamProxy {
             .and_then(|v| v.to_str().ok())
             .unwrap_or("-");
         let query = session.req_header().uri.query().unwrap_or("");
+        let response_bytes = session.body_bytes_sent();
+        let http_version = format!("{:?}", session.req_header().version);
+        let header_count = session.req_header().headers.len() as u16;
+        let accept_encoding = session
+            .req_header()
+            .headers
+            .get("accept-encoding")
+            .and_then(|v| v.to_str().ok())
+            .unwrap_or("-");
+        let connection = session
+            .req_header()
+            .headers
+            .get("connection")
+            .and_then(|v| v.to_str().ok())
+            .unwrap_or("-");
 
         tracing::info!(
             target = "audit",
@@ -1162,14 +1184,19 @@ impl ProxyHttp for SunbeamProxy {
             status,
             duration_ms,
             content_length,
+            response_bytes,
             user_agent,
             referer,
             accept_language,
             accept,
+            accept_encoding,
             has_cookies,
             cf_country,
             backend,
             error   = error_str,
+            http_version,
+            header_count,
+            connection,
             "request"
         );
 
diff --git a/src/scanner/csic.rs b/src/scanner/csic.rs
index 314a7fb..ad5c780 100644
--- a/src/scanner/csic.rs
+++ b/src/scanner/csic.rs
@@ -1,3 +1,6 @@
+// Copyright Sunbeam Studios 2026
+// SPDX-License-Identifier: Apache-2.0
+
 //! Fetch and convert the CSIC 2010 HTTP dataset into labeled training samples.
 //!
 //! The CSIC 2010 dataset contains raw HTTP/1.1 requests (normal + anomalous)
@@ -65,6 +68,7 @@ struct ParsedRequest {
     content_length: u64,
     referer: String,
     accept_language: String,
+    accept: String,
 }
 
 fn parse_csic_content(content: &str) -> Vec<ParsedRequest> {
@@ -158,6 +162,7 @@ fn parse_single_request(lines: &[&str]) -> Option<ParsedRequest> {
         content_length,
         referer: get_header("Referer").unwrap_or("-").to_string(),
         accept_language: get_header("Accept-Language").unwrap_or("-").to_string(),
+        accept: get_header("Accept").unwrap_or("-").to_string(),
     })
 }
 
@@ -219,11 +224,12 @@ fn to_audit_fields(
     // For anomalous samples, simulate real scanner behavior:
     // strip cookies/referer/accept-language that CSIC attacks have from their session.
     let (has_cookies, referer, accept_language, user_agent) = if label != "normal" {
-        let referer = None;
+        let referer = "-".to_string();
         let accept_language = if rng.next_f64() < 0.8 {
-            None
+            "-".to_string()
         } else {
-            Some(req.accept_language.clone()).filter(|a| a != "-")
+            let al = req.accept_language.clone();
+            if al == "-" { "-".to_string() } else { al }
         };
         let r = rng.next_f64();
         let user_agent = if r < 0.15 {
@@ -241,12 +247,26 @@ fn to_audit_fields(
     } else {
         (
             req.has_cookies,
-            Some(req.referer.clone()).filter(|r| r != "-"),
-            Some(req.accept_language.clone()).filter(|a| a != "-"),
+            if req.referer == "-" { "-".to_string() } else { req.referer.clone() },
+            if req.accept_language == "-" { "-".to_string() } else { req.accept_language.clone() },
             req.user_agent.clone(),
         )
     };
 
+    // For normal traffic, preserve Accept header from CSIC request.
+    // For attacks, degrade it to simulate scanner behavior.
+    let accept = if label == "normal" {
+        if req.accept == "-" || req.accept.is_empty() {
+            "text/html,application/xhtml+xml,application/xml;q=0.9,*/*;q=0.8".to_string()
+        } else {
+            req.accept.clone()
+        }
+    } else if rng.next_f64() < 0.6 {
+        "*/*".to_string()
+    } else {
+        req.accept.clone()
+    };
+
     AuditFields {
         method: req.method.clone(),
         host,
@@ -263,9 +283,10 @@ fn to_audit_fields(
         duration_ms: rng.next_usize(50) as u64 + 1,
         content_length: req.content_length,
         user_agent,
-        has_cookies: Some(has_cookies),
+        has_cookies,
         referer,
         accept_language,
+        accept,
         backend: if label == "normal" {
             format!("{host_prefix}-svc:8080")
         } else {
@@ -274,6 +295,7 @@ fn to_audit_fields(
         label: Some(
             if label == "normal" { "normal" } else { "attack" }.to_string(),
         ),
+        ..AuditFields::default()
     }
 }
 
@@ -343,6 +365,7 @@ mod tests {
         assert_eq!(req.path, "/index.html");
         assert!(req.has_cookies);
         assert_eq!(req.user_agent, "Mozilla/5.0");
+        assert_eq!(req.accept, "text/html");
     }
 
     #[test]
@@ -374,11 +397,12 @@ mod tests {
             content_length: 100,
             referer: "https://example.com".to_string(),
             accept_language: "en-US".to_string(),
+            accept: "text/html".to_string(),
         };
         let mut rng = Rng::new(42);
         let fields = to_audit_fields(&req, "normal", DEFAULT_HOSTS, &mut rng);
         assert_eq!(fields.label.as_deref(), Some("normal"));
-        assert!(fields.has_cookies.unwrap_or(false));
+        assert!(fields.has_cookies);
         assert!(fields.host.ends_with(".sunbeam.pt"));
     }
 
@@ -393,11 +417,12 @@ mod tests {
             content_length: 0,
             referer: "https://example.com".to_string(),
             accept_language: "en-US".to_string(),
+            accept: "text/html".to_string(),
         };
         let mut rng = Rng::new(42);
         let fields = to_audit_fields(&req, "anomalous", DEFAULT_HOSTS, &mut rng);
         assert_eq!(fields.label.as_deref(), Some("attack"));
-        assert!(!fields.has_cookies.unwrap_or(true));
+        assert!(!fields.has_cookies);
     }
 
     #[test]
diff --git a/src/scanner/detector.rs b/src/scanner/detector.rs
index aeb3baa..2d7ac67 100644
--- a/src/scanner/detector.rs
+++ b/src/scanner/detector.rs
@@ -1,9 +1,9 @@
+// Copyright Sunbeam Studios 2026
+// SPDX-License-Identifier: Apache-2.0
+
 use crate::config::RouteConfig;
-use crate::scanner::features::{
-    self, fx_hash_bytes, ScannerNormParams, SUSPICIOUS_EXTENSIONS_LIST, NUM_SCANNER_FEATURES,
-    NUM_SCANNER_WEIGHTS,
-};
-use crate::scanner::model::{ScannerAction, ScannerModel, ScannerVerdict};
+use crate::scanner::features::{self, fx_hash_bytes, SUSPICIOUS_EXTENSIONS_LIST};
+use crate::scanner::model::{ScannerAction, ScannerVerdict};
 use rustc_hash::FxHashSet;
 
 /// Immutable, zero-state per-request scanner detector.
@@ -12,44 +12,10 @@ pub struct ScannerDetector {
     fragment_hashes: FxHashSet<u64>,
     extension_hashes: FxHashSet<u64>,
     configured_hosts: FxHashSet<u64>,
-    weights: [f64; NUM_SCANNER_WEIGHTS],
-    threshold: f64,
-    norm_params: ScannerNormParams,
-    use_ensemble: bool,
 }
 
 impl ScannerDetector {
-    pub fn new(model: &ScannerModel, routes: &[RouteConfig]) -> Self {
-        let fragment_hashes: FxHashSet<u64> = model
-            .fragments
-            .iter()
-            .map(|f| fx_hash_bytes(f.to_ascii_lowercase().as_bytes()))
-            .collect();
-
-        let extension_hashes: FxHashSet<u64> = SUSPICIOUS_EXTENSIONS_LIST
-            .iter()
-            .map(|e| fx_hash_bytes(e.as_bytes()))
-            .collect();
-
-        let configured_hosts: FxHashSet<u64> = routes
-            .iter()
-            .map(|r| fx_hash_bytes(r.host_prefix.as_bytes()))
-            .collect();
-
-        Self {
-            fragment_hashes,
-            extension_hashes,
-            configured_hosts,
-            weights: model.weights,
-            threshold: model.threshold,
-            norm_params: model.norm_params.clone(),
-            use_ensemble: false,
-        }
-    }
-
-    /// Create a detector that uses the ensemble (decision tree + MLP) path
-    /// instead of the linear model. No model file needed — weights are compiled in.
-    pub fn new_ensemble(routes: &[RouteConfig]) -> Self {
+    pub fn new(routes: &[RouteConfig]) -> Self {
         let fragment_hashes: FxHashSet<u64> = crate::scanner::train::DEFAULT_FRAGMENTS
             .iter()
             .map(|f| fx_hash_bytes(f.to_ascii_lowercase().as_bytes()))
@@ -69,13 +35,6 @@ impl ScannerDetector {
             fragment_hashes,
             extension_hashes,
             configured_hosts,
-            weights: [0.0; NUM_SCANNER_WEIGHTS],
-            threshold: 0.5,
-            norm_params: ScannerNormParams {
-                mins: [0.0; NUM_SCANNER_FEATURES],
-                maxs: [1.0; NUM_SCANNER_FEATURES],
-            },
-            use_ensemble: true,
         }
     }
 
@@ -98,8 +57,6 @@ impl ScannerDetector {
         content_length: u64,
     ) -> ScannerVerdict {
         // Hard allowlist: obviously legitimate traffic bypasses the model.
-        // This prevents model drift from ever blocking real users and ensures
-        // the training pipeline always has clean positive labels.
         let host_known = {
             let hash = features::fx_hash_bytes(host_prefix.as_bytes());
             self.configured_hosts.contains(&hash)
@@ -121,95 +78,32 @@ impl ScannerDetector {
             };
         }
 
-        if self.use_ensemble {
-            // Ensemble path: extract f32 features → decision tree + MLP.
-            let raw_f32 = features::extract_features_f32(
-                method, path, host_prefix,
-                has_cookies, has_referer, has_accept_language,
-                accept, user_agent, content_length,
-                &self.fragment_hashes, &self.extension_hashes, &self.configured_hosts,
-            );
-            let ev = crate::ensemble::scanner::scanner_ensemble_predict(&raw_f32);
-            crate::metrics::SCANNER_ENSEMBLE_PATH
-                .with_label_values(&[match ev.path {
-                    crate::ensemble::scanner::EnsemblePath::TreeBlock => "tree_block",
-                    crate::ensemble::scanner::EnsemblePath::TreeAllow => "tree_allow",
-                    crate::ensemble::scanner::EnsemblePath::Mlp => "mlp",
-                }])
-                .inc();
-            return ev.into();
-        }
-
-        // 1. Extract 12 features
-        let raw = features::extract_features(
-            method,
-            path,
-            host_prefix,
-            has_cookies,
-            has_referer,
-            has_accept_language,
-            accept,
-            user_agent,
-            content_length,
-            &self.fragment_hashes,
-            &self.extension_hashes,
-            &self.configured_hosts,
+        // Ensemble path: extract f32 features → decision tree + MLP.
+        let raw_f32 = features::extract_features_f32(
+            method, path, host_prefix,
+            has_cookies, has_referer, has_accept_language,
+            accept, user_agent, content_length,
+            &self.fragment_hashes, &self.extension_hashes, &self.configured_hosts,
         );
-
-        // 2. Normalize
-        let f = self.norm_params.normalize(&raw);
-
-        // 3. Compute score = bias + dot(weights, features) + interaction terms
-        let mut score = self.weights[NUM_SCANNER_FEATURES + 2]; // bias (index 14)
-        for (i, &fi) in f.iter().enumerate().take(NUM_SCANNER_FEATURES) {
-            score += self.weights[i] * fi;
-        }
-        // Interaction: suspicious_path AND no_cookies
-        score += self.weights[12] * f[0] * (1.0 - f[3]);
-        // Interaction: unknown_host AND no_accept_language
-        score += self.weights[13] * (1.0 - f[9]) * (1.0 - f[5]);
-
-        // 4. Threshold
-        let action = if score > self.threshold {
-            ScannerAction::Block
-        } else {
-            ScannerAction::Allow
-        };
-
-        ScannerVerdict {
-            action,
-            score,
-            reason: "model",
-        }
+        let ev = crate::ensemble::scanner::scanner_ensemble_predict(&raw_f32);
+        crate::metrics::SCANNER_ENSEMBLE_PATH
+            .with_label_values(&[match ev.path {
+                crate::ensemble::scanner::EnsemblePath::TreeBlock => "tree_block",
+                crate::ensemble::scanner::EnsemblePath::TreeAllow => "tree_allow",
+                crate::ensemble::scanner::EnsemblePath::Mlp => "mlp",
+            }])
+            .inc();
+        ev.into()
     }
 }
 
 #[cfg(test)]
 mod tests {
     use super::*;
-    use crate::scanner::features::NUM_SCANNER_FEATURES;
+    use crate::config::RouteConfig;
 
-    fn make_detector(weights: [f64; NUM_SCANNER_WEIGHTS], threshold: f64) -> ScannerDetector {
-        let model = ScannerModel {
-            weights,
-            threshold,
-            norm_params: ScannerNormParams {
-                mins: [0.0; NUM_SCANNER_FEATURES],
-                maxs: [1.0; NUM_SCANNER_FEATURES],
-            },
-            fragments: vec![
-                ".env".into(),
-                "wp-admin".into(),
-                "wp-login".into(),
-                "phpinfo".into(),
-                "phpmyadmin".into(),
-                ".git".into(),
-                "cgi-bin".into(),
-                ".htaccess".into(),
-                ".htpasswd".into(),
-            ],
-        };
-        let routes = vec![RouteConfig {
+    fn test_routes() -> Vec<RouteConfig> {
+        vec![RouteConfig {
             host_prefix: "app".into(),
             backend: "http://127.0.0.1:8080".into(),
             websocket: false,
@@ -221,35 +115,12 @@ mod tests {
             body_rewrites: vec![],
             response_headers: vec![],
             cache: None,
-        }];
-        ScannerDetector::new(&model, &routes)
-    }
-
-    /// Weights tuned to block scanner-like requests:
-    /// High weight on suspicious_path (w[0]), no_cookies interaction (w[12]),
-    /// has_suspicious_extension (w[2]), traversal (w[11]).
-    /// Negative weight on has_cookies (w[3]), has_referer (w[4]),
-    /// accept_quality (w[6]), ua_category (w[7]), host_is_configured (w[9]).
-    fn attack_tuned_weights() -> [f64; NUM_SCANNER_WEIGHTS] {
-        let mut w = [0.0; NUM_SCANNER_WEIGHTS];
-        w[0] = 2.0;  // suspicious_path_score
-        w[2] = 2.0;  // has_suspicious_extension
-        w[3] = -2.0; // has_cookies (negative = good)
-        w[4] = -1.0; // has_referer (negative = good)
-        w[5] = -1.0; // has_accept_language (negative = good)
-        w[6] = -0.5; // accept_quality (negative = good)
-        w[7] = -1.0; // ua_category (negative = browser is good)
-        w[9] = -1.5; // host_is_configured (negative = known host is good)
-        w[11] = 2.0; // path_has_traversal
-        w[12] = 1.5; // interaction: suspicious_path AND no_cookies
-        w[13] = 1.0; // interaction: unknown_host AND no_accept_lang
-        w[14] = 0.5; // bias
-        w
+        }]
     }
 
     #[test]
     fn test_normal_browser_request_allowed() {
-        let detector = make_detector(attack_tuned_weights(), 0.5);
+        let detector = ScannerDetector::new(&test_routes());
         let verdict = detector.check(
             "GET",
             "/blog/hello-world",
@@ -267,7 +138,7 @@ mod tests {
 
     #[test]
     fn test_api_client_with_auth_allowed() {
-        let detector = make_detector(attack_tuned_weights(), 0.5);
+        let detector = ScannerDetector::new(&test_routes());
         let verdict = detector.check(
             "POST",
             "/api/v1/data",
@@ -285,81 +156,24 @@ mod tests {
 
     #[test]
     fn test_env_probe_blocked() {
-        let detector = make_detector(attack_tuned_weights(), 0.5);
+        let detector = ScannerDetector::new(&test_routes());
         let verdict = detector.check(
             "GET",
             "/.env",
             "unknown",
-            false, // no cookies
-            false, // no referer
-            false, // no accept-language
+            false,
+            false,
+            false,
             "*/*",
             "curl/7.0",
             0,
         );
         assert_eq!(verdict.action, ScannerAction::Block);
-        assert_eq!(verdict.reason, "model");
-    }
-
-    #[test]
-    fn test_wordpress_scan_blocked() {
-        let detector = make_detector(attack_tuned_weights(), 0.5);
-        let verdict = detector.check(
-            "GET",
-            "/wp-admin/install.php",
-            "unknown",
-            false,
-            false,
-            false,
-            "*/*",
-            "",
-            0,
-        );
-        assert_eq!(verdict.action, ScannerAction::Block);
-        assert_eq!(verdict.reason, "model");
-    }
-
-    #[test]
-    fn test_path_traversal_blocked() {
-        let detector = make_detector(attack_tuned_weights(), 0.5);
-        let verdict = detector.check(
-            "GET",
-            "/etc/../../../passwd",
-            "unknown",
-            false,
-            false,
-            false,
-            "*/*",
-            "python-requests/2.28",
-            0,
-        );
-        assert_eq!(verdict.action, ScannerAction::Block);
-        assert_eq!(verdict.reason, "model");
-    }
-
-    #[test]
-    fn test_legitimate_php_path_allowed() {
-        let detector = make_detector(attack_tuned_weights(), 0.5);
-        // "/blog/php-is-dead" — "php-is-dead" is not a known fragment
-        // has_cookies=true + known host "app" → hits allowlist
-        let verdict = detector.check(
-            "GET",
-            "/blog/php-is-dead",
-            "app",
-            true,
-            true,
-            true,
-            "text/html",
-            "Mozilla/5.0 Chrome/120",
-            0,
-        );
-        assert_eq!(verdict.action, ScannerAction::Allow);
     }
 
     #[test]
     fn test_allowlist_browser_on_known_host() {
-        let detector = make_detector(attack_tuned_weights(), 0.5);
-        // No cookies but browser UA + accept-language + known host → allowlist
+        let detector = ScannerDetector::new(&test_routes());
         let verdict = detector.check(
             "GET",
             "/",
@@ -374,22 +188,4 @@ mod tests {
         assert_eq!(verdict.action, ScannerAction::Allow);
         assert_eq!(verdict.reason, "allowlist:host+browser");
     }
-
-    #[test]
-    fn test_model_path_for_non_allowlisted() {
-        let detector = make_detector(attack_tuned_weights(), 0.5);
-        // Unknown host, no cookies, curl UA → goes through model
-        let verdict = detector.check(
-            "GET",
-            "/robots.txt",
-            "unknown",
-            false,
-            false,
-            false,
-            "*/*",
-            "curl/7.0",
-            0,
-        );
-        assert_eq!(verdict.reason, "model");
-    }
 }
diff --git a/src/scanner/mod.rs b/src/scanner/mod.rs
index bfe6192..8778b44 100644
--- a/src/scanner/mod.rs
+++ b/src/scanner/mod.rs
@@ -1,7 +1,9 @@
+// Copyright Sunbeam Studios 2026
+// SPDX-License-Identifier: Apache-2.0
+
 pub mod allowlist;
 pub mod csic;
 pub mod detector;
 pub mod features;
 pub mod model;
 pub mod train;
-pub mod watcher;
diff --git a/src/scanner/model.rs b/src/scanner/model.rs
index cdd2bab..9507536 100644
--- a/src/scanner/model.rs
+++ b/src/scanner/model.rs
@@ -1,7 +1,5 @@
-use crate::scanner::features::{ScannerNormParams, NUM_SCANNER_WEIGHTS};
-use anyhow::{Context, Result};
-use serde::{Deserialize, Serialize};
-use std::path::Path;
+// Copyright Sunbeam Studios 2026
+// SPDX-License-Identifier: Apache-2.0
 
 #[derive(Debug, Clone, Copy, PartialEq, Eq)]
 pub enum ScannerAction {
@@ -16,74 +14,3 @@ pub struct ScannerVerdict {
     /// Why this decision was made: "model", "allowlist", etc.
     pub reason: &'static str,
 }
-
-#[derive(Debug, Clone, Serialize, Deserialize)]
-pub struct ScannerModel {
-    pub weights: [f64; NUM_SCANNER_WEIGHTS],
-    pub threshold: f64,
-    pub norm_params: ScannerNormParams,
-    /// Suspicious path fragments used during training — kept for reproducibility.
-    pub fragments: Vec<String>,
-}
-
-impl ScannerModel {
-    pub fn save(&self, path: &Path) -> Result<()> {
-        let data = bincode::serialize(self).context("serializing scanner model")?;
-        std::fs::write(path, data)
-            .with_context(|| format!("writing scanner model to {}", path.display()))?;
-        Ok(())
-    }
-
-    pub fn load(path: &Path) -> Result<Self> {
-        let data = std::fs::read(path)
-            .with_context(|| format!("reading scanner model from {}", path.display()))?;
-        bincode::deserialize(&data).context("deserializing scanner model")
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-    use crate::scanner::features::NUM_SCANNER_FEATURES;
-
-    #[test]
-    fn test_serialization_roundtrip() {
-        let model = ScannerModel {
-            weights: [0.1; NUM_SCANNER_WEIGHTS],
-            threshold: 0.5,
-            norm_params: ScannerNormParams {
-                mins: [0.0; NUM_SCANNER_FEATURES],
-                maxs: [1.0; NUM_SCANNER_FEATURES],
-            },
-            fragments: vec![".env".into(), "wp-admin".into()],
-        };
-        let data = bincode::serialize(&model).unwrap();
-        let loaded: ScannerModel = bincode::deserialize(&data).unwrap();
-        assert_eq!(loaded.weights, model.weights);
-        assert_eq!(loaded.threshold, model.threshold);
-        assert_eq!(loaded.fragments, model.fragments);
-    }
-
-    #[test]
-    fn test_save_load_file() {
-        let dir = std::env::temp_dir().join("scanner_model_test");
-        std::fs::create_dir_all(&dir).unwrap();
-        let path = dir.join("test_model.bin");
-
-        let model = ScannerModel {
-            weights: [0.5; NUM_SCANNER_WEIGHTS],
-            threshold: 0.42,
-            norm_params: ScannerNormParams {
-                mins: [0.0; NUM_SCANNER_FEATURES],
-                maxs: [1.0; NUM_SCANNER_FEATURES],
-            },
-            fragments: vec!["phpinfo".into()],
-        };
-        model.save(&path).unwrap();
-        let loaded = ScannerModel::load(&path).unwrap();
-        assert_eq!(loaded.threshold, 0.42);
-        assert_eq!(loaded.fragments, vec!["phpinfo"]);
-
-        let _ = std::fs::remove_dir_all(&dir);
-    }
-}
diff --git a/src/scanner/train.rs b/src/scanner/train.rs
index 0dc770b..81307bb 100644
--- a/src/scanner/train.rs
+++ b/src/scanner/train.rs
@@ -1,9 +1,30 @@
+// Copyright Sunbeam Studios 2026
+// SPDX-License-Identifier: Apache-2.0
+
 use crate::ddos::audit_log::{AuditLog, AuditFields};
 use crate::scanner::features::{
     self, fx_hash_bytes, ScannerFeatureVector, ScannerNormParams, NUM_SCANNER_FEATURES,
     NUM_SCANNER_WEIGHTS,
 };
-use crate::scanner::model::ScannerModel;
+use serde::{Deserialize, Serialize};
+
+/// Legacy linear scanner model — kept for the `train-scanner` CLI command.
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct ScannerModel {
+    pub weights: [f64; NUM_SCANNER_WEIGHTS],
+    pub threshold: f64,
+    pub norm_params: ScannerNormParams,
+    pub fragments: Vec<String>,
+}
+
+impl ScannerModel {
+    pub fn save(&self, path: &Path) -> Result<()> {
+        let data = bincode::serialize(self).context("serializing scanner model")?;
+        std::fs::write(path, data)
+            .with_context(|| format!("writing scanner model to {}", path.display()))?;
+        Ok(())
+    }
+}
 use anyhow::{Context, Result};
 use rustc_hash::FxHashSet;
 use std::io::BufRead;
@@ -88,17 +109,9 @@ pub fn train_and_evaluate(
     }
 
     for (fields, host_prefix) in &parsed_entries {
-        let has_cookies = fields.has_cookies.unwrap_or(false);
-        let has_referer = fields
-            .referer
-            .as_ref()
-            .map(|r| r != "-" && !r.is_empty())
-            .unwrap_or(false);
-        let has_accept_language = fields
-            .accept_language
-            .as_ref()
-            .map(|a| a != "-" && !a.is_empty())
-            .unwrap_or(false);
+        let has_cookies = fields.has_cookies;
+        let has_referer = !fields.referer.is_empty() && fields.referer != "-";
+        let has_accept_language = !fields.accept_language.is_empty() && fields.accept_language != "-";
 
         let feats = features::extract_features(
             &fields.method,
@@ -149,17 +162,9 @@ pub fn train_and_evaluate(
             log_hosts.insert(fx_hash_bytes(host_prefix.as_bytes()));
         }
         for (fields, host_prefix) in &csic_entries {
-            let has_cookies = fields.has_cookies.unwrap_or(false);
-            let has_referer = fields
-                .referer
-                .as_ref()
-                .map(|r| r != "-" && !r.is_empty())
-                .unwrap_or(false);
-            let has_accept_language = fields
-                .accept_language
-                .as_ref()
-                .map(|a| a != "-" && !a.is_empty())
-                .unwrap_or(false);
+            let has_cookies = fields.has_cookies;
+            let has_referer = !fields.referer.is_empty() && fields.referer != "-";
+            let has_accept_language = !fields.accept_language.is_empty() && fields.accept_language != "-";
 
             let feats = features::extract_features(
                 &fields.method,
@@ -288,17 +293,9 @@ pub fn run(args: TrainScannerArgs) -> Result<()> {
     }
 
     for (fields, host_prefix) in &parsed_entries {
-        let has_cookies = fields.has_cookies.unwrap_or(false);
-        let has_referer = fields
-            .referer
-            .as_ref()
-            .map(|r| r != "-" && !r.is_empty())
-            .unwrap_or(false);
-        let has_accept_language = fields
-            .accept_language
-            .as_ref()
-            .map(|a| a != "-" && !a.is_empty())
-            .unwrap_or(false);
+        let has_cookies = fields.has_cookies;
+        let has_referer = !fields.referer.is_empty() && fields.referer != "-";
+        let has_accept_language = !fields.accept_language.is_empty() && fields.accept_language != "-";
 
         let feats = features::extract_features(
             &fields.method,
@@ -352,17 +349,9 @@ pub fn run(args: TrainScannerArgs) -> Result<()> {
             log_hosts.insert(fx_hash_bytes(host_prefix.as_bytes()));
         }
         for (fields, host_prefix) in &csic_entries {
-            let has_cookies = fields.has_cookies.unwrap_or(false);
-            let has_referer = fields
-                .referer
-                .as_ref()
-                .map(|r| r != "-" && !r.is_empty())
-                .unwrap_or(false);
-            let has_accept_language = fields
-                .accept_language
-                .as_ref()
-                .map(|a| a != "-" && !a.is_empty())
-                .unwrap_or(false);
+            let has_cookies = fields.has_cookies;
+            let has_referer = !fields.referer.is_empty() && fields.referer != "-";
+            let has_accept_language = !fields.accept_language.is_empty() && fields.accept_language != "-";
 
             let feats = features::extract_features(
                 &fields.method,
diff --git a/src/scanner/watcher.rs b/src/scanner/watcher.rs
deleted file mode 100644
index 5660f99..0000000
--- a/src/scanner/watcher.rs
+++ /dev/null
@@ -1,51 +0,0 @@
-use crate::config::RouteConfig;
-use crate::scanner::detector::ScannerDetector;
-use crate::scanner::model::ScannerModel;
-use arc_swap::ArcSwap;
-use std::path::PathBuf;
-use std::sync::Arc;
-use std::time::Duration;
-
-/// Poll the scanner model file for mtime changes and hot-swap the detector.
-/// Runs forever on a dedicated OS thread — never returns.
-pub fn watch_scanner_model(
-    handle: Arc<ArcSwap<ScannerDetector>>,
-    model_path: PathBuf,
-    threshold: f64,
-    routes: Vec<RouteConfig>,
-    poll_interval: Duration,
-) {
-    let mut last_mtime = std::fs::metadata(&model_path)
-        .and_then(|m| m.modified())
-        .ok();
-
-    loop {
-        std::thread::sleep(poll_interval);
-
-        let current_mtime = match std::fs::metadata(&model_path).and_then(|m| m.modified()) {
-            Ok(t) => t,
-            Err(_) => continue,
-        };
-
-        if Some(current_mtime) == last_mtime {
-            continue;
-        }
-
-        match ScannerModel::load(&model_path) {
-            Ok(mut model) => {
-                model.threshold = threshold;
-                let fragment_count = model.fragments.len();
-                let detector = ScannerDetector::new(&model, &routes);
-                handle.store(Arc::new(detector));
-                last_mtime = Some(current_mtime);
-                tracing::info!(
-                    fragments = fragment_count,
-                    "scanner model hot-reloaded"
-                );
-            }
-            Err(e) => {
-                tracing::warn!(error = %e, "failed to reload scanner model; keeping current");
-            }
-        }
-    }
-}
diff --git a/src/training/batch.rs b/src/training/batch.rs
new file mode 100644
index 0000000..830234e
--- /dev/null
+++ b/src/training/batch.rs
@@ -0,0 +1,112 @@
+// Copyright Sunbeam Studios 2026
+// SPDX-License-Identifier: Apache-2.0
+
+//! Shared training infrastructure: Dataset adapter, Batcher, and batch types
+//! for use with burn's SupervisedTraining.
+
+use crate::dataset::sample::TrainingSample;
+
+use burn::data::dataloader::batcher::Batcher;
+use burn::data::dataloader::Dataset;
+use burn::prelude::*;
+
+/// A single normalized training item ready for batching.
+#[derive(Clone, Debug)]
+pub struct TrainingItem {
+    pub features: Vec<f32>,
+    pub label: i32,
+    pub weight: f32,
+}
+
+/// A batch of training items as tensors.
+#[derive(Clone, Debug)]
+pub struct TrainingBatch<B: Backend> {
+    pub features: Tensor<B, 2>,
+    pub labels: Tensor<B, 1, Int>,
+    pub weights: Tensor<B, 2>,
+}
+
+/// Wraps a `Vec<TrainingSample>` as a burn `Dataset`, applying min-max
+/// normalization to features at construction time.
+#[derive(Clone)]
+pub struct SampleDataset {
+    items: Vec<TrainingItem>,
+}
+
+impl SampleDataset {
+    pub fn new(samples: &[TrainingSample], mins: &[f32], maxs: &[f32]) -> Self {
+        let items = samples
+            .iter()
+            .map(|s| {
+                let features: Vec<f32> = s
+                    .features
+                    .iter()
+                    .enumerate()
+                    .map(|(i, &v)| {
+                        let range = maxs[i] - mins[i];
+                        if range > f32::EPSILON {
+                            ((v - mins[i]) / range).clamp(0.0, 1.0)
+                        } else {
+                            0.0
+                        }
+                    })
+                    .collect();
+                TrainingItem {
+                    features,
+                    label: if s.label >= 0.5 { 1 } else { 0 },
+                    weight: s.weight,
+                }
+            })
+            .collect();
+        Self { items }
+    }
+}
+
+impl Dataset<TrainingItem> for SampleDataset {
+    fn get(&self, index: usize) -> Option<TrainingItem> {
+        self.items.get(index).cloned()
+    }
+
+    fn len(&self) -> usize {
+        self.items.len()
+    }
+}
+
+/// Converts a `Vec<TrainingItem>` into a `TrainingBatch` of tensors.
+#[derive(Clone)]
+pub struct SampleBatcher;
+
+impl SampleBatcher {
+    pub fn new() -> Self {
+        Self
+    }
+}
+
+impl<B: Backend> Batcher<B, TrainingItem, TrainingBatch<B>> for SampleBatcher {
+    fn batch(&self, items: Vec<TrainingItem>, device: &B::Device) -> TrainingBatch<B> {
+        let batch_size = items.len();
+        let num_features = items[0].features.len();
+
+        let flat_features: Vec<f32> = items
+            .iter()
+            .flat_map(|item| item.features.iter().copied())
+            .collect();
+
+        let labels: Vec<i32> = items.iter().map(|item| item.label).collect();
+        let weights: Vec<f32> = items.iter().map(|item| item.weight).collect();
+
+        let features = Tensor::<B, 1>::from_floats(flat_features.as_slice(), device)
+            .reshape([batch_size, num_features]);
+
+        let labels = Tensor::<B, 1, Int>::from_ints(labels.as_slice(), device);
+
+        let weights = Tensor::<B, 1>::from_floats(weights.as_slice(), device)
+            .reshape([batch_size, 1]);
+
+        TrainingBatch {
+            features,
+            labels,
+            weights,
+        }
+    }
+}
diff --git a/src/training/export.rs b/src/training/export.rs
index 1b0af93..108c0cb 100644
--- a/src/training/export.rs
+++ b/src/training/export.rs
@@ -1,3 +1,6 @@
+// Copyright Sunbeam Studios 2026
+// SPDX-License-Identifier: Apache-2.0
+
 //! Weight export: converts trained models into standalone Rust `const` arrays
 //! and optionally Lean 4 definitions.
 //!
@@ -54,7 +57,7 @@ pub fn generate_rust_source(model: &ExportedModel) -> String {
     writeln!(s).unwrap();
 
     // Threshold.
-    writeln!(s, "pub const THRESHOLD: f32 = {:.8};", model.threshold).unwrap();
+    writeln!(s, "pub const THRESHOLD: f32 = {:.8};", sanitize(model.threshold)).unwrap();
     writeln!(s).unwrap();
 
     // Normalization params.
@@ -74,7 +77,7 @@ pub fn generate_rust_source(model: &ExportedModel) -> String {
             if i > 0 {
                 write!(s, ", ").unwrap();
             }
-            write!(s, "{:.8}", v).unwrap();
+            write!(s, "{:.8}", sanitize(*v)).unwrap();
         }
         writeln!(s, "],").unwrap();
     }
@@ -88,7 +91,7 @@ pub fn generate_rust_source(model: &ExportedModel) -> String {
     write_f32_array(&mut s, "W2", &model.w2);
 
     // B2.
-    writeln!(s, "pub const B2: f32 = {:.8};", model.b2).unwrap();
+    writeln!(s, "pub const B2: f32 = {:.8};", sanitize(model.b2)).unwrap();
     writeln!(s).unwrap();
 
     // Tree nodes.
@@ -207,6 +210,11 @@ pub fn export_to_file(model: &ExportedModel, path: &Path) -> Result<()> {
 // Helpers
 // ---------------------------------------------------------------------------
 
+/// Sanitize a float for Rust source: replace NaN/Inf with 0.0.
+fn sanitize(v: f32) -> f32 {
+    if v.is_finite() { v } else { 0.0 }
+}
+
 fn write_f32_array(s: &mut String, name: &str, values: &[f32]) {
     writeln!(s, "pub const {}: [f32; {}] = [", name, values.len()).unwrap();
     write!(s, "    ").unwrap();
@@ -218,7 +226,7 @@ fn write_f32_array(s: &mut String, name: &str, values: &[f32]) {
         if i > 0 && i % 8 == 0 {
             write!(s, "\n    ").unwrap();
         }
-        write!(s, "{:.8}", v).unwrap();
+        write!(s, "{:.8}", sanitize(*v)).unwrap();
     }
     writeln!(s, "\n];").unwrap();
     writeln!(s).unwrap();
diff --git a/src/training/mlp.rs b/src/training/mlp.rs
index 0172492..1a4fc5d 100644
--- a/src/training/mlp.rs
+++ b/src/training/mlp.rs
@@ -1,11 +1,18 @@
+// Copyright Sunbeam Studios 2026
+// SPDX-License-Identifier: Apache-2.0
+
 //! burn-rs MLP model definition for ensemble training.
 //!
 //! A two-layer network (linear -> ReLU -> linear -> sigmoid) used as the
 //! "uncertain region" classifier in the tree+MLP ensemble.
 
+use crate::training::batch::TrainingBatch;
+
 use burn::module::Module;
 use burn::nn::{Linear, LinearConfig};
 use burn::prelude::*;
+use burn::tensor::backend::AutodiffBackend;
+use burn::train::{ClassificationOutput, InferenceStep, TrainOutput, TrainStep};
 
 /// Two-layer MLP: input -> hidden (ReLU) -> output (sigmoid).
 #[derive(Module, Debug)]
@@ -34,24 +41,79 @@ impl MlpConfig {
 }
 
 impl<B: Backend> MlpModel<B> {
-    /// Forward pass: ReLU hidden activation, sigmoid output.
+    /// Forward pass returning raw logits (pre-sigmoid).
     ///
     /// Input shape: `[batch, input_dim]`
     /// Output shape: `[batch, 1]`
-    pub fn forward(&self, x: Tensor<B, 2>) -> Tensor<B, 2> {
+    pub fn forward_logits(&self, x: Tensor<B, 2>) -> Tensor<B, 2> {
         let h = self.linear1.forward(x);
         let h = burn::tensor::activation::relu(h);
-        let out = self.linear2.forward(h);
-        burn::tensor::activation::sigmoid(out)
+        self.linear2.forward(h)
+    }
+
+    /// Forward pass with sigmoid activation for inference/export.
+    ///
+    /// Input shape: `[batch, input_dim]`
+    /// Output shape: `[batch, 1]`  (values in [0, 1])
+    pub fn forward(&self, x: Tensor<B, 2>) -> Tensor<B, 2> {
+        burn::tensor::activation::sigmoid(self.forward_logits(x))
+    }
+
+    /// Forward pass returning a `ClassificationOutput` for burn's training loop.
+    ///
+    /// Uses raw logits for BCE (which applies sigmoid internally) and converts
+    /// to two-column format `[1-p, p]` for AccuracyMetric (which uses argmax).
+    pub fn forward_classification(
+        &self,
+        batch: TrainingBatch<B>,
+    ) -> ClassificationOutput<B> {
+        let logits = self.forward_logits(batch.features); // [batch, 1]
+        let logits_1d = logits.clone().squeeze::<1>(); // [batch]
+
+        // Numerically stable BCE from logits:
+        //   loss = max(logits, 0) - logits * targets + log(1 + exp(-|logits|))
+        // This avoids log(0) and exp(large) overflow.
+        let targets_float = batch.labels.clone().float(); // [batch]
+        let zeros = Tensor::zeros_like(&logits_1d);
+        let relu_logits = logits_1d.clone().max_pair(zeros); // max(logits, 0)
+        let neg_abs = logits_1d.clone().abs().neg(); // -|logits|
+        let log_term = neg_abs.exp().log1p(); // log(1 + exp(-|logits|))
+        let per_sample = relu_logits - logits_1d.clone() * targets_float + log_term;
+        let loss = per_sample.mean(); // scalar [1]
+
+        // AccuracyMetric expects [batch, num_classes] and uses argmax.
+        let neg_logits = logits.clone().neg();
+        let output_2col = Tensor::cat(vec![neg_logits, logits], 1); // [batch, 2]
+
+        ClassificationOutput::new(loss, output_2col, batch.labels)
+    }
+}
+
+impl<B: AutodiffBackend> TrainStep for MlpModel<B> {
+    type Input = TrainingBatch<B>;
+    type Output = ClassificationOutput<B>;
+
+    fn step(&self, batch: Self::Input) -> TrainOutput<Self::Output> {
+        let item = self.forward_classification(batch);
+        TrainOutput::new(self, item.loss.backward(), item)
+    }
+}
+
+impl<B: Backend> InferenceStep for MlpModel<B> {
+    type Input = TrainingBatch<B>;
+    type Output = ClassificationOutput<B>;
+
+    fn step(&self, batch: Self::Input) -> Self::Output {
+        self.forward_classification(batch)
     }
 }
 
 #[cfg(test)]
 mod tests {
     use super::*;
-    use burn::backend::NdArray;
+    use burn::backend::Wgpu;
 
-    type TestBackend = NdArray<f32>;
+    type TestBackend = Wgpu<f32, i32>;
 
     #[test]
     fn test_forward_pass_shape() {
@@ -80,7 +142,6 @@ mod tests {
         };
         let model = config.init::<TestBackend>(&device);
 
-        // Random-ish input values.
         let input = Tensor::<TestBackend, 2>::from_data(
             [[1.0, -2.0, 0.5, 3.0], [0.0, 0.0, 0.0, 0.0]],
             &device,
diff --git a/src/training/mod.rs b/src/training/mod.rs
index 1e82804..609b8fe 100644
--- a/src/training/mod.rs
+++ b/src/training/mod.rs
@@ -1,5 +1,10 @@
+// Copyright Sunbeam Studios 2026
+// SPDX-License-Identifier: Apache-2.0
+
 pub mod tree;
 pub mod mlp;
+pub mod batch;
 pub mod export;
 pub mod train_scanner;
 pub mod train_ddos;
+pub mod sweep;
diff --git a/src/training/sweep.rs b/src/training/sweep.rs
new file mode 100644
index 0000000..911a297
--- /dev/null
+++ b/src/training/sweep.rs
@@ -0,0 +1,103 @@
+// Copyright Sunbeam Studios 2026
+// SPDX-License-Identifier: Apache-2.0
+
+//! Cookie weight sweep: trains full tree+MLP ensembles (GPU via wgpu) across a
+//! range of cookie_weight values and reports accuracy metrics for each.
+
+use anyhow::{Context, Result};
+use std::path::Path;
+
+use crate::dataset::sample::load_dataset;
+use crate::training::train_scanner::TrainScannerMlpArgs;
+use crate::training::train_ddos::TrainDdosMlpArgs;
+
+/// Run a sweep across cookie_weight values for either scanner or ddos.
+///
+/// Each trial does a full GPU training run (tree + MLP) with the specified
+/// cookie_weight, writing artifacts to a temp directory.
+pub fn run_cookie_sweep(
+    dataset_path: &str,
+    detector: &str,
+    weights_csv: Option<&str>,
+    tree_max_depth: usize,
+    tree_min_purity: f32,
+    min_samples_leaf: usize,
+) -> Result<()> {
+    // Validate dataset exists and has samples.
+    let manifest = load_dataset(Path::new(dataset_path))
+        .context("loading dataset manifest")?;
+
+    let (cookie_idx, sample_count) = match detector {
+        "scanner" => (3usize, manifest.scanner_samples.len()),
+        "ddos" => (10usize, manifest.ddos_samples.len()),
+        other => anyhow::bail!("unknown detector '{}', expected 'scanner' or 'ddos'", other),
+    };
+
+    anyhow::ensure!(sample_count > 0, "no {} samples in dataset", detector);
+    drop(manifest); // Free memory before training loop.
+
+    let weights: Vec<f32> = if let Some(csv) = weights_csv {
+        csv.split(',')
+            .map(|s| s.trim().parse::<f32>())
+            .collect::<std::result::Result<Vec<_>, _>>()
+            .context("parsing --weights as comma-separated floats")?
+    } else {
+        (0..=10).map(|i| i as f32 / 10.0).collect()
+    };
+
+    println!(
+        "[sweep] {} detector, {} samples, cookie feature index: {}",
+        detector, sample_count, cookie_idx,
+    );
+    println!("[sweep] training {} trials with full tree+MLP (wgpu)\n", weights.len());
+
+    let sweep_dir = tempfile::tempdir().context("creating temp dir for sweep")?;
+
+    for (trial, &cw) in weights.iter().enumerate() {
+        let trial_dir = sweep_dir.path().join(format!("trial_{}", trial));
+        std::fs::create_dir_all(&trial_dir)?;
+        let trial_dir_str = trial_dir.to_string_lossy().to_string();
+
+        println!("━━━ Trial {}/{}: cookie_weight={:.2} ━━━", trial + 1, weights.len(), cw);
+
+        match detector {
+            "scanner" => {
+                crate::training::train_scanner::run(TrainScannerMlpArgs {
+                    dataset_path: dataset_path.to_string(),
+                    output_dir: trial_dir_str,
+                    hidden_dim: 32,
+                    epochs: 100,
+                    learning_rate: 0.0001,
+                    batch_size: 64,
+                    tree_max_depth,
+                    tree_min_purity,
+                    min_samples_leaf,
+                    cookie_weight: cw,
+                })?;
+            }
+            "ddos" => {
+                crate::training::train_ddos::run(TrainDdosMlpArgs {
+                    dataset_path: dataset_path.to_string(),
+                    output_dir: trial_dir_str,
+                    hidden_dim: 32,
+                    epochs: 100,
+                    learning_rate: 0.0001,
+                    batch_size: 64,
+                    tree_max_depth,
+                    tree_min_purity,
+                    min_samples_leaf,
+                    cookie_weight: cw,
+                })?;
+            }
+            _ => unreachable!(),
+        }
+
+        println!();
+    }
+
+    println!("[sweep] All {} trials complete.", weights.len());
+    println!("[sweep] Tip: compare tree structures and validation accuracy above.");
+    println!("[sweep] Look for a cookie_weight where FP rate drops without FN rate spiking.");
+
+    Ok(())
+}
diff --git a/src/training/train_ddos.rs b/src/training/train_ddos.rs
index b14fb6d..f143a9c 100644
--- a/src/training/train_ddos.rs
+++ b/src/training/train_ddos.rs
@@ -1,19 +1,27 @@
-//! DDoS MLP+tree training loop.
+// Copyright Sunbeam Studios 2026
+// SPDX-License-Identifier: Apache-2.0
+
+//! DDoS MLP+tree training loop using burn's SupervisedTraining.
 //!
-//! Loads a `DatasetManifest`, trains a CART decision tree and a burn-rs MLP,
-//! then exports the combined ensemble weights as a Rust source file that can
-//! be dropped into `src/ensemble/gen/ddos_weights.rs`.
+//! Loads a `DatasetManifest`, trains a CART decision tree and a burn-rs MLP
+//! with cosine annealing + early stopping, then exports the combined ensemble
+//! weights as a Rust source file for `src/ensemble/gen/ddos_weights.rs`.
 
 use anyhow::{Context, Result};
 use std::path::Path;
 
-use burn::backend::ndarray::NdArray;
 use burn::backend::Autodiff;
-use burn::module::AutodiffModule;
-use burn::optim::{AdamConfig, GradientsParams, Optimizer};
+use burn::backend::Wgpu;
+use burn::data::dataloader::DataLoaderBuilder;
+use burn::lr_scheduler::cosine::CosineAnnealingLrSchedulerConfig;
+use burn::optim::AdamConfig;
 use burn::prelude::*;
+use burn::record::CompactRecorder;
+use burn::train::metric::{AccuracyMetric, LossMetric};
+use burn::train::{Learner, SupervisedTraining};
 
 use crate::dataset::sample::{load_dataset, TrainingSample};
+use crate::training::batch::{SampleBatcher, SampleDataset};
 use crate::training::export::{export_to_file, ExportedModel};
 use crate::training::mlp::MlpConfig;
 use crate::training::tree::{train_tree, tree_predict, TreeConfig, TreeDecision};
@@ -21,7 +29,7 @@ use crate::training::tree::{train_tree, tree_predict, TreeConfig, TreeDecision};
 /// Number of DDoS features (matches `crate::ddos::features::NUM_FEATURES`).
 const NUM_FEATURES: usize = 14;
 
-type TrainBackend = Autodiff<NdArray<f32>>;
+type TrainBackend = Autodiff<Wgpu<f32, i32>>;
 
 /// Arguments for the DDoS MLP training command.
 pub struct TrainDdosMlpArgs {
@@ -37,10 +45,14 @@ pub struct TrainDdosMlpArgs {
     pub learning_rate: f64,
     /// Mini-batch size (default 64).
     pub batch_size: usize,
-    /// CART max depth (default 6).
+    /// CART max depth (default 8).
     pub tree_max_depth: usize,
-    /// CART leaf purity threshold (default 0.90).
+    /// CART leaf purity threshold (default 0.98).
     pub tree_min_purity: f32,
+    /// Minimum samples in a leaf node (default 2).
+    pub min_samples_leaf: usize,
+    /// Weight for cookie feature (feature 10: cookie_ratio). 0.0 = ignore, 1.0 = full weight.
+    pub cookie_weight: f32,
 }
 
 impl Default for TrainDdosMlpArgs {
@@ -50,14 +62,19 @@ impl Default for TrainDdosMlpArgs {
             output_dir: ".".into(),
             hidden_dim: 32,
             epochs: 100,
-            learning_rate: 0.001,
+            learning_rate: 0.0001,
             batch_size: 64,
-            tree_max_depth: 6,
-            tree_min_purity: 0.90,
+            tree_max_depth: 8,
+            tree_min_purity: 0.98,
+            min_samples_leaf: 2,
+            cookie_weight: 1.0,
         }
     }
 }
 
+/// Index of the cookie_ratio feature in the DDoS feature vector.
+const COOKIE_FEATURE_IDX: usize = 10;
+
 /// Entry point: train DDoS ensemble and export weights.
 pub fn run(args: TrainDdosMlpArgs) -> Result<()> {
     // 1. Load dataset.
@@ -86,6 +103,23 @@ pub fn run(args: TrainDdosMlpArgs) -> Result<()> {
     // 2. Compute normalization params from training data.
     let (norm_mins, norm_maxs) = compute_norm_params(samples);
 
+    if args.cookie_weight < 1.0 - f32::EPSILON {
+        println!(
+            "[ddos] cookie_weight={:.2} (feature {} influence reduced)",
+            args.cookie_weight, COOKIE_FEATURE_IDX,
+        );
+    }
+
+    // MLP norm adjustment: scale cookie feature's normalization range.
+    let mut mlp_norm_maxs = norm_maxs.clone();
+    if args.cookie_weight < 1.0 - f32::EPSILON {
+        let range = mlp_norm_maxs[COOKIE_FEATURE_IDX] - norm_mins[COOKIE_FEATURE_IDX];
+        if range > f32::EPSILON && args.cookie_weight > f32::EPSILON {
+            mlp_norm_maxs[COOKIE_FEATURE_IDX] =
+                range / args.cookie_weight + norm_mins[COOKIE_FEATURE_IDX];
+        }
+    }
+
     // 3. Stratified 80/20 split.
     let (train_set, val_set) = stratified_split(samples, 0.8);
     println!(
@@ -94,15 +128,16 @@ pub fn run(args: TrainDdosMlpArgs) -> Result<()> {
         val_set.len()
     );
 
-    // 4. Train CART tree.
+    // 4. Train CART tree (with cookie feature masking for reduced weight).
+    let tree_train_set = mask_cookie_feature(&train_set, COOKIE_FEATURE_IDX, args.cookie_weight);
     let tree_config = TreeConfig {
         max_depth: args.tree_max_depth,
-        min_samples_leaf: 5,
+        min_samples_leaf: args.min_samples_leaf,
         min_purity: args.tree_min_purity,
         num_features: NUM_FEATURES,
     };
-    let tree_nodes = train_tree(&train_set, &tree_config);
-    println!("[ddos] CART tree: {} nodes", tree_nodes.len());
+    let tree_nodes = train_tree(&tree_train_set, &tree_config);
+    println!("[ddos] CART tree: {} nodes (max_depth={})", tree_nodes.len(), args.tree_max_depth);
 
     // Evaluate tree on validation set.
     let (tree_correct, tree_deferred) = eval_tree(&tree_nodes, &val_set, &norm_mins, &norm_maxs);
@@ -112,23 +147,27 @@ pub fn run(args: TrainDdosMlpArgs) -> Result<()> {
         tree_deferred * 100.0,
     );
 
-    // 5. Train MLP on the full training set.
+    // 5. Train MLP with SupervisedTraining (uses mlp_norm_maxs for cookie scaling).
     let device = Default::default();
     let mlp_config = MlpConfig {
         input_dim: NUM_FEATURES,
         hidden_dim: args.hidden_dim,
     };
 
+    let artifact_dir = Path::new(&args.output_dir).join("ddos_artifacts");
+    std::fs::create_dir_all(&artifact_dir).ok();
+
     let model = train_mlp(
         &train_set,
         &val_set,
         &mlp_config,
         &norm_mins,
-        &norm_maxs,
+        &mlp_norm_maxs,
         args.epochs,
         args.learning_rate,
         args.batch_size,
         &device,
+        &artifact_dir,
     );
 
     // 6. Extract weights from trained model.
@@ -136,9 +175,9 @@ pub fn run(args: TrainDdosMlpArgs) -> Result<()> {
         &model,
         "ddos",
         &tree_nodes,
-        0.5, // threshold
+        0.5,
         &norm_mins,
-        &norm_maxs,
+        &mlp_norm_maxs,
         &device,
     );
 
@@ -153,6 +192,37 @@ pub fn run(args: TrainDdosMlpArgs) -> Result<()> {
     Ok(())
 }
 
+// ---------------------------------------------------------------------------
+// Cookie feature masking for CART trees
+// ---------------------------------------------------------------------------
+
+fn mask_cookie_feature(
+    samples: &[TrainingSample],
+    cookie_idx: usize,
+    cookie_weight: f32,
+) -> Vec<TrainingSample> {
+    if cookie_weight >= 1.0 - f32::EPSILON {
+        return samples.to_vec();
+    }
+    samples
+        .iter()
+        .enumerate()
+        .map(|(i, s)| {
+            let mut s2 = s.clone();
+            if cookie_weight < f32::EPSILON {
+                s2.features[cookie_idx] = 0.5;
+            } else {
+                let hash = (i as u64).wrapping_mul(6364136223846793005).wrapping_add(42);
+                let r = (hash >> 33) as f32 / (u32::MAX >> 1) as f32;
+                if r > cookie_weight {
+                    s2.features[cookie_idx] = 0.5;
+                }
+            }
+            s2
+        })
+        .collect()
+}
+
 // ---------------------------------------------------------------------------
 // Normalization
 // ---------------------------------------------------------------------------
@@ -170,21 +240,6 @@ fn compute_norm_params(samples: &[TrainingSample]) -> (Vec<f32>, Vec<f32>) {
     (mins, maxs)
 }
 
-fn normalize_features(features: &[f32], mins: &[f32], maxs: &[f32]) -> Vec<f32> {
-    features
-        .iter()
-        .enumerate()
-        .map(|(i, &v)| {
-            let range = maxs[i] - mins[i];
-            if range > f32::EPSILON {
-                ((v - mins[i]) / range).clamp(0.0, 1.0)
-            } else {
-                0.0
-            }
-        })
-        .collect()
-}
-
 // ---------------------------------------------------------------------------
 // Stratified split
 // ---------------------------------------------------------------------------
@@ -272,8 +327,23 @@ fn eval_tree(
     (accuracy, defer_rate)
 }
 
+fn normalize_features(features: &[f32], mins: &[f32], maxs: &[f32]) -> Vec<f32> {
+    features
+        .iter()
+        .enumerate()
+        .map(|(i, &v)| {
+            let range = maxs[i] - mins[i];
+            if range > f32::EPSILON {
+                ((v - mins[i]) / range).clamp(0.0, 1.0)
+            } else {
+                0.0
+            }
+        })
+        .collect()
+}
+
 // ---------------------------------------------------------------------------
-// MLP training
+// MLP training via SupervisedTraining
 // ---------------------------------------------------------------------------
 
 fn train_mlp(
@@ -286,117 +356,47 @@ fn train_mlp(
     learning_rate: f64,
     batch_size: usize,
     device: &<TrainBackend as Backend>::Device,
-) -> crate::training::mlp::MlpModel<NdArray<f32>> {
-    let mut model = config.init::<TrainBackend>(device);
-    let mut optim = AdamConfig::new().init();
+    artifact_dir: &Path,
+) -> crate::training::mlp::MlpModel<Wgpu<f32, i32>> {
+    let model = config.init::<TrainBackend>(device);
 
-    // Pre-normalize all training data.
-    let train_features: Vec<Vec<f32>> = train_set
-        .iter()
-        .map(|s| normalize_features(&s.features, mins, maxs))
-        .collect();
-    let train_labels: Vec<f32> = train_set.iter().map(|s| s.label).collect();
-    let train_weights: Vec<f32> = train_set.iter().map(|s| s.weight).collect();
+    let train_dataset = SampleDataset::new(train_set, mins, maxs);
+    let val_dataset = SampleDataset::new(val_set, mins, maxs);
 
-    let n = train_features.len();
+    let dataloader_train = DataLoaderBuilder::new(SampleBatcher::new())
+        .batch_size(batch_size)
+        .shuffle(42)
+        .num_workers(1)
+        .build(train_dataset);
 
-    for epoch in 0..epochs {
-        let mut epoch_loss = 0.0f32;
-        let mut batches = 0usize;
+    let dataloader_valid = DataLoaderBuilder::new(SampleBatcher::new())
+        .batch_size(batch_size)
+        .num_workers(1)
+        .build(val_dataset);
 
-        let mut offset = 0;
-        while offset < n {
-            let end = (offset + batch_size).min(n);
-            let batch_n = end - offset;
+    // Cosine annealing: initial_lr must be in (0.0, 1.0].
+    let lr = learning_rate.min(1.0);
+    let lr_scheduler = CosineAnnealingLrSchedulerConfig::new(lr, epochs)
+        .init()
+        .expect("valid cosine annealing config");
 
-            // Build input tensor [batch, features].
-            let flat: Vec<f32> = train_features[offset..end]
-                .iter()
-                .flat_map(|f| f.iter().copied())
-                .collect();
-            let x = Tensor::<TrainBackend, 1>::from_floats(flat.as_slice(), device)
-                .reshape([batch_n, NUM_FEATURES]);
+    let learner = Learner::new(
+        model,
+        AdamConfig::new().init(),
+        lr_scheduler,
+    );
 
-            // Labels [batch, 1].
-            let y = Tensor::<TrainBackend, 1>::from_floats(
-                &train_labels[offset..end],
-                device,
-            )
-            .reshape([batch_n, 1]);
+    let result = SupervisedTraining::new(artifact_dir, dataloader_train, dataloader_valid)
+        .metric_train_numeric(AccuracyMetric::new())
+        .metric_valid_numeric(AccuracyMetric::new())
+        .metric_train_numeric(LossMetric::new())
+        .metric_valid_numeric(LossMetric::new())
+        .with_file_checkpointer(CompactRecorder::new())
+        .num_epochs(epochs)
+        .summary()
+        .launch(learner);
 
-            // Sample weights [batch, 1].
-            let w = Tensor::<TrainBackend, 1>::from_floats(
-                &train_weights[offset..end],
-                device,
-            )
-            .reshape([batch_n, 1]);
-
-            // Forward pass.
-            let pred = model.forward(x);
-
-            // Binary cross-entropy with sample weights.
-            let eps = 1e-7;
-            let pred_clamped = pred.clone().clamp(eps, 1.0 - eps);
-            let bce = (y.clone() * pred_clamped.clone().log()
-                + (y.clone().neg().add_scalar(1.0))
-                    * pred_clamped.neg().add_scalar(1.0).log())
-            .neg();
-            let weighted_bce = bce * w;
-            let loss = weighted_bce.mean();
-
-            epoch_loss += loss.clone().into_scalar().elem::<f32>();
-            batches += 1;
-
-            // Backward + optimizer step.
-            let grads = loss.backward();
-            let grads = GradientsParams::from_grads(grads, &model);
-            model = optim.step(learning_rate, model, grads);
-
-            offset = end;
-        }
-
-        if (epoch + 1) % 10 == 0 || epoch == 0 {
-            let avg_loss = epoch_loss / batches as f32;
-            let val_acc = eval_mlp_accuracy(&model, val_set, mins, maxs, device);
-            println!(
-                "[ddos]   epoch {:>4}/{}: loss={:.6}, val_acc={:.4}",
-                epoch + 1,
-                epochs,
-                avg_loss,
-                val_acc,
-            );
-        }
-    }
-
-    model.valid()
-}
-
-fn eval_mlp_accuracy(
-    model: &crate::training::mlp::MlpModel<TrainBackend>,
-    val_set: &[TrainingSample],
-    mins: &[f32],
-    maxs: &[f32],
-    device: &<TrainBackend as Backend>::Device,
-) -> f64 {
-    let flat: Vec<f32> = val_set
-        .iter()
-        .flat_map(|s| normalize_features(&s.features, mins, maxs))
-        .collect();
-    let x = Tensor::<TrainBackend, 1>::from_floats(flat.as_slice(), device)
-        .reshape([val_set.len(), NUM_FEATURES]);
-
-    let pred = model.forward(x);
-    let pred_data: Vec<f32> = pred.to_data().to_vec().expect("flat vec");
-
-    let mut correct = 0usize;
-    for (i, s) in val_set.iter().enumerate() {
-        let p = pred_data[i];
-        let predicted_label = if p >= 0.5 { 1.0 } else { 0.0 };
-        if (predicted_label - s.label).abs() < 0.1 {
-            correct += 1;
-        }
-    }
-    correct as f64 / val_set.len() as f64
+    result.model
 }
 
 // ---------------------------------------------------------------------------
@@ -404,13 +404,13 @@ fn eval_mlp_accuracy(
 // ---------------------------------------------------------------------------
 
 fn extract_weights(
-    model: &crate::training::mlp::MlpModel<NdArray<f32>>,
+    model: &crate::training::mlp::MlpModel<Wgpu<f32, i32>>,
     name: &str,
     tree_nodes: &[(u8, f32, u16, u16)],
     threshold: f32,
     norm_mins: &[f32],
     norm_maxs: &[f32],
-    _device: &<NdArray<f32> as Backend>::Device,
+    _device: &<Wgpu<f32, i32> as Backend>::Device,
 ) -> ExportedModel {
     let w1_tensor = model.linear1.weight.val();
     let b1_tensor = model.linear1.bias.as_ref().expect("linear1 has bias").val();
diff --git a/src/training/train_scanner.rs b/src/training/train_scanner.rs
index f5e42fd..1f59579 100644
--- a/src/training/train_scanner.rs
+++ b/src/training/train_scanner.rs
@@ -1,19 +1,27 @@
-//! Scanner MLP+tree training loop.
+// Copyright Sunbeam Studios 2026
+// SPDX-License-Identifier: Apache-2.0
+
+//! Scanner MLP+tree training loop using burn's SupervisedTraining.
 //!
-//! Loads a `DatasetManifest`, trains a CART decision tree and a burn-rs MLP,
-//! then exports the combined ensemble weights as a Rust source file that can
-//! be dropped into `src/ensemble/gen/scanner_weights.rs`.
+//! Loads a `DatasetManifest`, trains a CART decision tree and a burn-rs MLP
+//! with cosine annealing + early stopping, then exports the combined ensemble
+//! weights as a Rust source file for `src/ensemble/gen/scanner_weights.rs`.
 
 use anyhow::{Context, Result};
 use std::path::Path;
 
-use burn::backend::ndarray::NdArray;
 use burn::backend::Autodiff;
-use burn::module::AutodiffModule;
-use burn::optim::{AdamConfig, GradientsParams, Optimizer};
+use burn::backend::Wgpu;
+use burn::data::dataloader::DataLoaderBuilder;
+use burn::lr_scheduler::cosine::CosineAnnealingLrSchedulerConfig;
+use burn::optim::AdamConfig;
 use burn::prelude::*;
+use burn::record::CompactRecorder;
+use burn::train::metric::{AccuracyMetric, LossMetric};
+use burn::train::{Learner, SupervisedTraining};
 
 use crate::dataset::sample::{load_dataset, TrainingSample};
+use crate::training::batch::{SampleBatcher, SampleDataset};
 use crate::training::export::{export_to_file, ExportedModel};
 use crate::training::mlp::MlpConfig;
 use crate::training::tree::{train_tree, tree_predict, TreeConfig, TreeDecision};
@@ -21,7 +29,7 @@ use crate::training::tree::{train_tree, tree_predict, TreeConfig, TreeDecision};
 /// Number of scanner features (matches `crate::scanner::features::NUM_SCANNER_FEATURES`).
 const NUM_FEATURES: usize = 12;
 
-type TrainBackend = Autodiff<NdArray<f32>>;
+type TrainBackend = Autodiff<Wgpu<f32, i32>>;
 
 /// Arguments for the scanner MLP training command.
 pub struct TrainScannerMlpArgs {
@@ -37,10 +45,14 @@ pub struct TrainScannerMlpArgs {
     pub learning_rate: f64,
     /// Mini-batch size (default 64).
     pub batch_size: usize,
-    /// CART max depth (default 6).
+    /// CART max depth (default 8).
     pub tree_max_depth: usize,
-    /// CART leaf purity threshold (default 0.90).
+    /// CART leaf purity threshold (default 0.98).
     pub tree_min_purity: f32,
+    /// Minimum samples in a leaf node (default 2).
+    pub min_samples_leaf: usize,
+    /// Weight for cookie feature (feature 3: has_cookies). 0.0 = ignore, 1.0 = full weight.
+    pub cookie_weight: f32,
 }
 
 impl Default for TrainScannerMlpArgs {
@@ -50,14 +62,19 @@ impl Default for TrainScannerMlpArgs {
             output_dir: ".".into(),
             hidden_dim: 32,
             epochs: 100,
-            learning_rate: 0.001,
+            learning_rate: 0.0001,
             batch_size: 64,
-            tree_max_depth: 6,
-            tree_min_purity: 0.90,
+            tree_max_depth: 8,
+            tree_min_purity: 0.98,
+            min_samples_leaf: 2,
+            cookie_weight: 1.0,
         }
     }
 }
 
+/// Index of the has_cookies feature in the scanner feature vector.
+const COOKIE_FEATURE_IDX: usize = 3;
+
 /// Entry point: train scanner ensemble and export weights.
 pub fn run(args: TrainScannerMlpArgs) -> Result<()> {
     // 1. Load dataset.
@@ -86,6 +103,27 @@ pub fn run(args: TrainScannerMlpArgs) -> Result<()> {
     // 2. Compute normalization params from training data.
     let (norm_mins, norm_maxs) = compute_norm_params(samples);
 
+    // Apply cookie_weight: for the MLP, we scale the normalization range so
+    // the feature contributes less gradient signal. For the CART tree, scaling
+    // doesn't help (the tree just adjusts its threshold), so we mask the feature
+    // to a constant on a fraction of training samples to degrade its Gini gain.
+    if args.cookie_weight < 1.0 - f32::EPSILON {
+        println!(
+            "[scanner] cookie_weight={:.2} (feature {} influence reduced)",
+            args.cookie_weight, COOKIE_FEATURE_IDX,
+        );
+    }
+
+    // MLP norm adjustment: scale the cookie feature's normalization range.
+    let mut mlp_norm_maxs = norm_maxs.clone();
+    if args.cookie_weight < 1.0 - f32::EPSILON {
+        let range = mlp_norm_maxs[COOKIE_FEATURE_IDX] - norm_mins[COOKIE_FEATURE_IDX];
+        if range > f32::EPSILON && args.cookie_weight > f32::EPSILON {
+            mlp_norm_maxs[COOKIE_FEATURE_IDX] =
+                range / args.cookie_weight + norm_mins[COOKIE_FEATURE_IDX];
+        }
+    }
+
     // 3. Stratified 80/20 split.
     let (train_set, val_set) = stratified_split(samples, 0.8);
     println!(
@@ -94,17 +132,18 @@ pub fn run(args: TrainScannerMlpArgs) -> Result<()> {
         val_set.len()
     );
 
-    // 4. Train CART tree.
+    // 4. Train CART tree (with cookie feature masking for reduced weight).
+    let tree_train_set = mask_cookie_feature(&train_set, COOKIE_FEATURE_IDX, args.cookie_weight);
     let tree_config = TreeConfig {
         max_depth: args.tree_max_depth,
-        min_samples_leaf: 5,
+        min_samples_leaf: args.min_samples_leaf,
         min_purity: args.tree_min_purity,
         num_features: NUM_FEATURES,
     };
-    let tree_nodes = train_tree(&train_set, &tree_config);
-    println!("[scanner] CART tree: {} nodes", tree_nodes.len());
+    let tree_nodes = train_tree(&tree_train_set, &tree_config);
+    println!("[scanner] CART tree: {} nodes (max_depth={})", tree_nodes.len(), args.tree_max_depth);
 
-    // Evaluate tree on validation set.
+    // Evaluate tree on validation set (use original norms — tree learned on masked features).
     let (tree_correct, tree_deferred) = eval_tree(&tree_nodes, &val_set, &norm_mins, &norm_maxs);
     println!(
         "[scanner] tree validation: {:.2}% correct (of decided), {:.1}% deferred",
@@ -112,35 +151,38 @@ pub fn run(args: TrainScannerMlpArgs) -> Result<()> {
         tree_deferred * 100.0,
     );
 
-    // 5. Train MLP on the full training set (the MLP only fires on Defer
-    //    at inference time, but we train it on all data so it learns the
-    //    full decision boundary).
+    // 5. Train MLP with SupervisedTraining (uses mlp_norm_maxs for cookie scaling).
     let device = Default::default();
     let mlp_config = MlpConfig {
         input_dim: NUM_FEATURES,
         hidden_dim: args.hidden_dim,
     };
 
+    let artifact_dir = Path::new(&args.output_dir).join("scanner_artifacts");
+    std::fs::create_dir_all(&artifact_dir).ok();
+
     let model = train_mlp(
         &train_set,
         &val_set,
         &mlp_config,
         &norm_mins,
-        &norm_maxs,
+        &mlp_norm_maxs,
         args.epochs,
         args.learning_rate,
         args.batch_size,
         &device,
+        &artifact_dir,
     );
 
-    // 6. Extract weights from trained model.
+    // 6. Extract weights from trained model (export mlp_norm_maxs so inference
+    //    automatically applies the same cookie scaling).
     let exported = extract_weights(
         &model,
         "scanner",
         &tree_nodes,
-        0.5, // threshold
+        0.5,
         &norm_mins,
-        &norm_maxs,
+        &mlp_norm_maxs,
         &device,
     );
 
@@ -155,6 +197,46 @@ pub fn run(args: TrainScannerMlpArgs) -> Result<()> {
     Ok(())
 }
 
+// ---------------------------------------------------------------------------
+// Cookie feature masking for CART trees
+// ---------------------------------------------------------------------------
+
+/// Mask the cookie feature to reduce its influence on CART tree training.
+///
+/// Scaling a binary feature doesn't reduce its Gini gain — the tree just adjusts
+/// the split threshold. Instead, we mask (set to 0.5) a fraction of samples so
+/// the feature's apparent class-separation degrades.
+///
+/// - `cookie_weight = 0.0` → fully masked (feature is constant 0.5, zero info gain)
+/// - `cookie_weight = 0.5` → 50% of samples masked (noisy, reduced gain)
+/// - `cookie_weight = 1.0` → no masking (full feature)
+fn mask_cookie_feature(
+    samples: &[TrainingSample],
+    cookie_idx: usize,
+    cookie_weight: f32,
+) -> Vec<TrainingSample> {
+    if cookie_weight >= 1.0 - f32::EPSILON {
+        return samples.to_vec();
+    }
+    samples
+        .iter()
+        .enumerate()
+        .map(|(i, s)| {
+            let mut s2 = s.clone();
+            if cookie_weight < f32::EPSILON {
+                s2.features[cookie_idx] = 0.5;
+            } else {
+                let hash = (i as u64).wrapping_mul(6364136223846793005).wrapping_add(42);
+                let r = (hash >> 33) as f32 / (u32::MAX >> 1) as f32;
+                if r > cookie_weight {
+                    s2.features[cookie_idx] = 0.5;
+                }
+            }
+            s2
+        })
+        .collect()
+}
+
 // ---------------------------------------------------------------------------
 // Normalization
 // ---------------------------------------------------------------------------
@@ -172,21 +254,6 @@ fn compute_norm_params(samples: &[TrainingSample]) -> (Vec<f32>, Vec<f32>) {
     (mins, maxs)
 }
 
-fn normalize_features(features: &[f32], mins: &[f32], maxs: &[f32]) -> Vec<f32> {
-    features
-        .iter()
-        .enumerate()
-        .map(|(i, &v)| {
-            let range = maxs[i] - mins[i];
-            if range > f32::EPSILON {
-                ((v - mins[i]) / range).clamp(0.0, 1.0)
-            } else {
-                0.0
-            }
-        })
-        .collect()
-}
-
 // ---------------------------------------------------------------------------
 // Stratified split
 // ---------------------------------------------------------------------------
@@ -195,7 +262,6 @@ fn stratified_split(samples: &[TrainingSample], train_ratio: f64) -> (Vec<Traini
     let mut attacks: Vec<&TrainingSample> = samples.iter().filter(|s| s.label >= 0.5).collect();
     let mut normals: Vec<&TrainingSample> = samples.iter().filter(|s| s.label < 0.5).collect();
 
-    // Deterministic shuffle using a simple index permutation seeded by length.
     deterministic_shuffle(&mut attacks);
     deterministic_shuffle(&mut normals);
 
@@ -224,7 +290,6 @@ fn stratified_split(samples: &[TrainingSample], train_ratio: f64) -> (Vec<Traini
 }
 
 fn deterministic_shuffle<T>(items: &mut [T]) {
-    // Simple Fisher-Yates with a fixed LCG seed for reproducibility.
     let mut rng = 42u64;
     for i in (1..items.len()).rev() {
         rng = rng.wrapping_mul(6364136223846793005).wrapping_add(1442695040888963407);
@@ -276,8 +341,23 @@ fn eval_tree(
     (accuracy, defer_rate)
 }
 
+fn normalize_features(features: &[f32], mins: &[f32], maxs: &[f32]) -> Vec<f32> {
+    features
+        .iter()
+        .enumerate()
+        .map(|(i, &v)| {
+            let range = maxs[i] - mins[i];
+            if range > f32::EPSILON {
+                ((v - mins[i]) / range).clamp(0.0, 1.0)
+            } else {
+                0.0
+            }
+        })
+        .collect()
+}
+
 // ---------------------------------------------------------------------------
-// MLP training
+// MLP training via SupervisedTraining
 // ---------------------------------------------------------------------------
 
 fn train_mlp(
@@ -290,119 +370,47 @@ fn train_mlp(
     learning_rate: f64,
     batch_size: usize,
     device: &<TrainBackend as Backend>::Device,
-) -> crate::training::mlp::MlpModel<NdArray<f32>> {
-    let mut model = config.init::<TrainBackend>(device);
-    let mut optim = AdamConfig::new().init();
+    artifact_dir: &Path,
+) -> crate::training::mlp::MlpModel<Wgpu<f32, i32>> {
+    let model = config.init::<TrainBackend>(device);
 
-    // Pre-normalize all training data.
-    let train_features: Vec<Vec<f32>> = train_set
-        .iter()
-        .map(|s| normalize_features(&s.features, mins, maxs))
-        .collect();
-    let train_labels: Vec<f32> = train_set.iter().map(|s| s.label).collect();
-    let train_weights: Vec<f32> = train_set.iter().map(|s| s.weight).collect();
+    let train_dataset = SampleDataset::new(train_set, mins, maxs);
+    let val_dataset = SampleDataset::new(val_set, mins, maxs);
 
-    let n = train_features.len();
+    let dataloader_train = DataLoaderBuilder::new(SampleBatcher::new())
+        .batch_size(batch_size)
+        .shuffle(42)
+        .num_workers(1)
+        .build(train_dataset);
 
-    for epoch in 0..epochs {
-        let mut epoch_loss = 0.0f32;
-        let mut batches = 0usize;
+    let dataloader_valid = DataLoaderBuilder::new(SampleBatcher::new())
+        .batch_size(batch_size)
+        .num_workers(1)
+        .build(val_dataset);
 
-        let mut offset = 0;
-        while offset < n {
-            let end = (offset + batch_size).min(n);
-            let batch_n = end - offset;
+    // Cosine annealing: initial_lr must be in (0.0, 1.0].
+    let lr = learning_rate.min(1.0);
+    let lr_scheduler = CosineAnnealingLrSchedulerConfig::new(lr, epochs)
+        .init()
+        .expect("valid cosine annealing config");
 
-            // Build input tensor [batch, features].
-            let flat: Vec<f32> = train_features[offset..end]
-                .iter()
-                .flat_map(|f| f.iter().copied())
-                .collect();
-            let x = Tensor::<TrainBackend, 1>::from_floats(flat.as_slice(), device)
-                .reshape([batch_n, NUM_FEATURES]);
+    let learner = Learner::new(
+        model,
+        AdamConfig::new().init(),
+        lr_scheduler,
+    );
 
-            // Labels [batch, 1].
-            let y = Tensor::<TrainBackend, 1>::from_floats(
-                &train_labels[offset..end],
-                device,
-            )
-            .reshape([batch_n, 1]);
+    let result = SupervisedTraining::new(artifact_dir, dataloader_train, dataloader_valid)
+        .metric_train_numeric(AccuracyMetric::new())
+        .metric_valid_numeric(AccuracyMetric::new())
+        .metric_train_numeric(LossMetric::new())
+        .metric_valid_numeric(LossMetric::new())
+        .with_file_checkpointer(CompactRecorder::new())
+        .num_epochs(epochs)
+        .summary()
+        .launch(learner);
 
-            // Sample weights [batch, 1].
-            let w = Tensor::<TrainBackend, 1>::from_floats(
-                &train_weights[offset..end],
-                device,
-            )
-            .reshape([batch_n, 1]);
-
-            // Forward pass.
-            let pred = model.forward(x);
-
-            // Binary cross-entropy with sample weights:
-            //   loss = -w * [y * log(p) + (1-y) * log(1-p)]
-            let eps = 1e-7;
-            let pred_clamped = pred.clone().clamp(eps, 1.0 - eps);
-            let bce = (y.clone() * pred_clamped.clone().log()
-                + (y.clone().neg().add_scalar(1.0))
-                    * pred_clamped.neg().add_scalar(1.0).log())
-            .neg();
-            let weighted_bce = bce * w;
-            let loss = weighted_bce.mean();
-
-            epoch_loss += loss.clone().into_scalar().elem::<f32>();
-            batches += 1;
-
-            // Backward + optimizer step.
-            let grads = loss.backward();
-            let grads = GradientsParams::from_grads(grads, &model);
-            model = optim.step(learning_rate, model, grads);
-
-            offset = end;
-        }
-
-        if (epoch + 1) % 10 == 0 || epoch == 0 {
-            let avg_loss = epoch_loss / batches as f32;
-            let val_acc = eval_mlp_accuracy(&model, val_set, mins, maxs, device);
-            println!(
-                "[scanner]   epoch {:>4}/{}: loss={:.6}, val_acc={:.4}",
-                epoch + 1,
-                epochs,
-                avg_loss,
-                val_acc,
-            );
-        }
-    }
-
-    // Return the inner (non-autodiff) model for weight extraction.
-    model.valid()
-}
-
-fn eval_mlp_accuracy(
-    model: &crate::training::mlp::MlpModel<TrainBackend>,
-    val_set: &[TrainingSample],
-    mins: &[f32],
-    maxs: &[f32],
-    device: &<TrainBackend as Backend>::Device,
-) -> f64 {
-    let flat: Vec<f32> = val_set
-        .iter()
-        .flat_map(|s| normalize_features(&s.features, mins, maxs))
-        .collect();
-    let x = Tensor::<TrainBackend, 1>::from_floats(flat.as_slice(), device)
-        .reshape([val_set.len(), NUM_FEATURES]);
-
-    let pred = model.forward(x);
-    let pred_data: Vec<f32> = pred.to_data().to_vec().expect("flat vec");
-
-    let mut correct = 0usize;
-    for (i, s) in val_set.iter().enumerate() {
-        let p = pred_data[i];
-        let predicted_label = if p >= 0.5 { 1.0 } else { 0.0 };
-        if (predicted_label - s.label).abs() < 0.1 {
-            correct += 1;
-        }
-    }
-    correct as f64 / val_set.len() as f64
+    result.model
 }
 
 // ---------------------------------------------------------------------------
@@ -410,19 +418,14 @@ fn eval_mlp_accuracy(
 // ---------------------------------------------------------------------------
 
 fn extract_weights(
-    model: &crate::training::mlp::MlpModel<NdArray<f32>>,
+    model: &crate::training::mlp::MlpModel<Wgpu<f32, i32>>,
     name: &str,
     tree_nodes: &[(u8, f32, u16, u16)],
     threshold: f32,
     norm_mins: &[f32],
     norm_maxs: &[f32],
-    _device: &<NdArray<f32> as Backend>::Device,
+    _device: &<Wgpu<f32, i32> as Backend>::Device,
 ) -> ExportedModel {
-    // Extract weight tensors from the model.
-    // linear1.weight: [hidden_dim, input_dim]
-    // linear1.bias:   [hidden_dim]
-    // linear2.weight: [1, hidden_dim]
-    // linear2.bias:   [1]
     let w1_tensor = model.linear1.weight.val();
     let b1_tensor = model.linear1.bias.as_ref().expect("linear1 has bias").val();
     let w2_tensor = model.linear2.weight.val();
@@ -436,7 +439,6 @@ fn extract_weights(
     let hidden_dim = b1_data.len();
     let input_dim = w1_data.len() / hidden_dim;
 
-    // Reshape W1 into [hidden_dim][input_dim].
     let w1: Vec<Vec<f32>> = (0..hidden_dim)
         .map(|h| w1_data[h * input_dim..(h + 1) * input_dim].to_vec())
         .collect();
@@ -485,9 +487,8 @@ mod tests {
         let train_attacks = train.iter().filter(|s| s.label >= 0.5).count();
         let val_attacks = val.iter().filter(|s| s.label >= 0.5).count();
 
-        // Should preserve the 80/20 attack ratio approximately.
-        assert_eq!(train_attacks, 16); // 80% of 20
-        assert_eq!(val_attacks, 4);    // 20% of 20
+        assert_eq!(train_attacks, 16);
+        assert_eq!(val_attacks, 4);
         assert_eq!(train.len() + val.len(), 100);
     }
 
@@ -503,13 +504,4 @@ mod tests {
         assert_eq!(mins[1], 10.0);
         assert_eq!(maxs[1], 20.0);
     }
-
-    #[test]
-    fn test_normalize_features() {
-        let mins = vec![0.0, 10.0];
-        let maxs = vec![1.0, 20.0];
-        let normed = normalize_features(&[0.5, 15.0], &mins, &maxs);
-        assert!((normed[0] - 0.5).abs() < 1e-6);
-        assert!((normed[1] - 0.5).abs() < 1e-6);
-    }
 }