Gym-Arc turns what a business already has (rulebooks, policy tables, case logs, agent trajectories and outcome feedback) into a Context Graph. From the graph it synthesizes whole families of worlds, compiles them into RL environments and Judges, trains the agent, and feeds every deployed conversation back in. The loop keeps closing, and the agent keeps getting better.
Pick an industry and a company above. Gym-Arc reads the company's raw material, infers its operating rules, checks itself, and builds a trainable environment for that company. No code is written for a new company, and no LLM sits in the loop, so the same pack always yields the same Context.
Handbooks written for people, policy tables kept by operations, years of case logs with what agents did and how each case ended, and a short questionnaire. Different formats, different owners, sometimes contradicting each other. This is the raw material.
Gym-Arc links each source to the evidence inside it, each piece of evidence to the operating rule it supports, and each rule to the environment dynamics and Judge checks that enforce it. Click any rule to see its provenance. Clauses with no slot in the schema are not dropped: they are flagged for review.
The Context Graph defines which rules can vary and how. Gym-Arc samples a family of companies around it, populates each with customers whose needs, trust and patience are hidden from the agent, and computes the best decision at every step from what the agent can actually see. Real companies slot into the same family.
Columns are companies, rows are rule dimensions. Brighter cells depart further from industry default. Outlined columns are the onboarded real-pack companies.
Each episode samples a customer. The agent sees the left column; the right column drives how the customer reacts and is only revealed by asking.
Expected value of each next action, computed by the environment from what the agent can observe. This is the supervision signal, and it changes with the company's rules.
E = BuildEnv(T, C, A, V): the task structure T is shared across the industry, the Context C comes from the graph, A is the action set, and the Verifier V is configured by the same Context. Breaking a company rule scores zero, whatever the outcome.
A compact model is trained on the synthetic company family, then evaluated on held-out companies: similar new companies (E2), new rule combinations (E3), values beyond the training range (E4), unseen rule dimensions (E5), and the onboarded real-pack companies (E6). Every hypothesis is pre-registered before the first training run.
Share of decisions that follow every company rule. Higher is better.
Value lost by the untrained model that training wins back.
The agent is deployed into a business world it was not built from, with a different customer mix and different dynamics. Each cycle it logs every trajectory and outcome. Gym-Arc recalibrates its world from those logs, retrains, gates the new model with the Judge, and redeploys.
Same held-out test decisions every cycle. Mean of two seeds.
Accuracy versus the number of logged business cases used for training.
Every trajectory and outcome that comes back is also evidence about the training ground itself: how customers really respond, which rules only show up in results, and where the gym is still guessing. Gym-Arc folds that evidence back into the Context Graph, so each version of the environment is closer to the real business than the last.
In the demo we onboard one of your data packs live, show the Context Graph it produces, and train against your rules.