The specific numerical value assigned to an AI agent's action during reinforcement learning to indicate whether the action moved the system closer to its designated goal.

Used by MLOps managers to gamify basic software testing, treating standard pass-fail metrics like a high-score arcade leaderboard.

• Arcade Mechanics: Making standard software evaluation sound like a video game.
• Metric Over-Simplification: Reducing complex qualitative software accuracy down to a single arbitrary number to impress executives.
• Kitchen Reality Check: Like eating a sandwich, holding up a scorecard with an '8' on it, and calling it a "lunch reward score."

Love Island (Season 7): Liberty Poole tracks a declining reward score with her partner Jake, constantly evaluating his subtle actions and realizing his public declarations are entirely performative rather than genuine.

A numerical grade given to a software model during training to tell it if it did a good job.

“Our training loop penalizes hallucinations by dropping the output reward score.”

“We set up a system that gives the chatbot a high score when it cites the correct document and a zero when it makes up a fact.”