rewardguard-premium 2.0.6

Auto-tuning reward system for production RL applications (Premium)

RewardGuard Premium

Automatic Reward Alignment for Production RL — with Statistical Detection and Auto-Correction

RewardGuard Premium is the paid tier of RewardGuard. Install it freely from PyPI — just sign in to your RewardGuard account to activate.

---

Installation

pip install rewardguard-premium

On first use you will be prompted to sign in:

RewardGuard Premium — Sign in to your account
  Visit https://rewardguard.ai to create an account

  Email: [e-mail]
  Password: ••••••••
  Signed in successfully!

Your session is saved to ~/.rewardguard/session.json and refreshed automatically. You only sign in once per machine.

Requires an active RewardGuard Premium subscription. Subscribe at: https://rewardguard.ai/premium

---

Sign-in CLI

rewardguard-premium login    # sign in (or switch accounts)
rewardguard-premium logout   # clear saved session
rewardguard-premium status   # show who you are signed in as

CI / automated environments — use env vars instead of the interactive prompt:

export REWARDGUARD_EMAIL=[e-mail]
export REWARDGUARD_PASSWORD=yourpassword

---

Quick Start

from rewardguard_premium import AutoMonitor

monitor = AutoMonitor(
    expected={"task": 0.7, "safety": 0.3},
    baseline_steps=300,   # warm-up before detection activates
    auto_correct=True,    # adjust weights automatically when flagged
)

for episode in range(num_episodes):
    for step in range(max_steps):
        r_task, r_safety = env.step(action)

        snapshot = monitor.step({"task": r_task, "safety": r_safety})

        if snapshot:
            if snapshot.flag == "critical":
                # Apply auto-corrected weights back to the environment
                env.set_reward_weights(monitor.weights)

monitor.print_report()

---

What AutoMonitor Does

Phase	What happens
Warm-up (baseline_steps)	Learns the normal ratio for your environment. Returns None from step().
Detection	Computes per-component z-scores against the learned baseline.
Alignment score	0 (misaligned) → 1 (fully aligned), sigmoid-mapped from the max z-score.
Flagging	ok (score > 0.75) / warning (> 0.5) / critical (≤ 0.5)
Auto-correction	Adjusts per-component weight multipliers when flagged.
Drift velocity	Linear regression slope over recent scores — distinguishes trends from spikes.

---

AlignmentSnapshot

Every call to step() after warm-up returns an AlignmentSnapshot:

snapshot.alignment_score      # float 0–1
snapshot.flag                 # "ok" / "warning" / "critical"
snapshot.z_scores             # {"task": -0.4, "safety": +2.8}
snapshot.drift_velocity       # negative = worsening trend
snapshot.corrections_applied  # {"safety": 1.24}  — weights changed this step
snapshot.component_ratios     # {"task": 68.3, "safety": 31.7}

---

Framework Integrations

Weights & Biases

from rewardguard_premium import AutoMonitor, make_wandb_callback
import wandb

wandb.init(project="my-rl-run")
monitor = AutoMonitor(
    expected={"task": 0.7, "safety": 0.3},
    callbacks=[make_wandb_callback()],
)

TensorBoard

from torch.utils.tensorboard import SummaryWriter
from rewardguard_premium import AutoMonitor, make_tensorboard_callback

writer = SummaryWriter("runs/my_run")
monitor = AutoMonitor(
    expected={"task": 0.7, "safety": 0.3},
    callbacks=[make_tensorboard_callback(writer)],
)

Stable-Baselines3

from stable_baselines3 import PPO
from rewardguard_premium import AutoMonitor, make_sb3_callback

monitor = AutoMonitor(expected={"task": 0.7, "safety": 0.3})
model = PPO("MlpPolicy", env)
model.learn(total_timesteps=500_000, callback=make_sb3_callback(monitor))

Your environment must include "reward_components" in its info dict for the SB3 callback.

---

Save / Load State

# Save after training
monitor.save("run_42_state.json")

# Resume later
monitor = AutoMonitor.load("run_42_state.json")

---

Export Data

monitor.to_json("results.json")   # full state
monitor.to_csv("snapshots.csv")   # one row per detection-phase step

---

AutoMonitor Parameters

Parameter	Default	Description
expected	required	Target distribution, e.g. {"task": 0.7, "safety": 0.3}
baseline_steps	300	Warm-up steps before detection activates
z_threshold	2.5	Z-score at which a component is flagged
auto_correct	True	Automatically adjust weights when flagged
correction_rate	0.2	Fraction of required correction applied per step
tolerance	5.0	Percentage-point tolerance for the free-tier check() method
window	200	Rolling window size for ratio computation
drift_window	30	Snapshots used to compute drift velocity
callbacks	[]	List of callables invoked with each AlignmentSnapshot

---

Free vs Premium

Feature	Free (rewardguard)	Premium
Live in-loop monitoring	✅	✅
Log-file analysis	✅	✅
Imbalance detection	✅	✅
Weight recommendations	✅	✅
Baseline learning	❌	✅
Z-score detection	❌	✅
Alignment score (0–1)	❌	✅
Drift velocity	❌	✅
Auto-correction	❌	✅
WandB / TensorBoard / SB3	❌	✅
Save / load state	❌	✅

---

License

Proprietary — requires an active RewardGuard Premium subscription. © 2026 RewardGuard | https://rewardguard.ai