What an RL environment needs from a real business archive: task, state, tool and outcome records

Short answer

An RL environment for business work needs five things from records: a task request, the starting state of the relevant systems, the tools and actions allowed, the policy that limits them, and an end state a program can check. Tickets and emails supply tasks, change histories supply starting states, system permissions supply tools, policy documents supply limits, and the final record values supply the outcome. Archives that keep field-level history are the ones that can supply all five.

Visual briefFull-size diagram
Five stacked requirement rows, task, starting state, allowed tools, policy and end state, each paired with the business record that supplies it, with the starting state row flagged as the one most archives lack.
Illustrative return request. Each environment component maps to a record type; point-in-time state is the usual gap.

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.

ComponentWhat the environment needsBusiness record that supplies itTest question for the owner
TaskA realistic request with enough detail to act onInbound emails, tickets, web forms, purchase requestsAre requests stored with their original wording?
Starting stateThe values of every relevant record when the request arrivedField-level change history or audit logsCan you see what a record looked like on a past date?
Tools and actionsThe operations a person was allowed to performSystem permissions, screen or API actions, transaction codesWhich systems did staff use to complete this work?
PolicyThe rules that limit which actions are correctVersioned policies, SOPs, approval thresholdsAre old policy versions kept with effective dates?
End stateA final, checkable resultFinal record values, postings, status codes, sign-offsIs 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.

ComponentValue drawn from records
TaskCustomer email: "Unit stopped charging, I'd like to return it for a refund."
Starting stateOrder delivered 38 days ago; item covered by a 12-month warranty; no prior returns on the account; refund status none
Allowed toolsOrder lookup, return authorization (RMA) creation, refund issue, warranty repair booking, customer reply
PolicyRefunds within 30 days of delivery; after 30 days, defective items go to warranty repair
Correct end stateRMA 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 endExpectedAgent that refundedAgent that followed policy
RMA typeWarranty repairReturn for refundWarranty repair
Refund postedNoYesNo
Repair bookedYesNoYes
ResultNot applicableFailPass

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.

AskStrong answerWhat it tells SourceX
Which systems hold the workflow, and since when?Named tools, five or more yearsDepth and continuity
Is change history or an audit log kept for key records?Yes, with retention knownWhether starting states can be rebuilt
Are requests stored in their original wording?Yes, emails or tickets linked to the recordSupplies realistic tasks
Are policies versioned with effective dates?YesLets each task be judged by the rule in force
Who can authorize a licensing discussion?CEO, usually the owner in an owner-run companyThe 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.

  1. Step 1Share your linkSend your personal link to a company you know.
  2. Step 2Company appliesThe company applies itself at /apply.
  3. Step 3Company signs and is paidThe company signs, delivers its data and is paid by SourceX under its agreement.
  4. Step 4You get your rewardthe greater of $50,000 plus 5% and 15% of the actual supplier payout, with no cap. A reward becomes payable after SourceX has paid the referred company under its signed agreement and the payout conditions in the program terms are met, including finance approval and your tax and payment details.

Common questions

Does SourceX build RL environments?

No. SourceX licenses approved business data directly from companies for AI training and evaluation, and manages the agreement, rights review, data preparation, delivery and payout. Environment building is done by AI developers and their vendors.

Which component is hardest for a company to supply?

Usually the starting state: what every relevant record looked like at the moment the task began. Systems that keep field-level change history make it possible; systems that only store the current value do not.

Does the company need to know what an RL environment is?

No. The owner only needs to know which systems hold the work, how long the history goes back and whether changes are logged. SourceX handles the technical assessment.

Sources

  1. gymnasium.farama.org/api/env
  2. arxiv.org/abs/2406.12045
  3. toloka.ai/agent-data
  4. refer.sourcex.si/terms

Free resources

Know a company with valuable proprietary data?

Become a referral partner, get your link and introduce the company’s CEO (in an owner-run company, usually the owner). Referral partners can join from most countries. We can't work with partners in countries or regions under U.S. sanctions or on U.S. restricted-party lists. The companies you introduce should be based in the U.S., Canada and the EU, or in the UK, Australia, New Zealand, India, Singapore, South Africa, Latin America, the UAE and Saudi Arabia with at least 70% of their data in English.

Refer a company →

I own a business

Explore licensing your company's data directly to SourceX for AI training and evaluation. Start a short assessment; no uploads needed.

Start an assessment