Reinforcement learning environment
The controlled setting where a model tries a task, its attempt is checked, and the check returns a reward.
Category
How a model is chosen and adapted
Related terms
A reinforcement learning (RL) environment is the controlled setting in which a model attempts a task, its attempt is checked, and the check returns a reward that training uses to make better attempts more likely.
What it is made of
Inputs drawn from the institution's own cases, the tools the model may call while it works, a checker that scores each attempt, and limits on how many steps an attempt may take. For institutional work it is a copy of the real setting with no live effect: no message is sent and no record changes outside it.
When it is built
Only when supervised fine-tuning still leaves a measured gap and the result can be checked reliably, for example that every figure in a draft matches its cited source. Where no reviewable reward exists, reinforcement learning is skipped. The corporate credit memo use case shows that condition: it runs only with a checkable reward, and is skipped if supervised fine-tuning already passes the test.
Want these terms worked through on one of your own decisions? Start with a Decision Blueprint