Which task, state, tool and outcome records can support an RL environment?
Each component of an environment has a matching record type in an ordinary company. The vocabulary comes from reinforcement learning practice. In the widely used Gymnasium API, an environment is reset to a starting observation, and each action the agent takes returns a new observation, a reward and a signal that the episode has ended. Benchmarks for business agents follow the same pattern with business systems. τ-bench, for example, gives an agent API tools and written policy guidelines, lets the hidden database change only through those tools, and scores the episode by comparing the final database with an annotated goal state.
Translate that into records and you get a requirements list an introducer can actually use.
| Component | What the environment needs | Business record that supplies it | Test question for the owner |
|---|---|---|---|
| Task | A realistic request with enough detail to act on | Inbound emails, tickets, web forms, purchase requests | Are requests stored with their original wording? |
| Starting state | The values of every relevant record when the request arrived | Field-level change history or audit logs | Can you see what a record looked like on a past date? |
| Tools and actions | The operations a person was allowed to perform | System permissions, screen or API actions, transaction codes | Which systems did staff use to complete this work? |
| Policy | The rules that limit which actions are correct | Versioned policies, SOPs, approval thresholds | Are old policy versions kept with effective dates? |
| End state | A final, checkable result | Final record values, postings, status codes, sign-offs | Is the final status recorded in a structured field? |
Vendors that build environments for AI developers describe this work in their own terms. Toloka, for example, says it builds environments and RL gyms and collects trajectories and graded evaluation signals. The RL environments explainer covers the concept, and why RL environments depend on real company workflows covers which workflows tend to become tasks. This page is narrower: which records each component needs.
A worked example: one return request, component by component
Illustrative. A fictional 400-person electronics distributor receives a customer email asking to return a device delivered 38 days earlier. Every detail is invented.
| Component | Value drawn from records |
|---|---|
| Task | Customer email: "Unit stopped charging, I'd like to return it for a refund." |
| Starting state | Order delivered 38 days ago; item covered by a 12-month warranty; no prior returns on the account; refund status none |
| Allowed tools | Order lookup, return authorization (RMA) creation, refund issue, warranty repair booking, customer reply |
| Policy | Refunds within 30 days of delivery; after 30 days, defective items go to warranty repair |
| Correct end state | RMA created with type "warranty repair"; no refund posted; customer reply sent; repair booked |
An agent that reads the email and issues a refund has done what the customer asked and broken the policy. An end-state check catches it without anyone reading the conversation:
| Field checked at the end | Expected | Agent that refunded | Agent that followed policy |
|---|---|---|---|
| RMA type | Warranty repair | Return for refund | Warranty repair |
| Refund posted | No | Yes | No |
| Repair booked | Yes | No | Yes |
| Result | Not applicable | Fail | Pass |
Every expected value comes from a record the company already keeps: the delivery date, the warranty term, the policy in force and, crucially, how its own staff resolved hundreds of similar emails. The verifiable rewards guide explains why end-state checks like this are attractive to model developers, and the tool use data page covers the action side.
Why is the starting state usually the gap?
Most systems store the current value of a record, not its history. To rebuild the moment the email arrived, an environment builder needs the order, account and warranty fields as they were 38 days after delivery, not as they are today after the return was processed. That needs field-level change history: an audit trail, a history table, or event logs that record each change with a timestamp.
Whether a company's systems kept such history, and for how long, varies by tool, configuration and retention settings. That is a question for the company's system owner during SourceX's review, not something a partner needs to establish. What a partner can usefully ask is the plain-language version: "If I asked what this order looked like on a date last year, could your team tell me?"
Company qualification questions
Discovery stays at metadata level. Partners never ask for or handle raw records.
| Ask | Strong answer | What it tells SourceX |
|---|---|---|
| Which systems hold the workflow, and since when? | Named tools, five or more years | Depth and continuity |
| Is change history or an audit log kept for key records? | Yes, with retention known | Whether starting states can be rebuilt |
| Are requests stored in their original wording? | Yes, emails or tickets linked to the record | Supplies realistic tasks |
| Are policies versioned with effective dates? | Yes | Lets each task be judged by the rule in force |
| Who can authorize a licensing discussion? | CEO, usually the owner in an owner-run company | The sponsor to introduce |
The introduced company needs 50+ full-time employees at peak (contractors excluded), a documented operating history, rights to license the records and an authorized contact. Companies in the U.S., Canada and the EU qualify, as do companies in the UK, Australia, New Zealand, India, Singapore, South Africa, Latin America, the UAE and Saudi Arabia when at least 70% of the data is in English.
What stays unknown until SourceX reviews it: whether buyers currently want this workflow, whether the history goes back far enough, how customer data will be redacted, the rights review result and any offer.
Limits and unanswered questions
This page explains a requirements pattern; it does not claim that any particular archive will be used to build an environment, or that any named developer buys such records. SourceX does not build environments or graders. It licenses approved business data directly from companies and manages the agreement, rights review, data preparation, delivery and payout.
Real archives also contain cases with no single correct end state, where a manager made a judgment call. Those can still be useful for rubric-based evaluation, but they are weaker for automatic checks. The case for enterprise data covers the broader demand picture.
Company compensation and referral rewards are separate. The company is paid under its own agreement with SourceX; the partner's reward is calculated on what SourceX actually pays the company and is funded separately. Illustrative only: on a first qualifying transaction paying USD 400,000, the reward would be USD 70,000. See how rewards work.
Next step
Refer a company. No cap on rewards. Default reward on a company's first qualifying transaction: the greater of $50,000 plus 5% and 15% of Actual Supplier Payout, with no cap; further qualifying transactions first agreed within 12 months of the first Closing pay 15%. Eligibility and payment conditions apply, including the USD 100,000 Minimum Transaction, and a signed Schedule A or addendum can differ. A lead, meeting or signature alone does not trigger payment.
If a client can say what its records looked like on a past date and keeps policies by version, register as a partner, check the company against who qualifies and read the program terms before introducing its decision-maker. More agent data topics are in the guides hub.