Jev Reinforcement Learning

A PPO-trained reinforcement-learning decision brain for the CONNECT TYCOON company simulator from CONNECT AI LAB: it runs staffing, inventory, marketing, refund and security judgments and answers in the JEV request format with probabilities.

Sammyliu's audio decision corpus for reinforcement learning and evaluation: causal audio windows paired with runtime action questions, each receiving a categorical label distribution. Train holds 168,494 decisions over 71,952 source windows; test, dev, calibration and an external diagnostic add about 122,600 more.

A Korean-language reinforcement-learning game that trains a company decision model: it reads a situation, a question and options, then answers with a choice plus probability and confidence, and asks a human when unsure. Built for a CONNECT AI LAB class; Apache-2.0, Python and Pygame.

A harness for evaluating TypeSafe Jev as a policy in Gymnasium and Atari environments: adapters turn observations into structured JSON and Jev picks a legal action through a single Choice question, with no training.

A runnable reward-model laboratory that trains DQN agents on CartPole, MountainCar, Acrobot and FrozenLake with official JEV scores as rewards: 36 runs across 3 seeds and 3.24 million steps, with checkpoint replays and a recorded JEV cost of about $0.00241.

A study of whether Jev can serve inside reinforcement learning training rather than as a single-domain black box. Frozen Jev fills reference-policy, exploration-judge, and replay-rater roles, and training with it beats a standard RL learner across nine MiniGrid tasks and three Atari games.

Dyad augments a pretrained LLM with an environment-conditioned action encoder that scores candidate action descriptions to yield a distribution over typed actions. Training only the encoder helps across four unseen environments; joint training adds a 3.80% absolute gain on ALFWorld with a 9B model.