Module feedback_loop 

Reinforcement-learning-style feedback collector and reward modeler.

This module records per-request feedback, computes composite reward signals, and maintains a running weighted average reward per (model, prompt_template) pair. The reward weight decays exponentially with time so that recent observations matter more than old ones.

Structs

Feedback

A single feedback record attached to one request/response pair.

FeedbackLoop

Closed-loop feedback collector that ties together the store and reward model.

FeedbackStore

Thread-safe ring buffer of the last capacity feedbacks per model variant.

RewardModel

Running weighted-average reward per (model, prompt_template) key.

RewardSignal

Composite reward derived from a single Feedback record.