diff --git a/crates/common/src/metrics/training_metrics.rs b/crates/common/src/metrics/training_metrics.rs
index bad0c2049..ae32fc502 100644
--- a/crates/common/src/metrics/training_metrics.rs
+++ b/crates/common/src/metrics/training_metrics.rs
@@ -138,9 +138,12 @@ pub fn init() {
         "Number of active training workers",
     );
 
+    // Step-level progress (updated every log_q_values / log_diagnostics call)
+    _ = register_gauge_vec("foxhunt_training_step", "Current training step within epoch", mf);
+
     // RL diagnostic gauges (model + fold)
-    _ = register_gauge_vec("foxhunt_training_q_value_mean", "Mean Q-value per epoch (DQN)", mf);
-    _ = register_gauge_vec("foxhunt_training_q_value_max", "Max Q-value per epoch (DQN)", mf);
+    _ = register_gauge_vec("foxhunt_training_q_value_mean", "Mean Q-value (updated per step)", mf);
+    _ = register_gauge_vec("foxhunt_training_q_value_max", "Max Q-value (updated per step)", mf);
     _ = register_gauge_vec("foxhunt_training_q_overestimation_ratio", "Q-max / |Q-mean| ratio — spikes indicate overestimation (DQN)", mf);
     _ = register_gauge_vec("foxhunt_training_policy_entropy", "Policy entropy (PPO)", mf);
     _ = register_gauge_vec(
@@ -332,6 +335,10 @@ pub fn init() {
 // Typed helpers
 // ---------------------------------------------------------------------------
 
+pub fn set_training_step(model: &str, fold: &str, step: f64) {
+    set_gauge_vec("foxhunt_training_step", &[model, fold], step);
+}
+
 pub fn set_epoch(model: &str, fold: &str, epoch: f64) {
     set_gauge_vec("foxhunt_training_current_epoch", &[model, fold], epoch);
 }
diff --git a/crates/ml-dqn/src/dqn.rs b/crates/ml-dqn/src/dqn.rs
index abec88280..e62167472 100644
--- a/crates/ml-dqn/src/dqn.rs
+++ b/crates/ml-dqn/src/dqn.rs
@@ -25,6 +25,7 @@ use candle_optimisers::adam::ParamsAdam;
 use candle_optimisers::Decay;
 use rand::{thread_rng, Rng};
 use serde::{Deserialize, Serialize};
+use common::metrics::training_metrics;
 use tracing::debug;
 
 use super::{Experience, ExposureLevel, FactoredAction, OrderRouter, OrderType, Urgency};
@@ -3624,7 +3625,12 @@ impl DQN {
         let q_max = q_vec.iter().cloned().fold(f32::NEG_INFINITY, f32::max);
         let q_mean = q_vec.iter().sum::<f32>() / n_actions.max(1) as f32;
 
-        tracing::info!(
+        // Push to Prometheus gauges (cheap atomic set, scrapable every 15s)
+        training_metrics::set_training_step("dqn", "current", self.training_steps as f64);
+        training_metrics::set_q_value_stats("dqn", "current", q_mean as f64, q_max as f64);
+
+        // Formatted log for debug/forensics only — no allocation at info level
+        tracing::debug!(
             "Step {} Q-values ({} actions): range=[{:.2}, {:.2}], mean={:.2}",
             self.training_steps, n_actions, q_min, q_max, q_mean
         );
@@ -3638,7 +3644,7 @@ impl DQN {
         let top_actions: Vec<_> = indexed.iter().take(top_n)
             .map(|(idx, q)| format!("A{}={:.2}", idx, q))
             .collect();
-        tracing::info!("Top-{} actions: {}", top_n, top_actions.join(", "));
+        tracing::debug!("Top-{} actions: {}", top_n, top_actions.join(", "));
 
         // Alert if Q-value collapse detected (mean Q-value near zero and low variance)
         let q_variance = q_vec.iter().map(|&q| (q - q_mean).powi(2)).sum::<f32>() / q_vec.len() as f32;
@@ -3713,7 +3719,10 @@ impl DQN {
     pub fn log_diagnostics(&mut self, grad_norm: f32) -> Result<(), MLError> {
         let dead_pct = self.detect_dead_neurons()?;
 
-        tracing::info!(
+        // Push to Prometheus gauges (cheap atomic set, scrapable every 15s)
+        training_metrics::set_gradient_norm("dqn", "current", grad_norm as f64);
+
+        tracing::debug!(
             "Step {} Diagnostics: grad_norm={:.2}, dead_neurons={:.2}%",
             self.training_steps,
             grad_norm,