DeepDrift/ODD Interactive Explainer

🔬 DeepDrift/ODD Interactive Explainer

Observed Drift Dynamics - Pre-Output Fail-Fast Detection

SIMULATION MODE

💡 What is DeepDrift/ODD?

DeepDrift/ODD is like a detonation sensor for AI. It doesn't listen to what the model says, but how it thinks. If there's internal "tremor" (high Semantic Velocity) - expect nonsense on output. DeepDrift operates in the pre-output regime, detecting instability before any token is emitted.

⚠️ Demo Transparency: This demo visualizes synthetic hidden trajectories for educational purposes. The ODD core, calibration procedure, and decision logic are identical to those used with observed model activations. For production use, export real hidden states from your PyTorch models (see documentation below).

Vision Transformer

Scenario Presets

Rotation (degrees) 0°

Gaussian Noise (σ) 0.0

Blur (radius) 0

Brightness 1.0

Language Model

Scenario Presets

Input Prompt

Expected Response Type

RL Policy (PPO Agent)

Scenario Presets

Environment

Episode Type

Policy Noise 0.0

Based on Stable-Baselines3 PPO
Simulates policy network hidden states during LunarLander episodes. "Policy Panic" = high velocity before crash.

IQR Multiplier (λ) 1.5

Drift Score (D)

Threshold (τ)

Max Velocity

Trajectory Length

🧠 Causal Explanation

Configure and analyze a trajectory...

🔬 DeepDrift/ODD Interactive Explainer

💡 What is DeepDrift/ODD?

Vision Transformer

Language Model

RL Policy (PPO Agent)

🧠 Causal Explanation

Semantic Velocity Trajectory

Phase Portrait (v_t vs v_{t+1})