diff --git a/engine/lib/__init__.py b/engine/lib/__init__.py
index 4e8fd99..d120204 100644
--- a/engine/lib/__init__.py
+++ b/engine/lib/__init__.py
@@ -1,3 +1,6 @@
 from .demand import estimate_demand, generate_demand_for_actor
 from .behavior import sample_behavior
 from .render import DashboardRenderer, style_axis
+from .wrappers import EconomicMetricsWrapper
+from .callbacks import MetricsCallback, EvalMetricsCallback
+from .providers import ProviderBenchmark, ProviderResult, BenchmarkConfig
diff --git a/engine/lib/behavior.py b/engine/lib/behavior.py
index f7dcc65..0f8c486 100644
--- a/engine/lib/behavior.py
+++ b/engine/lib/behavior.py
@@ -1,27 +1,39 @@
-from sim.rl.behavior_loader.models import BehaviorModel, AgentBehaviorModel, aggregate_event_transitions
+from sim.rl.behavior_loader.models import (
+    BehaviorModel,
+    AgentBehaviorModel,
+    aggregate_event_transitions,
+)
 import pandas as pd
 import numpy as np
 from .demand import generate_demand_for_actor
 
 base_dir = "/home/velocitatem/Documents/Projects/PHANTOM/experiments"
-human_dir, agent_dir = f"{base_dir}/collected_data/", f"{base_dir}/agents/collected_data/"
+human_dir, agent_dir = (
+    f"{base_dir}/collected_data/",
+    f"{base_dir}/agents/collected_data/",
+)
 
 _cache = {}  # lazy cache for models and base pivots
 
+
 def _get_base_pivot(human: bool):
-    key = 'human' if human else 'agent'
+    key = "human" if human else "agent"
     if key not in _cache:
         model = BehaviorModel(human_dir) if human else AgentBehaviorModel(agent_dir)
         mdp = model.build_MDP()
         _cache[key] = pd.DataFrame(aggregate_event_transitions(mdp)).fillna(0.0)
     return _cache[key]
 
+
 def adjust_behavior_to_condition(condition, transition_matrix):
     # expand NxN transition matrix to (N*P)x(N*P) weighted by demand condition
     cond_norm = condition / np.sum(condition)
     n_products = len(condition)
     base_vals = transition_matrix.values
-    base_cols, base_rows = transition_matrix.columns.tolist(), transition_matrix.index.tolist()
+    base_cols, base_rows = (
+        transition_matrix.columns.tolist(),
+        transition_matrix.index.tolist(),
+    )
 
     # expand via kronecker-like tiling: each cell becomes a P*P block weighted by outer product of cond_norm
     expanded = np.kron(base_vals, np.outer(cond_norm, cond_norm))
@@ -29,19 +41,24 @@ def adjust_behavior_to_condition(condition, transition_matrix):
     new_rows = [f"{r}_product{p}" for r in base_rows for p in range(n_products)]
     return pd.DataFrame(expanded, index=new_rows, columns=new_cols)
 
+
 def sample_behavior(condition, human=True, max_len=40):
     base_pivot = _get_base_pivot(human)
     adjusted_transitions = adjust_behavior_to_condition(condition, base_pivot)
 
     trajectory = [np.random.choice(adjusted_transitions.index)]
-    while len(trajectory) < max_len or 'checkout' in trajectory[-1]:
+    while len(trajectory) < max_len and "checkout" not in trajectory[-1]:
         probs = adjusted_transitions.loc[trajectory[-1]].values
-        sample = np.random.choice(adjusted_transitions.columns, p=probs/np.sum(probs) if np.sum(probs) > 0 else None)
+        sample = np.random.choice(
+            adjusted_transitions.columns,
+            p=probs / np.sum(probs) if np.sum(probs) > 0 else None,
+        )
         trajectory.append(sample)
     return trajectory
 
+
 if __name__ == "__main__":
-    t=sample_behavior(generate_demand_for_actor(np.array([10,20,30])), human=True)
+    t = sample_behavior(generate_demand_for_actor(np.array([10, 20, 30])), human=True)
     print(t)
-    t=sample_behavior(generate_demand_for_actor(np.array([10,20,30])), human=False)
+    t = sample_behavior(generate_demand_for_actor(np.array([10, 20, 30])), human=False)
     print(t)
diff --git a/engine/train.py b/engine/train.py
index 496ecfd..ebb14f4 100644
--- a/engine/train.py
+++ b/engine/train.py
@@ -1,21 +1,16 @@
+import wandb
 from stable_baselines3 import SAC
-from stable_baselines3.common.callbacks import EvalCallback, BaseCallback
+from stable_baselines3.common.callbacks import EvalCallback
 from .wrapper import PHANTOM
+from .lib import EconomicMetricsWrapper, MetricsCallback
 
+wandb.init(
+    project="phantom-pricing",
+    config={"alpha": 0.3, "n_products": 10, "total_timesteps": 50000}
+)
 
-class RenderCallback(BaseCallback):
-    """Renders environment on every step for live visualization."""
-    def __init__(self, env: PHANTOM):
-        super().__init__()
-        self.env = env
-
-    def _on_step(self) -> bool:
-        self.env.render()
-        return True
-
-
-env = PHANTOM(n_products=10, alpha=0.3, render_mode="human")
-eval_env = PHANTOM(n_products=10, alpha=0.3, render_mode=None)
+env = EconomicMetricsWrapper(PHANTOM(n_products=10, alpha=0.3, render_mode=None))
+eval_env = EconomicMetricsWrapper(PHANTOM(n_products=10, alpha=0.3, render_mode=None))
 
 model = SAC(
     "MultiInputPolicy",
@@ -28,11 +23,12 @@ model = SAC(
     gamma=0.99,
 )
 
-render_cb = RenderCallback(env)
+metrics_cb = MetricsCallback(log_histograms=True, log_freq=100)
 eval_cb = EvalCallback(eval_env, eval_freq=1000, n_eval_episodes=5, verbose=1)
 
-model.learn(total_timesteps=50000, callback=[render_cb, eval_cb])
+model.learn(total_timesteps=50000, callback=[metrics_cb, eval_cb])
 model.save("phantom_sac")
+wandb.finish()
 
 # test trained policy
 env = PHANTOM(n_products=10, alpha=0.3, render_mode="human")
diff --git a/engine/wrapper.py b/engine/wrapper.py
index 9bf3048..1dee8f5 100644
--- a/engine/wrapper.py
+++ b/engine/wrapper.py
@@ -4,6 +4,7 @@ import numpy as np
 from .engine import Limbo, MarketEngine, PricingEngine
 from .lib.render import DashboardRenderer
 from .lib.coi import compute_coi_proxy
+from .lib.wrappers import EconomicMetricsWrapper
 
 
 class PHANTOM(gym.Env):
@@ -134,11 +135,43 @@ class PHANTOM(gym.Env):
 
 
 if __name__ == "__main__":
-    env = PHANTOM(n_products=15, alpha=0.3, N=100, render_mode="human")
-    obs, _ = env.reset()
-    for step in range(100):
-        action = env.action_space.sample()
-        obs, reward, term, trunc, info = env.step(action)
-        env.render()
-        if term: break
+    import wandb
+    from .lib import MetricsCallback
+
+    class RandomPolicy:
+        """Minimal SB3-compatible random policy for baseline testing."""
+        def __init__(self, env):
+            self.env = env
+            self.num_timesteps = 0
+
+        def learn(self, total_timesteps, callback=None):
+            callback.model = self
+            callback.num_timesteps = 0
+            callback.locals = {}
+            callback.on_training_start({}, {})
+
+            obs, _ = self.env.reset()
+            for step in range(total_timesteps):
+                action = self.env.action_space.sample()
+                obs, reward, term, trunc, info = self.env.step(action)
+                self.num_timesteps = step + 1
+                callback.num_timesteps = self.num_timesteps
+                callback.locals = {"infos": [info]}
+                callback.on_step()
+                if term or trunc:
+                    callback.on_rollout_end()
+                    obs, _ = self.env.reset()
+            return self
+
+        def predict(self, obs, **kwargs):
+            return self.env.action_space.sample(), None
+
+    wandb.init(project="phantom-pricing", config={"policy": "random", "alpha": 0.3})
+    env = EconomicMetricsWrapper(PHANTOM(n_products=15, alpha=0.3, render_mode=None))
+
+    model = RandomPolicy(env)
+    model.learn(total_timesteps=1000, callback=MetricsCallback())
+
+    print(f"Episode revenue: {env.episode_revenue:.1f}")
+    wandb.finish()
     env.close()
diff --git a/requirements.txt b/requirements.txt
index e7eaf09..247121e 100644
--- a/requirements.txt
+++ b/requirements.txt
@@ -12,3 +12,4 @@ uv
 scikit-learn
 supabase
 pymc
+wandb