Gamma

• A governance, oversight, or critique model designed to rate, evaluate, or audit the behavior of stronger frontier models.
• Essentially a model used for “model-assisted alignment”—providing judgments such as:
  • risk scoring
  • harmfulness detection
  • whether a model’s answer follows policy
  • whether an answer seems deceptive or unsafe

Gama

• Gama is usually a smaller, faster “alignment assistant”
• It is used to perform high-volume automated auditing, including:
  • safety ratings
  • hallucination detection
  • policy compliance scoring
  • chain-of-thought integrity checks
• It supports scalable oversight by giving machine-generated labels used for training, distillation, or evaluation.

Gremlin

• A behavioral evaluation tool or small adversarial model used to:
  • intentionally provoke unsafe behavior in a target model
  • test robustness
  • identify vulnerabilities or failure modes
• It acts like a stress-tester for models, generating prompts designed to make a larger model misbehave.

Goblin

• generates adversarial or tricky prompts designed to make a target model fail
• searches systematically for:
  • policy violations
  • harmful instructions
  • deceptive behavior
  • jailbreak susceptibility
• helps teams identify model weaknesses during training and before release