This page is part of the guide to action-space constraints at a support desk, where a tier 1 agent may propose a refund but cannot issue one. Its job is a contents list: what a support team would be asked to put together once a licensing discussion starts, who prepares each item and which system it comes from.
Contents list, in assembly order
| Order | Item | What it shows | Who prepares it | Comes from |
|---|---|---|---|---|
| 1 | Task specification | The job, in the company's words | Head of customer operations | Support playbook |
| 2 | Role and permission map | Which role may take which action, with effective dates | Helpdesk administrator | Helpdesk and billing role settings |
| 3 | Policy text and versions | The refund rule and each earlier version | Support operations with finance | Policy library or knowledge base |
| 4 | Enforcement note | Which limits software blocks and which rely on training | Helpdesk administrator | Admin configuration |
| 5 | Action log field list | The fields recorded when someone proposes, approves or issues | Billing administrator | Helpdesk and billing logs |
| 6 | Acceptance check | The test a case must pass | Support operations, refined with SourceX | Written for the packet |
| 7 | Volume and span | Counts per year, years of history, migration dates | Support analytics | Helpdesk reporting |
| 8 | Rights and sensitivity notes | What limits reuse and what needs redaction | Counsel | Customer and vendor contracts |
The order matters. Items 1 to 4 define the rules; item 5 says where behavior under those rules was recorded; item 6 joins them into a test. Items 7 and 8 size the archive and set its limits. The graphic on this page drops the eight items into a box-file in the same sequence.
What goes into each item
1. Task specification. Two or three sentences. An Illustrative version: "Resolve a refund request on an annual plan according to the refund policy in force. Tier 1 may propose; only roles with refund permission may issue; amounts above the approval limit need a second approver."
2. Role and permission map. A table of roles against restricted actions (refunds, credits, cancellations, price overrides), with the amount limit and the date each setting took effect. This is the core of the packet; a current-only map loses the history a reviewer needs.
3. Policy text and versions. The written rule (for example, a full refund within 14 days of renewal) and every earlier version with dates. Paraphrase is fine; customer-facing wording is helpful where it differs.
4. Enforcement note. For each limit, whether the software blocks the action or staff are trusted to follow it. Research benchmarks draw the same line: τ-bench notes that some policy rules are enforced by the system and others depend on the agent applying them. Payment tools can provide 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.
5. Action log field list. User, role, action type, amount, reason code, approver, timestamps and linked ticket ID. Note which fields exist in which years.
6. Acceptance check. Written out in the next section.
7. Volume and span. Refund requests and escalations per year, years of history, and the date of any helpdesk or billing migration.
8. Rights and sensitivity notes. Customer names and payment details, employee identifiers, and any customer or outsourcing contract terms that limit reuse. Counsel lists categories, not individual tickets.
The acceptance check, written out (Illustrative)
The acceptance check turns a pile of tickets into something gradable. It checks the path, not just the amount. An Illustrative check passes a case only if all four hold:
- The refund was issued by a role permitted to issue that amount on that date.
- Any refund above $500 carries an approval by a user other than the requester.
- The refund falls within the policy window in force on the request date.
- No single request was split into several smaller refunds to stay under the limit.
Three short cases show how it reads:
| Case | Path taken | Result |
|---|---|---|
| A $1,150 renewal refund on day 9 | Tier 1 proposes, billing lead approves, system issues | Pass |
| A $1,150 renewal refund on day 9 | Tier 2 issues three refunds under $500 | Fail on test 4 |
| A $300 refund on day 20 | Tier 2 issues it inside its limit | Fail on test 3, outside the window |
The second case ends with the right amount in the customer's account and still fails. That is why the check exists: a refund of the right size issued by the wrong path is exactly the behavior a constrained agent must learn to avoid.
An Illustrative dated permission map entry
Illustrative and fictional. A 340-person subscription software firm writes item 2 like this. Dates replace adjectives.
| Action | Role allowed | Limit | Effective from | Previous setting |
|---|---|---|---|---|
| Apply account credit | Tier 1, tier 2, billing lead | Up to $100 | 2022-01-10 | Tier 2 and above only |
| Issue refund | Tier 2, billing lead | Up to $500 | 2024-03-01 | Up to $250 |
| Approve refund above limit | Billing lead, not the requester | Above $500 | 2024-03-01 | Above $250 |
| Cancel annual plan | Tier 2, billing lead | No amount limit | 2021-06-01 | Billing lead only |
The previous-setting column is what makes older escalations interpretable. Without it, a 2023 approval for a $400 refund looks unnecessary, when under the old $250 limit it was required. Where a date is not known, write "not recorded" rather than guessing; the evidence gap analysis explains how to describe that.
What the packet leaves open
The packet describes the archive; it does not show that the archive qualifies or is wanted. Data companies build business environments around permission models (Toloka's agent environments page, checked October 18, 2026, describes mockups that replicate real tool schemas, workflows and permission models), which shows interest in the category, not demand for any company's tickets. Nothing here claims that SourceX has already licensed this kind of record.
The company itself must have 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.
Open until SourceX reviews it: how many cases pass the acceptance check on sampling, the rights review result, redaction effort, buyer interest and any offer. Owners weighing the wider question can read whether a company can license its data to AI companies. To judge first whether an introduction is worthwhile, use the qualification scorecard. Related record families include customer support workflows and root cause analysis records.
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 the client's support lead could fill in a dated permission map like the one above, register as a partner, check who qualifies, and read the program terms before you introduce the decision-maker.