Explain Eligibility Decisions Without Hiding Exceptions

An explanation begins before the notice is written

Public-program decisions become difficult to explain when the final status is the only durable result. A case system may store facts, a worksheet may calculate eligibility, a desk guide may describe exceptions, a ranking tool may score eligible cases, a financial system may track funds, and a document template may produce a notice. If those implementations drift, a clear notice can confidently explain the wrong rule.

The use case is not “automate eligibility.” It is to preserve a reviewable chain from governing source to accepted facts, effective rule, eligibility gates, exception state, score or priority, resource constraint, draft explanation, authorized determination, issued notice, and any later correction or review. Those are related acts, but they are not interchangeable.

Selected federal sources make the responsibilities concrete, but their scope matters. 5 U.S.C. § 555(e) generally requires notice of the denial of a written request in an agency proceeding and, subject to stated exceptions, a brief statement of grounds. It does not supply a program's eligibility rule.

For federal Privacy Act systems of records, 5 U.S.C. § 552a addresses access, amendment, purpose limitation, and records used in determinations. Applicability depends on the actual system and program.

44 U.S.C. § 3101 addresses adequate documentation of federal policies, decisions, procedures, and essential transactions. The Plain Writing Act and Revised Section 508 Standards establish requirements for covered communications and technology. Documentation, clarity, or digital delivery alone does not make a determination correct.

The current workflow hides unlike decisions inside one status

The risky shortcut is a field called approved, eligible, or denied that has lost the reasons and authority behind it. Several distinct questions may have been collapsed:

1. Are the submitted and retrieved facts sufficiently complete and reliable for the represented use?
2. Do those facts satisfy every hard eligibility gate under the policy version effective for the case?
3. Does an exception path exist, and has the role authorized to decide it acted?
4. If eligible, how is the case scored or prioritized?
5. Do represented funds, capacity, or other program constraints permit selection?
6. Who may issue the official determination, notice, award, payment, license, or benefit?

An external-data match should not answer all six. GAO-26-107466 found data-quality issues in all nine eligibility sources it examined and inconsistencies across seven. GAO cautioned that identified exclusion-and-award matches did not themselves prove an improper award or fraud; each case required evaluation. A match is evidence, not self-executing authority.

Model the decision as states, not one score

A governed eligibility model should preserve at least these states:

State	What it means	What it must not imply
Facts unresolved	Required evidence is missing, stale, disputed, or contradictory	Failure, fraud, or ineligibility
Ineligible finding	Accepted facts fail a represented hard gate	Issued denial or exhausted review rights
Exception review required	A represented exception path may apply but lacks its authorized disposition	Automatic waiver or eligibility
Eligible finding	All represented hard gates pass	Selection, entitlement, available funds, or payment
Ranked or prioritized	An authored scoring or priority rule has been applied to eligible cases	Authority to select or obligate funds
Funding constrained	The represented request does not fit the modeled ceiling or allocation rule	Legal unavailability, denial, or authority to change the request
Determined	The authorized program process records an official disposition	Proof that the decision is lawful, fair, or immune from correction or review

This separation makes explanations more accurate. It also prevents a high score from curing failed eligibility, an available budget from curing missing evidence, or a generated recommendation from becoming a public act.

A transparent synthetic grant fixture

The following is a synthetic arithmetic fixture for design and testing. It is not a real program, policy, applicant, legal interpretation, award recommendation, or Grid deployment. Lettered applicants are rows, not fictitious organizations. The criteria and dollar amounts are invented only to make the expected states reproducible.

The fixture has four hard gates: the applicant is a public or nonprofit operator; the proposed site is inside the service area; the package was complete by the cutoff; and acceptable evidence of site control is present. Standard evidence satisfies the fourth gate automatically. Alternate evidence may be accepted only by the designated exception reviewer. Until that person acts, the case is exception review required rather than eligible or ineligible.

The component arithmetic can be checked for every row, but only eligible cases enter the comparative ranking:

score = heat exposure + public access + backup capability + implementation readiness

The maximum components are 40, 25, 20, and 15 points, respectively, for a maximum score of 100. The modeled planning ceiling is $250,000. Partial awards are not represented. For capacity analysis, the fixture considers resolved eligible cases in descending score order. That provisional order is advisory; it does not authorize selection or an award.

Synthetic row	Hard-gate facts	Exception state	Score arithmetic	Request	Represented result before determination
Applicant A	Operator, area, and cutoff pass; alternate site-control evidence supplied	Awaiting authorized review	36 + 20 + 12 + 13 = 81 if eligible	$120,000	Exception review required; not yet ranked
Applicant B	All four gates pass with standard evidence	None	30 + 22 + 18 + 14 = 84	$150,000	Eligible and highest currently resolved score
Applicant C	Site is outside the represented service area	Not applicable	40 + 25 + 20 + 15 = 100, shown only to test separation	$100,000	Ineligible finding; score cannot cure the failed gate

At the initial review, Applicant B is the only resolved eligible case. Its $150,000 request fits within the $250,000 ceiling, leaving 250,000 - 150,000 = 100,000 in represented planning capacity. That does not create an award; the selecting and fiscal authorities still have to act through the actual process.

Suppose the designated exception reviewer later accepts Applicant A's alternate site-control evidence under the synthetic rule. A new case revision records the evidence reviewed, the reviewer, the authority asserted, the disposition, and the effective time. Applicant A becomes eligible with a score of 81. It ranks behind Applicant B, but its $120,000 full request does not fit the remaining $100,000. The model therefore returns eligible, funding constrained. It must not silently reduce the request, move funds, invent a partial award, or convert the funding constraint into an eligibility denial.

Applicant C's explanation should identify the service-area gate and accepted location fact. The 100-point arithmetic is deliberately immaterial. If the location fact is corrected through the applicable process, the system should create a new revision and reevaluate the gate; it should not overwrite the record that supported the earlier finding.

This fixture exposes three errors that a single score hides: an unresolved exception, an eligibility failure despite the highest score, and an eligible case that does not fit the represented funding ceiling.

Build an explanation packet, not an answer-shaped paragraph

A reviewer or notice author needs structured reasons before prose is generated.

Packet element	Question it answers
Governing source and effective version	Which represented rule applied, and why did it apply to this case?
Material accepted facts	Which facts changed the finding, from which source, and as of when?
Missing, stale, or disputed facts	What remains unresolved rather than treated as false?
Gate-by-gate result	Which eligibility predicates passed, failed, or stopped?
Exception record	What path exists, who may decide it, what evidence was reviewed, and what happened?
Score and constraint calculation	How was a score computed, and which resource rule affected selection?
Authority record	Who reviewed, recommended, determined, issued, or corrected—and under what delegation?
Review and correction path	What next step is configured, and which earlier state must remain reconstructable?

Grid can produce separate staff, applicant, and oversight views from that packet, plus report and API outputs. It should not publish the same fields to every audience: internal notes, protected information, legal advice, and security data may follow different rules from public reasons.

A draft explanation is not an issued notice. The program must determine content, service, language access, accessibility, timing, and review instructions, then decide whether the reasons are sufficient for the action.

The Grid operating pattern and authority map

The model begins with a governed policy package: sources, approved interpretations, effective intervals, transition rules, owners, tests, and superseded versions. Case facts retain provenance, time, quality state, disputes, and corrections. Hard gates remain separate from scoring, exceptions, and funding constraints. Each finding binds to its fact and rule revisions.

Roles remain explicit. Policy owners approve rules; data stewards govern mappings; caseworkers verify facts; delegated reviewers decide exceptions; program officials determine cases; and fiscal systems and officials control funds. Notice, records, privacy, accessibility, civil-rights, security, legal, and review responsibilities remain external.

Grid can apply represented rules consistently and retain an explicitly modeled execution record. Call that record a runtime receipt only when it satisfies a documented receipt contract and binds captured runtime provenance. Grid cannot interpret ambiguous law, establish a delegation, decide credibility, approve an exception, obligate funds, issue a legally operative notice, waive appeal rights, or certify compliance. A consistent execution of a defective or unlawful policy remains defective or unlawful.

OMB Circular A-123 and the separately issued 2025 GAO Green Book address federal internal control. Neither endorses Grid; both reinforce the need for control owners, evidence, monitoring, and corrective action.

Integration, privacy, accessibility, and records boundaries

The model should not replace authoritative case, identity, document, financial, payment, or notice systems by accident. Each integration needs a defined source owner, schema, permitted purpose, identity resolution rule, validation, retry behavior, correction path, and acknowledgment. A successful API response proves transport, not truth or authorized action.

Collect and expose only what the purpose requires. When sensitive data is lawfully needed, the organization must define use, access, retention, disclosure, and review. Privacy, records, accessibility, civil-rights, security, and legal evaluations apply to the complete system.

Test accessibility on every intended surface and document. A shared model does not make an interface accessible. Records treatment must cover sources, case revisions, actions, issued artifacts, and corrections under the applicable schedule—not just the final screen.

Failure modes that should fail the evaluation

• The latest policy version is applied to a case governed by an earlier effective version.
• Missing evidence is converted to false, or conflicting facts are resolved by “latest wins” without authority.
• An external match is treated as conclusive proof of ineligibility or fraud.
• Scoring runs before hard gates, allowing a high score to hide a failed requirement.
• An exception is implemented as an undocumented override.
• A reviewer can decide an exception or determination outside the role's delegation.
• A funding constraint is described as an eligibility failure.
• A correction overwrites the evidence and explanation for an earlier action.
• A generated explanation omits the material reason or invents a reason not present in the model.
• A technically correct notice is inaccessible, sent through the wrong process, or exposes protected data.

Acceptance criteria for a no-consequence pilot

Use synthetic or appropriately controlled historical fixtures before any live public consequence. Freeze expected outcomes and stop states with qualified program, legal, fiscal, privacy, records, accessibility, security, and technical reviewers.

Test area	Acceptance criterion
Policy versioning	Every effective-date and transition test applies the preapproved rule version
Facts and provenance	Required sources, observation times, disputes, corrections, and missing states remain distinguishable
Calculation	Independent reviewers reproduce every gate, score, ranking, and funding result
Exceptions	Every seeded exception stops for the designated role and retains its evidence and disposition
Explanation	Qualified reviewers identify the same material reasons and find no unsupported reason in the packet
Authority	Every seeded unauthorized official-state transition is blocked and recorded
Replay	Earlier findings, issued artifacts, corrections, and review links remain reconstructable
Surface controls	Accessibility, privacy, records, security, and disclosure tests pass for each intended channel

Compare rekeys, rule-version conflicts, unresolved cases presented as complete, exception cycle time, correction effort, explanation defects, and reconstruction time. Set thresholds before the run. A pass applies only to the tested rules, data, interfaces, roles, and conditions—not legal sufficiency, fairness, compliance, authority to operate, or program effectiveness.

Continue with Policy That Can Explain Itself for the deeper legal and institutional boundary and One Program Rule, Many Public Surfaces for staff, public, report, and API consistency. Use the AI decision-authority assessment when generated assistance participates, and the decision evaluation worksheet to establish the oracle and evidence plan.

The standard is not a polished eligibility answer. It is a determination lifecycle in which facts, gates, exceptions, scores, constraints, explanations, authority, notices, and corrections remain distinct enough that the program—and the person affected—can understand what actually happened.