Getting started
Workflow first: how to request SourceX specialist operational datasets
By SourceX Editorial · Reviewed by Noah Loul ·
10 sections
Short answer
Request SourceX specialist operational datasets by naming one decision first: its trigger, the conditional answers allowed, the records that existed when it was made, how the right answer became known, the look-alike cases, the rights evidence and the acceptance test. The industry field comes last because it settles none of these. In public 2024 single-audit data, each of the five agencies with the most findings spanned all 13 compliance areas.
Checklist
Workflow first
Key takeaways
- A specialist request should name one operational decision and its conditional answers before it names an industry.
- Seven parts make the packet: decision, answers, decision-time evidence, outcome and lag, hard negatives, rights evidence and acceptance test.
- Among the five agencies with the most 2024 single-audit findings, no compliance area covered more than 38.1 percent of any one agency's findings.
- Score requests out of 16 and send at 12 or more, with points on the decision, rights and acceptance gates.
- SourceX sources on request and checks buyer fit during review; a precise request improves the match but promises no supply or result.
What should a specialist data request lead with?
Lead with one decision. A request for SourceX specialist operational datasets works when it names the operational call a model must make, the conditional answers that call allows, the records that existed when it was made, how the right answer became known, the cases that only look alike, the evidence that the holder may license the records, and the test the delivery must pass. The industry comes last, because it settles none of those.
This series has built that structure for one decision at a time: a construction change order approved with open items, a track defect that can wait only on stated terms, a superseded product on a wholesale order, a legal instruction from someone whose authority is unproven, a repair report that tenants dispute, and a grant application whose eligibility evidence conflicts. In each, the right answer is often partial or deferred: ask for a named document, hold, approve with conditions, substitute or escalate.
This brief turns those cases into a request a buyer can send: a seven-part workflow packet, a scoring rubric with a send gate, a domain review checklist and the fields on SourceX's request form that carry each part.
Why a sector label settles none of the hard questions
A sector label answers no question a licensing decision turns on. It does not say who may license the records, whether the records fit a safe use, or whether a model trained on them will generalize to the buyer's setting.
Rights first. SourceX's screening page asks for original, well-documented proprietary data, a decision maker who can authorize a licensing discussion and the rights to license that material, and it lists material that belongs to the company's clients rather than the company as usually not a fit. Two firms in the same industry can sit on opposite sides of that line: one created its review records, the other holds its clients' files.
Suitability next. NIST's AI Risk Management Framework asks that intended purposes, context-specific laws, norms and expectations, and the settings where a system will be deployed be understood and documented. An eligibility triage assistant and an automated denial tool may use identical records and carry very different risks, and only the request's statement of intended use tells a supplier which one it is serving.
Then buyer fit. SourceX's screening page says the proprietary material has to match a real program with AI developers and that SourceX checks this during review. A request that names the decision is what lets that check happen; a request that names a sector leaves both sides guessing.
In public audit data, the sector label barely predicts the workflow
One public record set shows how weakly a sector label predicts the work inside it. Single-audit findings filed with the Federal Audit Clearinghouse carry two labels at once: the federal agency whose program was audited, a sector label, and the compliance requirement the finding concerns, a workflow label such as eligibility, procurement, reporting or special tests and provisions, as OMB's Compliance Supplement defines them.
We cross-tabulated audit year 2024: 16,189 distinct findings, counted once under each agency whose award they touched, giving 18,956 finding-agency pairs across 42 agencies. Each of the five agencies with the most findings spanned all 13 requirement areas, and no single area covered more than 38.1 percent of any of those agencies' findings. Read the other way, every requirement area appeared under 17 to 39 agencies, and no single agency held more than 37.6 percent of any area's pairs. Eligibility, the workflow behind this series' grant case, appeared under 26 agencies, with Housing and Urban Development holding the largest share at 32.4 percent.
The practical reading: a buyer who asks for housing data, or education data, has described a mix spread across all 13 compliance areas. The decision names the work. The same files and denominators feed the grant administration article in this series, so the eligibility figures in both pages match:
| Funding agency (a sector label) | Findings under its programs | Requirement areas present, of 13 | Largest area and its share of the agency's findings |
|---|---|---|---|
| Health and Human Services | 3,834 | 13 | Reporting (L), 26.9% |
| Education | 3,717 | 13 | Special tests and provisions (N), 38.1% |
| Housing and Urban Development | 3,518 | 13 | Special tests and provisions (N), 32.7% |
| Treasury | 2,396 | 13 | Reporting (L), 34.5% |
| Agriculture | 1,779 | 13 | Reporting (L), 26.2% |
The specialist workflow packet: seven parts to send
Send seven parts, each answering one question a supplier and a reviewer must settle before records move. The right-hand column shows each part for the grant eligibility case; the middle column shows what a sector-first request usually says instead, and why it stalls:
| Part | Question it answers | Sector-first version | Workflow-first version (grant eligibility) |
|---|---|---|---|
| Decision | Which call, triggered by what, at which point? | Government grants data | Eligibility calls at intake where two evidence items disagree |
| Conditional answers | Which actions count as right? | Approved or denied | Refer, escalate, decline, request a named document, condition, eligible |
| Decision-time evidence | What did the decision maker have? | All application records | Application, dated registry captures, rule version, reviewer checklist, requests and replies |
| Outcome and lag | How is the right answer known, and when? | Outcome labels | Condition clearance, appeal and monitoring results with dates; open files marked censored |
| Hard negatives | Which cases only look alike? | Not mentioned | Assumed names, near-name exclusion hits, stale determination letters |
| Rights evidence | Who may license these records? | Licensed data | Records the administrator created, with funder permission; personal details removed or excluded |
| Acceptance test | What must the delivery pass? | High quality | Blind sample relabeled by two domain reviewers; pass rule fixed before delivery |
One decision shape, many trades: the cases in this series
What changes from trade to trade is the rulebook and the record system, and both belong in the packet as inputs, not as things a model should learn. A track defect waits under federal track rules, a grant file under the funding notice in force that round, a superseded part under the customer's contract. A request that asks the supplier to attach the governing rule version to every case lets one label design travel across suppliers while each case is still judged by the rule that actually applied to it.
Across trades the records change and the decision shape does not. Every case below needs a label set wider than yes or no, evidence frozen at the decision date and an outcome read from a later record. That shared shape is why a buyer can specify a workflow precisely without knowing the industry's jargon, and why a supplier in an unexpected industry may still fit:
| Trade | Decision | Conditional answers the labels must allow | Records that settle it |
|---|---|---|---|
| Construction | Approve a change order that leaves items open | Approve with open scope, price basis, time or dependency | The change order family that followed |
| Rail track maintenance | Let a defect wait | Restrict and schedule with a dated repair limit, monitor, escalate | Restriction, authorized decision maker, repair limit |
| Wholesale distribution | Fill an order for a superseded product | Ship the successor with notice, hold for approval, quote alternatives | Dated item status, interchangeability, customer terms |
| Legal intake | Act on an instruction of unproven authority | Proceed with conditions, hold for a named document, escalate | The authority document and later matter records |
| Property maintenance | Act on disputed repair reports | Make safe now, inspect first, request evidence | Competing claims with source, time and basis |
| Grant administration | Decide eligibility when evidence conflicts | Request a named document, approve with conditions, escalate | Dated registry captures, rule version, requests and replies |
Score the request before you send it
Score a request on eight criteria worth up to two points each, and send it when it reaches 12 of 16 with at least one point on each gate: the decision, the rights evidence and the acceptance test. A request that misses a gate fails for a reason no volume figure can fix.
A script scored three illustrative requests for grant records. A sector-only request, government grants data at large volume in JSONL to train an agent, scores 1 of 16 and is held: add the decision first. A sector-plus-task request, intake review records labeled approve or deny, scores 4 and is held for a missing acceptance test. The workflow-first request built in the packet above scores 13 and passes; its weakest part is the intended use, which the packet does not state. The rubric:
| Criterion | 0 points | 1 point | 2 points |
|---|---|---|---|
| Decision (gate) | A sector or archive only | A task without its trigger | One decision with its trigger and decision point |
| Conditional answers | Approve or deny only | Some middle answers, undefined | Every allowed action defined, including ask, hold, condition, substitute and escalate |
| Decision-time evidence | All records | Systems named | Fields named, each with its capture date |
| Outcome and lag | None | An outcome field named | The record that shows the right answer, its typical lag, and censoring for open cases |
| Hard negatives | None | Mentioned | Look-alike families named with a target share |
| Rights evidence (gate) | Not mentioned | Licensed data asked for | Who created the records, who may license them and how personal data is handled |
| Acceptance test (gate) | Quality asked for in general | A sample review named | A blind relabel by domain reviewers with a pass rule set before delivery |
| Intended use | Not stated | Training or evaluation named | Use bounded, with excluded high-stakes uses written down |
Domain review checklist before records move
Ask a practitioner who has made the decision in that trade to review a sample before any volume is agreed. The decision is explicit: proceed when every item holds on the sample; otherwise ask for the missing evidence and rescore the request:
Checklist · 9 items
- The sample cases are the decision named in the request, not neighboring decisions from the same system.
- Each case freezes the evidence available at the decision date; later documents sit only in the outcome fields.
- Labels use only the request's action set, and the order or rule that assigns them is written down.
- Two reviewers labeled part of the sample blind, and their disagreements were adjudicated and kept.
- Outcome fields name their source record and date, and open cases are marked censored, not clean.
- Hard negatives appear at roughly the share the request set.
- The supplier created the records or holds written permission to license them; client-owned material is out.
- Personal details are removed, or the case is excluded where removal would break it.
- The intended use is written, and high-stakes uses the data cannot support are listed as excluded.
0 of 9 ticked. Tick each item as you go.
What SourceX looks for, in its own published words
SourceX describes its fit test in a few phrases, and a workflow-first request maps onto each. Its screening page looks for "original, well-documented proprietary data", "a decision maker who can authorize a licensing discussion" and "the rights to license that material", and it adds that the material "has to match a real program with AI developers worldwide" and that "SourceX checks this during review." Its process page says SourceX "checks fit, completes the inventory with you, runs the rights check and sets the processing route", and that while it seeks commitment on approved terms, "positive interest is not yet a purchase."
On the buyer side, the request page says datasets "are sourced on request, not held in stock". Selecting workflow datasets on its form opens a workflow specification with fields for the business task, workflow type, input and context, desired action trajectory, required outcome labels and evaluation criteria, beside fields for source systems, licensing requirements and de-identification requirements. The packet's decision goes in the business task, its answers in the outcome labels, its acceptance test in the evaluation criteria, and its rights evidence in the licensing requirements.
What a workflow-first request still cannot promise
A precise request improves the match; it promises nothing about supply, acceptance or results. SourceX sources on request, so a packet describes candidate records, not inventory, and whether a supplier exists for a given decision is learned during sourcing.
A narrow decision also bounds the claims a dataset can support. Records from one trade's version of a decision show how that trade made it under its rules; they do not establish that a model will transfer to another trade, another jurisdiction or a higher-stakes use. State the population the records cover, test on held-out cases from it, and treat anything beyond it as a new question needing new evidence. Rights, privacy and contract questions belong with counsel and the relevant domain experts before any prescriptive use.
Methods note: the audit cross-tab, the rubric and the sources
Audit cross-tab: the script behind the grant administration article in this series (bin/xs0824_fac_eligibility.py) read the audit year 2024 general, federal awards and findings files from a third-party GitHub copy of Federal Audit Clearinghouse data, cwert11/FAC_Data_FY21-FY26 at commit d037e6c, refreshed through 22 October 2026. A finding is a distinct report and reference number pair; its requirement letters are the union of its rows' type_requirement characters, from the 13 letters the clearinghouse's schema allows. A finding counts once under each agency prefix among its award rows, which gives finding-agency pairs; agency names come from the clearinghouse code repository's agency list. For each agency the script counts the pairs carrying each letter, and the table's shares are of that agency's pairs; for each letter it also counts the agencies present and the largest agency's share of that letter's pairs. Shares are rounded half up once. The figure's shading bands come from bin/xs0825_heatmap_values.py and bin/xs0825_make_svg.py, which read the script's aggregate output. Area names follow OMB's Compliance Supplement. The official bulk download was not reachable here, so a rerun on it is a release check.
Rubric: bin/xs0825_rubric_v2.py scores the three invented requests and applies the send gate; the criteria and points are this article's design, not a SourceX standard. The cross-trade table summarizes this series' own case pages. Not measured: how often any trade's records support these decisions, reviewer agreement on any label set, and any SourceX supply, demand or acceptance figure.
Frequently asked questions
What are SourceX specialist operational datasets?
They are datasets of real operational decisions in a specific trade, sourced from businesses that hold the records and may license them, described by the decision they capture rather than by industry. SourceX sources them on request; they are not held in stock. A request names the decision, its allowed answers, the evidence, the outcome source and an acceptance test.
Why not request data by industry?
An industry label does not establish who may license records, whether they suit an intended use, or whether a model will generalize. In public 2024 single-audit data, each of the five agencies with the most findings spanned all 13 compliance areas, so a sector label described a mix of workflows rather than one.
What goes in the workflow specification on SourceX's request form?
Selecting workflow datasets opens fields for the business task, workflow type, input and context, desired action trajectory, required outcome labels and evaluation criteria. Put the decision in the business task, the full action set in the outcome labels and the acceptance test in the evaluation criteria.
How do I know a request is ready to send?
Score it on the eight-criterion rubric: send at 12 of 16 or more, with at least one point each for the decision, the rights evidence and the acceptance test. In the article's illustrative scoring, a sector-only request scored 1 and a workflow-first request scored 13.
Does a precise request secure a dataset or a better model?
No. It improves the match and speeds review, but whether a supplier exists is learned during sourcing, SourceX checks buyer fit during review, and a narrow dataset supports claims only about the population it covers.
Sources
- Federal Audit Clearinghouse findings and awards for audit year 2024, in a third-party copy refreshed through 22 October 2026, were cross-tabulated by agency and compliance requirement for this brief. Source
- The clearinghouse schema allows compliance requirement codes built from 13 letters, from A to P. Source
- Assistance listing prefixes map to federal agencies, such as 84 for Education and 21 for Treasury. Source
- OMB's Compliance Supplement letters its requirements, including A activities allowed or unallowed, B allowable costs and C cash management. Source
- The Supplement's guidance describes eligibility, procurement and suspension and debarment, reporting, subrecipient monitoring and special tests and provisions requirements. Source
- NIST's AI RMF asks that intended purposes, context-specific laws, norms and expectations, and deployment settings be understood and documented. Source
- SourceX's screening page lists its fit criteria and says client-owned material is usually not a fit and that buyer-program fit is checked during review. Source
- SourceX checks fit, completes the inventory, runs the rights check and sets the processing route, and positive interest is not yet a purchase. Source
- SourceX's request page says datasets are sourced on request, not held in stock, and its form collects workflow specifications. Source
Tell us what data you need
Share scope, volume, format, timing and licensing needs. Datasets are sourced on request.