Your agent shines in the demo, then breaks at the next client with a different rulebook. Gym-Arc turns each company's rules into a decision environment you can train on and verify against, so your model learns to read the rules before it acts, and proves it on companies it has never seen.
Product shelves, approval thresholds, eligibility policy, disclosure timing, house preferences. A general model can read the handbook, yet it cannot tell what a rule is worth at this step.
We tuned it for our first client. At the second one it started breaking rules.
The model memorised one company's playbook instead of learning to read rules.
GYM-ARCTrain across many company environments; sign off on companies the model has never seen.
We rewarded closed deals, and the model learned to push.
Pressure and guaranteed-return promises close deals too. Rewarding the outcome rewards the violation.
GYM-ARCAny violation scores zero, reward follows step-level decision quality, and deals are only reported.
Expert labels are slow and costly, and they only say which option to pick.
A hard label never says what a wrong choice costs, so the model cannot learn trade-offs.
GYM-ARCThe environment prices every action and produces soft labels that carry the cost.
Start with evaluation, move to training environments, then to cross-company adaptation. All three share the same environments, Verifier and data pipeline, so you can go deeper as you need.
For · vertical agent companies, enterprise AI teams, model selection
Delivered as eval sets + Verifier + report and console
For · industry model and agent teams with a training budget
Delivered as environment + labelled datasets + training recipe
For · frontier labs, general-purpose agent teams
Delivered as environment family + held-out evals + transfer evidence
Onboarding does not rely on an LLM skimming your handbook. Every rule value has a source, and every conflict or ambiguity comes back to you for sign-off.
Compliance handbook, product and approval tables, case history, plus a structured intake questionnaire. We provide the templates.
Tables are parsed, behavioural rules are measured from cases, and questionnaire answers are checked line by line against the handbook. Same pack, same result, every time.
Task structure, your rules, actions and Verifier combine into one environment. From rules written to environment ready takes about 0.06 s, with no code changes.
Train your model, then sign off on held-out customers and companies under criteria fixed in advance. Every number in the report is traceable.
Your rules, structured and versioned
contexts/<company>.yamlSource and confidence of every rule, plus uncovered clauses verbatim
report.jsonStandard interfaces with action masks and illegal-action handling
GymArcEnv · make_trl_env_classFour-layer verdicts; any failed constraint scores 0
constraints · quality · workflow · outcomeQ values and soft labels for every action at every step
SFT · RL readyNew customers, firms, combinations, ranges, dimensions, outside partners
E1 – E6 · OODExpert-trajectory SFT, GRPO configs and eval scripts
ga grpo · configs/trainingStep-by-step replay: what the model saw vs the hidden truth
console · offline snapshotSame model, different training environments: one company versus twelve, then tested on held-out companies. The study design was committed before training began and the results were judged by rules written in advance.
12-company vs single-company training; value lost per decision falls from 3.10 to 0.97
Companies whose rules were each seen, but never together (2.60 → 0.72 value lost)
Fine-tuned decision model vs a general LLM given the rules, and 112× faster
Six company data packs onboarded automatically; the remaining 2 rules were flagged for sign-off
Measured on Gym-Arc benchmark environments built from real-world industry rulebooks, regulations and sales workflows. Full methodology, 95% confidence intervals and per-seed results are available on request.
Insurance decision benchmark, 600 held-out questions. A compact model trained in Gym-Arc against leading general-purpose LLMs. Up and to the left is better.
Load an insurance brokerage, a card centre, an online store or a business bank and follow its own data through Gym-Arc: raw handbooks and case logs become a Context Graph, then synthetic worlds, an RL environment and Judge, a trained agent, and a deployment loop that keeps improving it.
The same rule costs different amounts at different decision points. Gym-Arc computes that consequence structure, and that is what it takes to train a model that reads rules.
At every decision point, every legal action is rolled out counterfactually to get its expected value and regret. Models learn trade-offs, beyond right and wrong.
Labels use only what the model can observe. Switching to these labels cut decision loss in bancassurance sales by more than half.
Fail any constraint and the episode scores 0. Business outcomes are reported apart and never enter the reward, so violations never pay.
Companies differ by one rules file. Adding one takes zero code and under a second from rules to a ready environment.
Describe customers, actions, rules and rewards in one YAML file; the framework supplies simulator, gold, Verifier and experiments. Three new industries took 0 lines of domain code.
Study designs are committed before training with git timestamps to prove it, and every number in the docs and console is re-checked by automated guards.
Each task family is built from real industry rulebooks and workflows, with lifelike customers, hidden state, company rule dimensions and multiple company environments. If your industry is missing, one declaration file adds it.
Needs analysis, disclosure, supervisor approval, visitor eligibility, referral to partner insurers.
Multi-turn branch conversations with customers who carry trust, patience and objections, and leave if handled badly.
Read one referral lead and choose the best next move.
Outbound offers of bill or cash instalments, where rate disclosure and do-not-disturb rules vary by bank.
Refunds, compensation, proof requirements and fresh-produce rules that differ by store.
Due-diligence documents, credit-check consent and eligibility thresholds. The same applicant can have a different best next step at another bank.
If decisions are multi-step, customer state can be described discretely and company rules fit a handful of dimensions, one domain.yaml brings it in, with an author guide that walks you through it.
You don't start with customer data. The environment is built from your rulebook and tables; case history, when available, calibrates rules that only show up in outcomes, such as approval turnaround or underwriting strictness.
The contract spells out five items: ownership of source data, the general framework, task and evaluation methods, synthetic assets derived from your rules, and whether anything may be licensed to other clients. It is in the agreement from day one.
Any model you can train through Gymnasium or Hugging Face TRL, plus custom structured decision models. General LLMs can be plugged in as baselines or judges.
An LLM that reads the rulebook still cannot tell what a rule is worth at this step. In our comparison, a decision model trained in the environment was 17.9 points more accurate than a general LLM given the rules, and 112× faster.
Yes. Environments, onboarding and training are Python and run on your GPU servers; the console is a static page that also works offline.
On the platform side there is no code to change, and rules become a ready environment in under a second. The overall timeline mostly depends on preparing documents; we provide a handbook skeleton, table templates and an intake questionnaire to get it right the first time.
In the demo we build an environment from one of your rulebooks on the spot: the same customer under your rules and under industry defaults, how the best next step differs, and whether the model picked it.