AI uses for records
Ask, condition or decline: a grant administration evidence dataset
By SourceX Editorial · Reviewed by Noah Loul ·
10 sections
Short answer
A grant administration evidence dataset should teach the action, not the verdict: refer, escalate, decline, request a named document, approve with conditions or confirm eligibility, tested in a fixed order with the program rule supplied as input. Outcomes arrive late. In audit year 2024 single audits, 1,572 of 16,189 findings involved eligibility, and those reports reached the clearinghouse a median 268 days after the fiscal year ended.
License terms
Ask, condition or decline
Key takeaways
- The unit is one eligibility fact on one application on one decision date, labeled with one of six actions in a fixed order.
- The program's rule version is an input to every case; a model that infers eligibility rules from old outcomes learns rules that have already changed.
- Keep two labels: a decision-time label from reviewers who see only the packet, and a dated outcome label that stays censored while open.
- In audit year 2024, eligibility findings were repeats 44.3 percent of the time against 34.1 percent for other findings, and carried questioned costs more often.
- Records a grant administrator holds for a funder are often the funder's; the licensable core is the administrator's own workflow record.
What decision does a grant administration evidence dataset capture?
It captures one call: at the eligibility step of a grant program, two or more pieces of evidence about the same eligibility fact disagree, and an administrator has to choose what happens next. The useful answer is often neither yes nor no. It is a request for one named document, an award with conditions, a question sent to the program, or a referral.
The unit is one eligibility fact on one application on one decision date. An application with three disputed facts yields three cases, because a model that learns to settle a headcount conflict has not learned to settle a tax-status conflict. The people making this call sit at funders, at pass-through entities and at companies that run grant or rebate programs on a funder's behalf.
Conditional answers are written into federal grant rules, not improvised. Under 2 CFR 200.206, a federal agency may make an award to a recipient that does not fully meet its risk standards when specific conditions can mitigate the risk under 200.208, and 200.208 says the agency must tell the recipient what the condition is, why it applies and what removes it. A dataset that records only approvals and denials throws that middle away.
The angle that keeps the dataset honest: restrict it to administrative records the holder may license, and never let a model infer a program's eligibility rules from past outcomes. Rules change between funding rounds and differ between programs. Each case carries the rule version in force on the decision date as an input, so the model learns to reconcile evidence against a supplied rule rather than to guess the rule.
Six actions, one fixed order: the label schema
Label each case with one of six actions, testing them in a fixed order and stopping at the first that fires: refer, escalate, decline, request, condition, eligible. The order matters because two careful reviewers who weigh the same tests in a different sequence will reach different labels on identical facts, and that disagreement becomes noise the model cannot learn from.
Integrity comes first because a possibly altered document makes every later test unreliable. Rule ambiguity comes next because no amount of evidence answers a question the program never decided. Only then does the authoritative source get read. A request is chosen only when one named document could settle the fact before the window closes; when it could not, the file escalates at that point instead. The table gives each label's trigger, the evidence the file must hold and the mislabel reviewers most often make:
| Label | Fires when | Evidence the file must hold | Mislabel to watch for |
|---|---|---|---|
| Refer | Two versions of one applicant document disagree, or an exclusion record matches on the unique entity identifier | Both versions with submission times; the identifier match | Treating a similar-name exclusion hit as a match |
| Escalate | The supplied rule does not settle the fact pattern; or, at the request test, the settling document cannot arrive before the window closes | Rule text and version; the open question; days left against usual reply time | Guessing what the program meant instead of asking it |
| Decline | The authoritative source named by the rule shows the criterion unmet on the decision date | Dated capture of that source; the criterion text | Declining on a stale or secondary document |
| Request | Status is unknown and one named document from an authoritative source can settle it in time | The document named, its source and the deadline set | A request that names no document |
| Condition | The criterion is met, a secondary item is open and the rule allows specific conditions | Condition text, clearance test, due date and who confirms | Conditioning on the very fact that decides eligibility |
| Eligible | The criterion is met and any mismatch is explained | The record that explains it, such as an assumed-name filing | Calling every mismatch a conflict |
Which records does a grant administrator keep, and what does each prove?
Seven record families make up a case file, and each proves less than it appears to. A registry capture proves status on the day it was taken; an applicant attestation proves only what was claimed. The dataset has to keep each family's own date, because the decision-time label may use only what existed on the decision date.
Tax status shows why the date matters. The IRS revokes exemption automatically when an organization fails to file a required return or notice for three consecutive years, and it publishes the names on an auto-revocation list. A determination letter from years earlier can therefore sit in an applicant's file looking authoritative after the status it records has ended. The capture rules below keep that from leaking into labels:
| Record family | What it proves | What it cannot show | Capture rule for the dataset |
|---|---|---|---|
| Application and attestations | What the applicant claimed, and when it signed | Whether the claim was true | Keep the submitted version and its time stamp, never a later edit |
| Registry captures: entity registration, exclusions, tax status | Status on the capture date | Status on any other date | Store the capture date; recapture at decision if the rule requires it |
| Program rule version | Which criteria applied to this round | How the program reads a case its text never mentions | Link each file to the notice, guidelines and answers in force that day |
| Reviewer checklist and notes | Which items were checked and which conflict was seen | Whether the reviewer's call was right | Keep the checklist as it stood at decision; remove reviewer names |
| Requests and replies | What was asked, by when, and what came back | Why an applicant never replied | Keep dates and the named document; record silence as silence |
| Determination record | The action, its date, the authority and any conditions | Whether conditions were later cleared | Store condition text and clearance test as fields, not prose |
| Later outcomes | Clearance, appeal results, monitoring and audit findings | What would have happened to a declined applicant | Keep in a separate outcome table with its own dates |
How is the right answer known? What 1,572 single-audit eligibility findings show
Use two labels, not one. The decision-time label comes from domain reviewers who see only the packet as it stood and adjudicate disagreements; the outcome label is assembled later from clearance records, appeals, monitoring and audits. The second arrives late and for only some files, which public single-audit data make concrete.
We counted the Federal Audit Clearinghouse tables for audit year 2024, the year in which the audited fiscal period ended, as the clearinghouse's own code sets it. Of 46,740 single audits accepted through 22 October 2026, 8,233 (17.6 percent) reported at least one finding and 1,103 (2.4 percent) at least one finding tagged with compliance requirement E, Eligibility. OMB's Compliance Supplement applies that requirement to most programs that provide benefits to individuals or groups or make subawards. Of 16,189 distinct findings, 1,572 (9.7 percent) carried the E tag.
Eligibility findings stick. They were marked as repeats of the prior audit in 44.3 percent of cases against 34.1 percent of other findings, carried questioned costs in 22.6 percent against 16.1 percent, and were a basis for a modified opinion in 25.2 percent against 19.6 percent. Material weakness rates were close, at 37.3 and 36.4 percent. Four agencies' programs held most of the 1,696 eligibility finding-agency pairs: Housing and Urban Development 32.4 percent, Health and Human Services 26.9 percent, Education 15.8 percent and Agriculture 14.6 percent, across 26 agencies in all.
The lag is the label problem. Reports with an eligibility finding were accepted a median 268 days after the fiscal year closed, with quartiles at 199 and 290 days. Federal rules set the filing deadline at the earlier of 30 days after the auditor's report or nine months after the period ends, and only entities above the audit threshold file at all: 750,000 dollars of federal spending in older text, 1,000,000 dollars in the current 2 CFR 200.501. A file decided early in a fiscal year therefore waits well over a year for this outcome, and small grantees never produce it.
Audit findings describe an auditor's tests at one grantee, not individual applications, so they cannot label a file. Use them to choose which programs need outcome tracking and how long to wait before treating an open file as settled, and keep open files as censored rather than counted as clean. The coding is noisy too: 1,226 eligibility findings carried E alone, while 25 were tagged with all 13 requirement letters. The comparison:
| Measure, audit year 2024 | Eligibility findings (1,572) | Other findings (14,617) | What it means for outcome labels |
|---|---|---|---|
| Repeat of the prior audit | 697 (44.3%) | 4,989 (34.1%) | Eligibility problems persist; one corrected file does not mean a fixed process |
| Questioned costs | 355 (22.6%) | 2,358 (16.1%) | Money is at stake more often, so false eligibles cost more than false requests |
| Basis for a modified opinion | 396 (25.2%) | 2,861 (19.6%) | Auditors escalate these more often than other findings |
| Material weakness | 586 (37.3%) | 5,315 (36.4%) | Control severity looks similar; the difference is persistence and cost |
| Reports accepted, days after year end | Median 268 (quartiles 199 to 290) | Not computed | Wait at least this long before calling an untouched file settled |
Nine illustrative files, labeled by script
The casebook applies the fixed order to nine invented files; a script assigned every label, so no row breaks the article's own rule. Each file states the conflict a reviewer would see. The program rule is an input in every row: where the rule allows conditions or names an authoritative source, the case says so rather than assuming it.
Three rows are hard negatives on purpose. Files A and F look like conflicts and are not, and file I holds a monitoring report dated after the decision that must be removed before labeling and kept only for the outcome table. Files D and H share a conflict and differ only in time left, which is the cleanest test of whether a model reads the window:
| File (illustrative) | Evidence that disagrees | Label | Why not the next label | Outcome to collect later |
|---|---|---|---|---|
| A | Application name differs from the legal name; the state registry lists it as an assumed name | Eligible | The mismatch is explained by a registry record, so nothing is open | Award performance; no eligibility follow-up |
| B | Paper determination letter says exempt; the IRS list shows automatic revocation and no reinstatement | Decline | The authoritative source shows the criterion unmet; a request would ask for nothing new | Reconsideration or reapplication after reinstatement |
| C | Entity registration lapsed after submission; renewal receipt attached; every other criterion met | Condition | Eligibility rests on the authoritative source; only the registration is open | Whether the condition cleared by its due date |
| D | Headcount on the form is under the program cap; last year's payroll summary is over it | Request | Status is unknown, and one named payroll record can arrive in time | Reply received, and the eligibility it showed |
| E | Unincorporated group applies through a fiscal sponsor; the notice is silent on sponsors | Escalate | No evidence answers a question the rule never settled | The program's ruling, recorded as a rule note |
| F | Exclusion search returns a similar name; identifier and address differ | Eligible | No identifier match, so there is no integrity signal | None beyond normal monitoring |
| G | Two board resolutions authorize the application with different dates and signers | Refer | Possible alteration outranks every later test | Integrity review result |
| H | The same headcount conflict as D, with two days left and replies usually taking ten | Escalate | The settling document cannot arrive before the window closes | Program decision on the late evidence |
| I | Packet includes a later monitoring report; decision-time evidence meets the criterion | Eligible | The later report is removed before labeling | The monitoring result, stored as outcome only |
Hard negatives: seven traps that test whether a model reads evidence or patterns
Hard negatives are cases that share surface features with the wrong label. Without them, a model scores well by learning that any mismatch means request and any registry warning means decline. Include each family below at a known rate in both training and evaluation splits:
At a glance
Hard negatives: seven traps that test whether a model reads evidence or patterns
- Assumed names: the applicant's trading name differs from its legal name, and a registry record explains it. Correct label: eligible.
- Mailing and service addresses: a post office box sits outside the service area while the delivery site sits inside it. The rule decides which address counts; the case must carry that rule.
- Stale authority: an old determination letter or certificate looks decisive but has been overtaken, as automatic revocation shows. Correct label depends on the dated authoritative capture, not the paper.
- Near-name exclusion hits: a similar entity appears on an exclusion list with a different identifier. Correct label: not an integrity signal.
- Leaked outcomes: a monitoring report, appeal or audit dated after the decision sits in the packet. Remove it from the decision-time view and keep it as outcome.
- Rule-permitted self-attestation: where a program accepts an applicant's signed statement, a missing document is not a gap. The rule input, not the model, says so.
- Same conflict, different clock: identical evidence with a closed or open window should flip between request and escalate.
What can a licensor's grant records show, and what can they not?
A licensor's records can show how its reviewers handled conflicts: which evidence they saw, what they asked for, how long replies took, which conditions they set and which cleared. They cannot show the true eligibility of anyone, the funder's final reading of an ambiguous rule unless the funder recorded it, or any outcome that arrived after the administrator's contract ended.
Ownership is the first filter. A company that administers grants for a funder often holds records on the funder's behalf, and SourceX's screening page lists material that belongs to the company's clients rather than the company among what is usually not a fit. The licensable core is narrower: the administrator's own checklists, request templates, timing and decision codes, released only where its contracts with funders allow that use.
Applicants are the second filter. Organizational applicants still name officers and signers, and programs that serve individuals hold personal and financial details. Those fields come out before anything moves, and some programs' files should not move at all. Rights, privacy and contract questions need counsel's review of the specific agreements; this page is general information.
A sector label settles none of this. Calling the records government grants data says nothing about who may license them, whether outcomes were ever captured, or whether a model trained on one program's files will transfer to another program with different rules.
Domain review checklist before a file enters the dataset
A domain reviewer, someone who has made eligibility determinations under the program type, signs off each batch against the checks below. The decision for a buyer is explicit: proceed only when every check passes on a sample; otherwise ask for the missing evidence before any volume is discussed:
Checklist · 9 items
- The file is administrative and the holder created it or may license it under its funder contracts; client-owned material is excluded.
- Each case names one eligibility fact, one decision date and the rule version in force that day.
- Every registry capture carries its own date, and nothing dated after the decision is in the decision-time view.
- The label is one of the six actions, assigned in the fixed order, with the order step recorded.
- Two reviewers labeled a blind sample independently, and disagreements were adjudicated and logged.
- Requests name a document and a deadline; conditions carry a clearance test and a due date.
- Outcome fields are separate, dated and marked censored where the outcome window is still open.
- Officer names, signatures, personal and bank details are removed or the file is excluded.
- Limits are written down: which programs, rounds and applicant types are covered, and that the dataset supports triage assistance, not automated denial.
0 of 9 ticked. Tick each item as you go.
Requesting a grant eligibility dataset through SourceX
Write the request around the decision, not the sector. On SourceX's request form, selecting workflow datasets opens a workflow specification: put the eligibility conflict in the business task, the packet fields in the input and context, the six actions in the required outcome labels, and the two-label test in the evaluation criteria. Licensing requirements and de-identification requirements have their own fields.
SourceX's request page says its datasets are sourced on request, not held in stock, so a request describes candidate records rather than reserving existing ones, and nothing here implies that a matching supplier exists or that a model will improve. The most useful acceptance test to attach is concrete: a blind sample relabeled by two domain reviewers against the schema above, with the pass rule written before delivery.
Methods note: the clearinghouse count, the casebook and what was not measured
Clearinghouse count: a script (bin/xs0824_fac_eligibility.py) read three files from a third-party GitHub copy of Federal Audit Clearinghouse API data, cwert11/FAC_Data_FY21-FY26 at commit d037e6c, refreshed with acceptances through 22 October 2026: general, federal awards and findings for audit year 2024. Rows with any other audit year would be dropped (none were); reports marked deprecated by resubmission would be dropped (none were). A finding is a distinct report and reference number pair, since the file repeats a finding once per award. Its requirement letters are the union of the type_requirement characters across those rows, counting only the 13 upper-case letters the clearinghouse allows (no other characters occurred, and every finding had at least one). A flag counts as yes when any row says yes; between 26 and 104 findings had rows that disagreed, depending on the flag. Agencies come from each award row's two-digit prefix, named from the FAC code repository's agency list; a finding counts once under each of its agencies, and 81 eligibility findings sat under more than one. Lag is calendar days from the general table's fiscal year end date to its acceptance date for each report with an eligibility finding, both parsed strictly as dates (none failed), with median and quartiles by nearest rank. Percentages are rounded half up once.
Flag meanings follow the clearinghouse's own submission instructions, for example that questioned costs is marked yes even when the amount cannot be determined and that a repeat finding is one carried from the immediate prior audit. The official bulk files at fac.gov were not reachable from our environment, so the count rests on the copy; a rerun on the official download is listed as a release check. The saved outputs hold aggregates only, with no entity names, identifiers or report numbers.
Casebook: bin/xs0824_casebook.py labels the nine invented files with the fixed order shown above, checking integrity, then whether the rule settles the case, then a decline on the authoritative source, then a request (or escalation when the document cannot arrive in time), then, for an open item, a condition where the rule allows one (otherwise a request or escalation by the same timing test), then eligibility; all six labels occur. Not measured: how often eligibility evidence conflicts in any real program, reviewer agreement on this schema, any model's performance, and any SourceX supply or demand. The audit figures describe findings at audited grantees, not decisions on applications.
Frequently asked questions
What is a grant administration evidence dataset?
It is a set of eligibility cases from grant programs in which evidence about one fact disagrees, each labeled with the action a qualified reviewer would take and, separately, with what happened later. The label set includes conditional answers such as requesting a named document or approving with conditions. The program rule in force is supplied with each case rather than learned.
Can single audit findings label individual grant decisions?
No. A finding records an auditor's test of a grantee's compliance, often on a sample, and it arrives months after the fiscal year ends; in audit year 2024 the median gap to acceptance for reports with eligibility findings was 268 days. Use findings to decide where outcome tracking matters and how long to wait, not as file-level labels.
Should an AI model learn a program's eligibility rules from past decisions?
It should not. Eligibility rules differ by program and change between rounds, so rules inferred from old outcomes go stale and can encode past errors. Give the model the rule text in force on the decision date and grade whether it reconciles evidence against that rule.
What does approving a grant with conditions mean?
In federal awards, 2 CFR 200.208 lets an agency add specific conditions, such as reimbursement-only payment or extra reporting, when risk calls for it, and it requires telling the recipient why and what removes the condition. In a dataset, the condition text, its clearance test and its due date should be separate fields so the outcome can be checked.
Who can license grant administration records for AI?
Only a holder with the right to do so, which often excludes records kept for a funder under an administration contract. Companies that run programs may hold their own workflow records, such as checklists, request templates and decision codes, that their contracts allow them to license after personal details are removed. Counsel should review the specific agreements.
Sources
- Federal Audit Clearinghouse tables for audit year 2024 (general, federal awards, findings), in a third-party copy refreshed through 22 October 2026, counted for this article. Source
- The clearinghouse's submission instructions define the questioned-cost, repeat-finding, material-weakness and modified-opinion flags. Source
- The clearinghouse sets a report's audit year from the year in which the auditee's fiscal period ends. Source
- Two-digit assistance listing prefixes map to federal agencies, such as 14 for Housing and Urban Development and 93 for Health and Human Services. Source
- The clearinghouse publishes full CSV exports of its general, federal awards and findings tables by audit year. Source
- OMB's Compliance Supplement applies the eligibility requirement to most programs that provide benefits to individuals or groups or make subawards. Source
- A federal agency may award to a recipient that does not fully meet risk standards when specific conditions under 200.208 can mitigate the risk. Source
- Agencies may impose specific conditions and must tell the recipient the reason and what is needed to remove them. Source
- Single audit packages are due the earlier of 30 days after the auditor's report or nine months after the audit period ends. Source
- The current single audit threshold is 1,000,000 dollars of federal expenditure in a fiscal year; earlier text set 750,000 dollars. Source
- An exempt organization that fails to file a required return or notice for three consecutive years loses its exemption automatically, and the IRS publishes an auto-revocation list. Source
- SourceX's screening guidance lists material that belongs to a company's clients rather than the company as usually not a fit. Source
Tell us what data you need
Share scope, volume, format, timing and licensing needs. Datasets are sourced on request.