The Model Card That Says 'High Accuracy'
A public-sector agency procures a document-classification system to route benefit applications. The vendor's model card reports 96% accuracy on a benchmark of 50,000 documents and lists intended use as 'administrative document triage.' The agency's own pilot on 3,000 real applications finds 91% accuracy overall, but 74% on handwritten forms, which are common among applicants over 65 and applicants in rural areas. The contract has a 30-day acceptance window, now at day 22. The vendor says the benchmark did not include handwriting because it is 'out of scope.' The agency's legal team notes that misrouting an application can delay benefits by weeks. The procurement officer wants to accept, citing the 91% figure as 'above the 90% threshold in the contract.' The contract's threshold clause does not define the population over which accuracy is measured.
1. Advise the procurement officer on whether the 91% pilot result satisfies the contract threshold, and what the ambiguity means.
Read the worked answer
The 91% figure does not clearly satisfy the threshold, because the threshold clause does not say what population it applies to, and the answer changes with the population.
On the full pilot, 91% clears a 90% bar. On handwritten forms, 74% does not. The contract is silent on subgroup performance, so the officer is reading a global number into a clause that never specified a global measure. That is a defensible but risky interpretation: if a dispute arises, the agency will have accepted a system that fails on a defined applicant subgroup, and the ambiguity in its own contract will not protect it.
The 'out of scope' defense is weak in context. The vendor defined scope in the model card, but the agency's actual applicant population includes handwritten forms. A scope exclusion that excludes a meaningful share of real users is not a neutral limitation; it is a gap between the intended use and the deployed use. The agency, not the vendor, owns the consequence of that gap.
The officer should not accept on day 22. The right move is to seek a written clarification or amendment defining the measurement population, require subgroup reporting, and treat the handwriting gap as a known limitation that must be mitigated before acceptance.
2. The 30-day window closes in 8 days. Outline the options and recommend one, with reasoning.
Read the worked answer
Three options.
Option A: accept now on the 91% reading. Fast, but locks in a system with a 74% subgroup accuracy, exposes the agency to appeals and reputational harm, and forfeits leverage. Not recommended.
Option B: reject and restart procurement. Clean, but slow, costly, and likely unnecessary, since the core system performs acceptably on printed forms.
Option C: conditional acceptance with a remediation plan. Accept the system for printed-form routing only, require a documented human-review step for handwritten forms, and make the vendor's continued acceptance contingent on meeting a stated subgroup target within a defined period. Recommended.
Reasoning: the agency's real risk is misrouting vulnerable applicants, not the headline number. Option C addresses the risk directly by removing handwritten forms from automated routing until performance improves, which is a mitigation the agency can implement itself without vendor cooperation. It also converts the ambiguous threshold into a defined one going forward. The 8-day constraint favors C because it does not require renegotiating the whole contract, only an amendment to the acceptance criteria and a routing rule. Document the decision, the population definitions, and the mitigation, so the agency can show it identified and managed the gap rather than ignoring it.