Tool-call provenance review packet: the eight metadata documents to assemble, in order

Short answer

A tool-call provenance review packet holds eight metadata-only documents, assembled in this order: a system inventory, an audit field dictionary, a rule catalog with revision history, an integration and credential map, a retention statement, an outcome linkage note, a rights summary, and a task specification with its acceptance check. Support operations, IT, finance and legal each prepare part. The packet describes what the records can show; it never contains the records themselves.

Visual briefFull-size diagram
Eight labeled documents, from the system inventory to the task specification with its acceptance check, slide one after another into an open folder marked Review packet, each tagged with the team that prepares it.
Illustrative packet contents. Every document describes records; none contains them.

This contents list belongs to the guide on tool-call provenance in a business workflow. That guide shows one traced status change; the packet is how a company proves its systems can produce thousands of them. Partners never assemble it. It is described here so an introducer knows what the company will be asked for later, under SourceX's agreement with the company, and can ask the right metadata questions up front.

Why the order matters

Each document depends on the one before it. A field dictionary means nothing until the reviewer knows which system it describes; a rule catalog cannot be read without knowing which audit field names the rule; an acceptance check cannot be written until the outcome linkage is known. Assembling in sequence also stops early: if the field dictionary shows no actor field on automated changes, the later documents can wait.

The sequence borrows its logic from W3C's provenance model, which separates the thing being changed, the change itself and whoever answers for it. Documents 1 to 5 pin down how the change and the responsible party are logged; document 6 follows the changed record forward to its consequences; documents 7 and 8 settle whether the archive may be licensed and what a correct case looks like.

Packet contents, in order

OrderDocumentWhat it showsWho prepares it
1System inventoryEach system involved (help desk, billing, CRM), its vendor, the years in use and any migrationsIT
2Audit field dictionaryThe fields each audit event carries: actor type, actor name, rule ID, revision, old and new value, timestampSupport operations with the platform admin
3Rule catalog and revision history summaryEach automated rule that changes status, when it was created, how often it was edited, and whether past versions surviveSupport operations
4Integration and credential mapWhich integrations write into the system, through which API user, and whether any login is sharedIT
5Retention statementHow many years of audit history exist, what is archived versus deleted, and any scheduled platform changeIT with legal
6Outcome linkage noteWhich key ties a ticket to its reopen, refund, dispute or next invoice, and where the link breaksFinance with support operations
7Rights and contract summaryWho owns the records, what customer and vendor terms say about reuse, and who can authorize licensingLegal, signed off by the CEO
8Task specification and acceptance checkOne routine automated task written as a testable case, with the conditions a correct run must meetSupport operations, drafted with SourceX

Document 2 is where a packet tends to succeed or stall. On some help desks it is short work: Zendesk's API documentation, for example, says each ticket audit carries a via object, and an individual event carries its own via when it differs from the audit's, as when a trigger makes a status change. A platform admin can describe those fields from the documentation and the configuration screen, without exporting a single ticket.

Document 8 in practice: one testable task

Document 8 turns the archive into something a reviewer can test. It is written before anyone looks at outcomes, so the standard cannot drift to fit the data.

Task specification (Illustrative). A fictional subscription company handles cancellation requests in its help desk. When a customer asks to cancel and the CRM integration has written the contract end date onto the request, the approved automation sets the request to Pending and schedules the plan to stop in the subscription system on that end date. Nobody stops a plan before the end date is confirmed, and nothing stops early.

Records the task depends on:

  1. The request's complete event history, every change rather than only the latest state.
  2. The automation's definition as of the version that fired, plus the dates each version was live.
  3. The CRM write that supplied the end date, with its source system and time.
  4. The scheduled stop in the subscription system, so the date can be compared independently.
  5. The follow-up window: reactivations, win-back requests from the same account, and card chargebacks within 60 days.

Acceptance check. A case passes only if all four hold: the named automation, not a person, scheduled the stop; the automation version that fired was the one approved on that date; the CRM end date was on the request before the stop was scheduled; and the account neither filed a chargeback nor complained of an early cutoff.

Four cancellations, scored against the check (Illustrative)

Every timestamp and version number below is invented to show how each document earns its place.

CaseWhat the records sayVerdictDocument that settles it
ACRM wrote end date 09:40; automation v12 scheduled the stop for that date 09:41; no complaintPass2, 3 and 6 read together
BAn agent stopped the plan by hand at 08:15, before the CRM had confirmed any end dateFail: acted before its input existed2 names a person as actor
CAutomation fired, but it was v11, which stopped plans the same day; v11 had been retiredFail: superseded logic ran3 gives the version dates
Dv12, right order, right date; the account later filed a chargeback for the final monthFail on follow-up6 ties the chargeback to the request

Case B is the one to remember. The plan would have ended anyway and the customer may never have noticed, so a check that looked only at outcomes would wave it through. Only the sequence of input and action, kept in the event fields, shows the step jumped ahead. Agent-evaluation vendors sell this kind of diagnosis: Appen, for one, advertises expert reference runs and failure analysis, both of which need to know which step broke. Treat that as a sign of market interest, not proof anyone wants a particular archive.

There is a reason documents 2 to 4 carry so much weight. When an AI agent invokes a tool under the Model Context Protocol, the call is valid only with an exactly listed name and schema-conforming arguments. A business automation's identity, version and input field play the same part, so a packet that cannot describe them cannot describe the call either.

What the partner does and does not touch

The partner's job ends at a qualified introduction. Before introducing, the partner can confirm that documents 1, 2 and 7 could be written, by asking the support or IT lead what systems exist, whether automated changes name their rule, and who could authorize licensing. The qualification scorecard turns those answers into a fit score, and the evidence gap analysis shows how to describe what is missing.

The partner never requests or forwards exports, screenshots, sample tickets or rule definitions. Partners sign the Channel Partner Agreement before submitting introductions. 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.

For the difference between tracing transformations and tracing origin, see data lineage vs data provenance. For how buyers judge what a packet describes, see how AI labs and data users evaluate datasets and why buyers want licensed data. Related record families are under tool-use training data.

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 company's support and IT leads say they could write the first two documents from what they already know, register as a partner, check who qualifies, and read the terms before you introduce the CEO, who would sign off document 7.

  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 the partner assemble the packet?

No. The company prepares it, after an introduction and under its own agreement with SourceX. The partner only needs to know the packet exists so the first conversation asks the right metadata questions.

Why does the packet contain no tickets?

Because it is a map of the records, used to decide whether a sample review is worth doing. Any review of actual records happens later, under the agreement, with de-identification agreed first.

How long does it take a company to prepare?

SourceX has not published a typical figure. Most of the documents are descriptions a support operations lead and an IT administrator can write from what they already know about their systems.

Sources

  1. developer.zendesk.com/api-reference/ticketing/tickets/ticket_audits
  2. w3.org/TR/prov-o
  3. appen.com/agentic-ai
  4. modelcontextprotocol.io/specification/2025-11-25/server/tools
  5. 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