# Action-space constraints: evidence gap analysis | SourceX

Canonical: https://refer.sourcex.si/questions/action-space-constraints-evidence-gaps
Last updated: 2026-11-04

Four gaps make action-space constraints hard to assess: missing context on the request, missing evidence of who acted and in what role, limits that changed without a dated record, and outcomes stored in another system with no link back. In a refund case they decide whether anyone can confirm that the person who issued a refund was allowed to. None of them rules a company out; each narrows how much of the history can be used, which SourceX measures in its review.

This page is part of the guide to [action-space constraints at a support desk](/guides/concept-action-space-constraints), where a tier 1 agent may propose a refund but cannot issue one. Its job is a gap analysis: set what a review would expect beside what support desks usually keep, then show how to describe the difference honestly.

## What a review of a permission history expects

An action space is the set of moves available in a task; in reinforcement learning libraries such as Gymnasium, every environment declares which actions are valid. A business history is useful when it shows the same thing for people: what each role could do, what they tried, and whether the limit held. For a refund, a reviewer looks for six pieces of evidence.

| Expected evidence | Why the review needs it |
|---|---|
| Request context on the ticket | Shows the policy test (for example, the 14-day renewal window) could be applied |
| Role of the person who issued the refund | Shows the action was within that role's permissions |
| Approver distinct from the requester | Shows a required second approval actually happened |
| Limit in force on the date | Shows whether an escalation was required at that amount |
| Linked outcome | Shows whether the refund stood, was reversed or became a dispute |
| Enforcement type | Shows whether software blocked the action or staff were trusted |

The last row echoes research practice. τ-bench notes that some policy rules are enforced by the software while others depend on the agent applying them, and the two need different evidence. Payment tools can supply the first kind: Stripe documents roles with refund permission removed (custom roles are in private preview) and two-party approvals, in public preview, that hold a refund until a different team member approves it.

## The four gaps in a refund history

Each gap below is shown in the main guide's refund case, with what support desks usually keep and what that loses.

| Gap | How it shows up | What desks usually keep | What a reviewer loses |
|---|---|---|---|
| Missing context | Tier 1 note says "refund requested" with no invoice or renewal date | The ticket text and status | Proof that the 14-day window was checked |
| Missing permissions | Refund recorded against a shared login, or with no role | The account name on the payment record | Any way to verify the role limit held |
| Unclear versions | Approval threshold moved from $250 to $500 in March with no record | The current setting only | Whether pre-March escalations were required |
| Unproven outcome | Refund issued, then a chargeback filed in a separate dispute tool | The refund ID | Whether the fix stood or was reversed |

Shared logins and undated threshold changes are the two to ask about first. Both are common at growing support desks, especially on night or weekend shifts, and both decide whether the constraint can be reconstructed at all.

## One broadband provider's refund history, measured (Illustrative)

**Illustrative and fictional.** A 280-person business broadband provider has handled about 5,200 refund and credit escalations since 2020 in one helpdesk, with refunds issued in a separate billing tool. An introducer asks the head of customer operations how often each piece of evidence is present, and records the answers as rough shares.

| Evidence | Expected | Available (owner's rough share) | The gap in one line |
|---|---|---|---|
| Request context | Every escalation | 65% | Invoice references became a required field in 2022 |
| Issuer role | Every refund | 50% | The weekend team shared one billing login until 2023 |
| Separate approver | Every refund above the limit | 70% | Approvals logged in billing; a few by email |
| Threshold version | Every escalation | 40% | The 2021 limit change was made but never dated |
| Linked outcome | Every refund | 55% | Chargebacks live in the payment processor without a ticket ID |

Read across, the usable core is clear: escalations from 2023 onward, after the shared login ended and once invoice references were required. Earlier years are thinner but not worthless; they may still show proposals and approvals, just not who issued each refund. The graphic on this page plots the same five shares against what a review would expect.

## Wording each gap for the introduction

Rounded-up descriptions create problems at review. Plain ones save time.

| Instead of | Say |
|---|---|
| "Every refund is approved properly." | "Refunds above the limit carry a logged approval in most cases; some approvals went by email." |
| "We know who did everything." | "Issuer roles are reliable from 2023; before that the weekend team shared a login." |
| "Limits have always been the same." | "The approval limit changed in 2021; the exact date is not recorded." |
| "We track outcomes." | "Refund IDs are kept; chargebacks sit in the payment processor without a ticket link." |

Two habits help. Give years, not adjectives, and say who gave each answer so the review knows whom to ask. Gaps found and explained by the company read better than gaps discovered later. Records of why limits changed, where they exist, are a record type of their own; see [architecture decision records](/data-types/architecture-decision-records) and [root cause analysis records](/data-types/root-cause-analysis-records) for neighboring families.

## What gaps do not change, and what stays unknown

Gaps do not change ownership or the company test. Support records usually belong to customer operations, with finance owning refunds and IT administering the tools; authority to discuss licensing sits with the CEO. The company needs 50+ full-time employees at peak (contractors excluded), documented operating history, rights to license 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.

Unknown until SourceX reviews it: the real share of usable escalations, how customer and payment data will be handled, whether an outsourced provider's contract affects rights, and whether any buyer wants this task type. Interest across the industry is real, as [why AI companies pay for data](/proof/ai-companies-pay-for-data) explains, but it does not mean every support archive qualifies. To decide whether the client is worth introducing, use the [qualification scorecard](/questions/does-action-space-constraints-data-qualify); for what the company assembles later, see the [review packet contents list](/resources/action-space-constraints-review-packet).

## 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 you can state a client's permission-history gaps in the plain terms above and name the records owner, [register as a partner](/register), check [who qualifies](/who-qualifies), and read the [program terms](/terms) before you make the introduction.

## Sources

- https://docs.stripe.com/account/approvals
- https://docs.stripe.com/get-started/account/teams/custom-roles
- https://arxiv.org/abs/2406.12045
- https://gymnasium.farama.org/api/env/
- https://refer.sourcex.si/terms
