# Clarification behavior review packet | SourceX

Canonical: https://refer.sourcex.si/resources/clarification-behavior-review-packet
Last updated: 2026-11-04

A clarification behavior review packet holds eight metadata-only documents that let SourceX judge whether a company's records show people pausing to resolve conflicting instructions before acting: a case class summary, a task sheet, the authority matrix by year, a conflict inventory, a map of where questions and answers live, an outcome linkage note, an acceptance check, and rights notes. It describes records; it contains none.

The packet belongs to the guide to [clarification behavior in an ordinary workflow](/guides/concept-clarification-behavior), which follows a renewal quote with a 12% and an 8% approval through a clerk's question to a clean invoice at 10%. Partners do not assemble this packet. It is set out here so an introducer knows what the company's records must be able to produce later, under SourceX's own agreement with the company.

## Eight documents, in order

The order runs from what kind of case this is, through who could decide and where each step is recorded, to how a correct handling is judged and who may license the records.

| # | Document | What it shows | Who prepares it |
|---|---|---|---|
| 1 | Case class summary | The kind of conflict, the systems involved, years covered | Finance or operations lead |
| 2 | Task sheet | The instruction staff were following, in one paragraph | Sales operations manager |
| 3 | Authority matrix by year | Who could approve which discount levels, for each year in scope | Finance controller |
| 4 | Conflict inventory | Rough counts of cases with two or more approvals for one customer and period | CRM or revenue operations administrator |
| 5 | Question trail map | Where questions and answers were recorded: CRM comment, ticket, email | Revenue operations administrator |
| 6 | Outcome linkage note | How quotes join to invoices, payments and credit notes | ERP administrator with finance |
| 7 | Acceptance check | The exact test for a correctly handled case | Finance and sales operations together |
| 8 | Rights notes | Customer contract limits, personal data present, parent-company terms | Legal counsel or the CFO |

## Inside each document

**1. Case class summary.** One paragraph: "Renewal and new-business quotes where two or more discount approvals exist for the same customer and period; CRM and ERP; 2021 to present."

**2. Task sheet.** The instruction as staff understood it. Illustrative: "Issue the renewal quote at the authorized discount. If approvals conflict, do not issue until the conflict is resolved by someone with authority for that discount level."

**3. Authority matrix by year.** Discount thresholds and approver roles, one version per year. Without it, nobody can check whether a question went to the right person.

**4. Conflict inventory.** Counts, not cases: how many quotes per year had more than one approval, and how many were put on hold. Estimates are fine if labeled.

**5. Question trail map.** Which system holds the question and which holds the reply, and whether they share the quote or customer ID. A question asked in a chat tool that is not linked to the deal should be named here, not hidden.

**6. Outcome linkage note.** Whether invoices, payments and credit notes carry the quote number or customer ID, and for which years.

**7. Acceptance check.** Set out in full in the next section.

**8. Rights notes.** Customer contracts that limit use of deal records, personal data in comments (names, direct lines), and any parent-company or reseller terms. Questions for review, not legal conclusions.

## The acceptance check, with a failing case (Illustrative)

A case passes only if all four conditions are true:

1. The quote was not issued while two approvals conflicted.
2. The question went to someone whose role, under the authority matrix in force that day, covered the final discount.
3. The issued quote matches the confirmed figure exactly.
4. The invoice and payment match the quote, with no later credit.

Two fictional cases from the same distributor show how it works.

| Condition | Case A: asked first | Case B: guessed first |
|---|---|---|
| Not issued during conflict | Pass: on hold for one day | Fail: issued at 12% the same morning |
| Question to authorized person | Pass: sales director, covers 10% | Not applicable: no question asked |
| Quote matches confirmed figure | Pass: 10% | Fail: 12% issued, 10% confirmed later |
| Invoice and payment match | Pass: paid in full | Fail: credit note issued to reach 10% |
| Result | Passes | Fails, though the customer ended at the right price |

Case B is the important one. It ends at the right number, and a record of final prices alone would rate it correct. The check fails it because the clerk guessed first. OpenAI's published Model Spec frames the same trade-off for AI systems: attempt by default, but ask when a wrong guess is costly. Researchers behind the CLAMBER benchmark found models weak at exactly this kind of clarifying move, which is why records that separate case A from case B can be useful.

## Preparing the packet without moving any records

Every document above can be written from what staff already know or can count inside their own systems. Nothing is exported, forwarded or shared with the introducer. The CRM administrator counts conflicts; nobody sends the conflicts. The ERP administrator describes how invoices link to quotes; nobody attaches an invoice.

That is the line for partners too. Discovery stays at metadata level: what the records are, roughly how many, how many years, which systems, who owns them. If an owner offers to forward example threads, decline and point them to SourceX's process.

## What the packet leaves open

Sales operations, finance and IT usually share ownership of these records, and the CEO, usually the owner in an owner-run company, authorizes a licensing discussion. The company must have 50 or more 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.

The packet cannot settle how many cases pass the check, what de-identification is needed, whether customer contracts allow use, or whether any buyer wants this task type. Vendors such as Appen market (on a page checked October 18, 2026) agent evaluation work for CRM and ERP workflows, but competitor interest in agent data does not mean every approval archive qualifies. Some buyers want material only for testing, which suits an [evaluation-only license](/compare/evaluation-only-data-license); for test material generally, see [test case data](/data-types/test-case-data-ai-training). Owners often ask first [whether a company can license its data to AI companies](/questions/can-a-company-license-its-data-to-ai-companies) and [why sales data is valuable](/proof/why-sales-data-is-valuable). Before any packet, rate the company with the [clarification behavior qualification scorecard](/questions/does-clarification-behavior-data-qualify) and describe weak spots using [clarification behavior evidence gaps](/questions/clarification-behavior-evidence-gaps).

## 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 finance lead believes the company could produce documents 3 to 6 for several years of quotes, [register as a partner](/register), confirm the basics on [who qualifies](/who-qualifies), and read the [terms](/terms) before you introduce the CEO, with the CFO as sponsor.

## Sources

- https://model-spec.openai.com/2026-08-18.html
- https://aclanthology.org/2024.acl-long.578/
- https://www.appen.com/agentic-ai
- https://refer.sourcex.si/terms
