attorneyrl.com / coming soon / reinforcement learning for legal ai

The gym for legal agents.

AttorneyRL is a reinforcement learning environment where AI agents work simulated legal matters — dockets, deadlines, discovery, motions — and earn reward for outcomes a lawyer would defend, not text that only sounds right.

episode
0
step
0
ε (explore)
—
return
0.00
last outcome
—
Background: an agent climbing a live reward surface. Real simulation, illustrative task.
§ 01 — the environment

Litigation, posed as a Markov decision process.

Every matter is an episode. The agent reads the record, picks a move, and the environment answers with the next state of the case — and a reward computed from rules, not vibes.

s

Observation

The state of the matter as a lawyer would see it on a given day.

  • pleadings & docket entries
  • governing rules & local rules
  • deadlines & the calendar
  • produced documents & privilege log
a

Action

Structured, auditable moves — each one checked before the state advances.

  • file · serve · respond
  • move · oppose · object
  • request · produce · withhold
  • negotiate · settle · appeal
r

Reward

Scored by verifiable checks on the outcome, with hard penalties for the mistakes that end careers.

  • favorable, defensible outcome
  • deadlines met, rules satisfied
  • missed deadline · waived right
  • fabricated citation · sanctions
§ 02 — watch it learn

Reward goes up. Sanctions go away.

Two live simulations, running in your browser right now. Left: episode returns from the agent in the background as its exploration decays. Right: a toy matter graph, walked one move at a time.

Episode return · moving avg (20)simulated · live
Matter graph · one trajectoryG = 0.00
§ 03 — the interface

Familiar API. Unfamiliar stakes.

A Gym-style loop your training stack already speaks. Swap the jurisdiction, the matter type, or the reward model — keep the loop.

import attorneyrl env = attorneyrl.make("FedCivPro-v0") obs, info = env.reset(seed=7) while True: action = policy(obs) # e.g. File(MotionToDismiss, ...) obs, reward, done, trunc, info = env.step(action) if done or trunc: break info["audit"] # every check, every rule cited

Preview of the planned interface. Names may change before release.

§ 04 — early access

Train agents that can stand up in court.