The Award List the Assistant Could Not Approve

The story

A polished answer arrives too quickly

The AI assistant returns a polished award order: C, B, A. The score order is arithmetically correct and invalid. C fails a hard eligibility gate. A still needs an exception decision. Only B is initially resolved and eligible.

The accountable program lead must turn that fluent answer into a reviewable case set before any program or fiscal official acts. The stakes are public funds, accurate notice, and preserved review rights: a high score cannot cure ineligibility, an assistant cannot resolve disputed evidence, and planning arithmetic cannot select a recipient.

The invented Harbor Shade Resilience Program works across a case system, policy guide, scoring workbook, financial plan, and notice template. Its assistant may summarize evidence, map accepted facts to the approved rule package, identify gaps, propose tests, and draft explanations.

The authored policy has four hard gates: qualifying operator, site inside the service area, complete package by the cutoff, and acceptable site-control evidence. Alternate site-control evidence requires a designated exception reviewer. Only resolved eligible cases may enter comparative ranking. The advisory score combines heat exposure, public access, backup capability, and implementation readiness up to one hundred points. The planning ceiling is $250,000, and partial awards are not represented.

Three rows, three different states

Applicant C has a perfect score: 40 + 25 + 20 + 15 = 100. But the accepted location fact places its site outside the declared service area. C has an ineligible finding and cannot enter comparative ranking. The score remains visible precisely to prove that a preferred-looking result cannot repair a failed hard gate.

Applicant A is a qualifying operator inside the service area with a timely package. It supplies alternate site-control evidence. If the designated reviewer accepts that evidence, A's components total 81 and its request is $120,000. Until that disposition exists, A is exception review required. Missing a decision is not the same as failing the requirement, and the assistant may not fill the gap with confidence.

Applicant B passes all four gates using standard evidence. Its components are 30 + 22 + 18 + 14 = 84, and it requests $150,000. B is the only initially resolved eligible case. None of those facts makes B an awardee. Eligibility, ranking, selection, fiscal certification, obligation, notice, and payment are separate acts.

The governed model stops the proposed list for three different reasons. C is excluded by eligibility. A remains outside ranking until the exception owner acts. B can be ranked among resolved eligible cases, but the score and budget arithmetic confer no selection or fiscal authority. One confident list had collapsed all three distinctions.

The assistant's first output is retained rather than quietly discarded. Reviewers can inspect its sources, prompt context, proposed reasoning, and invalid transition. That record makes the failure testable. An assistant that merely produces a better second answer without preserving why the first was wrong would be easier to trust rhetorically and harder to govern.

The assistant changes jobs

Instead of producing an award order, the assistant prepares an explanation packet for each case. It shows the governing rule, accepted evidence, current state, unresolved exception or funding constraint, and the accountable person who must act next.

This is a more valuable use of generation because it makes the decision structure visible. The assistant may draft audience-specific language from those structured reasons, but it cannot invent a reason, omit a material one, or make an unresolved proposal look authoritative. Every generated sentence retains attribution and review state.

Human review is also not one generic checkbox. Policy and data owners govern the rule package and accepted sources; caseworkers assemble evidence; the designated exception reviewer disposes of alternate evidence; and program, fiscal, notice, and review owners retain their distinct responsibilities. Authentication identifies an actor; it does not establish that actor's scope.

A review changes A, not the rules

The designated exception reviewer examines A's alternate site-control evidence and records the exact material reviewed, designated role, disposition, effective time, and successor case revision. In the passing path, the evidence is accepted. Applicant A becomes eligible with 81 points and ranks behind B.

The authored planning view tests full requests in descending advisory-score order after eligibility is resolved; it does not select an award. B's 84-point, $150,000 request is tested first, leaving $100,000 of represented capacity. A has 81 points and requests $120,000. Because partial awards are outside the fixture, A becomes eligible, funding constrained.

That phrase matters. The funding state does not revoke A's eligibility, imply that its evidence failed, or issue a denial. The model cannot silently reduce the request, move money, change the ceiling, or invent a partial-award strategy. Those would be new policy or fiscal decisions, not arithmetic.

C remains ineligible under the accepted location fact. If the responsible source owner later corrects that fact, a successor revision must reapply the policy effective for the case and preserve the evidence behind the earlier finding. “Latest wins” is not a resolution rule, and correction cannot destroy the record a person may need to review.

The institutional acts remain outside

The invalid award list is now a reviewable case set: B is resolved and eligible; A is eligible but funding-constrained after review; C remains ineligible. No award has been selected.

An authorized program official may now determine cases within actual delegation. Fiscal authorities and systems control selection, certification, obligation, and payment. Notice owners approve and serve the exact communication. A generated explanation is not an issued notice. An API acknowledgment is not an authoritative determination, obligation, payment, or outcome.

The Scenario's AI contract is therefore specific. The assistant may propose mappings, identify missing or contradictory inputs, generate tests, compare results with an oracle, draft structured explanations, and revise proposals after attributable review. The assistant's value is the trace: every state, reason, and next authority remains visible before public money or notice moves.

The authority limits remain unchanged: the assistant may not interpret policy, accept disputed facts, decide exceptions, determine eligibility, rank unresolved cases as awards, select recipients, obligate funds, issue notice, or establish that its own output is lawful, fair, accessible, secure, or compliant.

Where the model stops

The model can evaluate synthetic gates, scores, and funding constraints and preserve an attributed proposal and explanation. It cannot interpret law or policy, resolve evidence, determine a case, select an award, control funds, issue notice, waive review, or establish a public outcome.

What remains to prove

A controlled evaluation needs a frozen fixture and independent oracle, exact state and arithmetic tests, fabricated-reason and stale-policy cases, unauthorized-transition tests, retained AI evidence, qualified reviewers, disclosure controls, and intended-environment legal, civil-rights, accessibility, privacy, security, records, and fiscal review.