Legal Governance of Classification Thresholds
Two people with similar risk scores can face different outcomes because a cutoff falls between them. This paper examines the decision to set that boundary in federal prison classification, New Jersey pretrial decisions and English facial recognition. The records studied identify boundaries without comparing alternatives or explaining how the choices distribute mistakes. The analysis shows where existing law governs that choice and where gaps remain. It proposes a Threshold Decision Record to document who is responsible, the alternatives considered, expected consequences and when the rule should be reviewed.
Abstract
That the thresholds converting algorithmic scores into governmental action are policy choices is settled in scholarship. The act fixing those thresholds remains at most partially governed. This article isolates the operating point, the decision rule converting scores, ranks, similarities and constrained selections into official action, as a distinct unit of public-law analysis. Documentary records of PATTERN, New Jersey’s pretrial framework and English live facial recognition show operating points adopted and revised with, at best, authorship, publication and bounded reasons. Yet no located adoption record states the error allocation, compares alternatives or supports meaningful threshold-specific review. This article specifies when existing doctrine already reaches the threshold-setting act and where it runs out. It develops the Threshold Decision Record, a canonical record of adoption and material revision connecting the rule in its native technical form to an identified owner, comparative error-allocation reasons, upstream dependencies, an adopting legal vehicle and the review route.
How the inquiry is constructed
A documentary and doctrinal study examines public records, judgments, legislation and published guidance from the United States and England and Wales. A defined search method and coded treatment of absence claims support the documentary analysis. Decision theory is used narrowly to explain cost-sensitive thresholds and the conditions under which a threshold can imply a benchmark error-cost ratio. The institutional proposal is an eight-element Threshold Decision Record.
What the analysis establishes
- The operating point is a distinct public act because it converts an epistemic score into an official consequence.
- Published systems can identify an owner and a boundary while still lacking comparative reasons about the errors created by that boundary.
- Existing administrative law reaches some threshold-setting acts but does not provide continuous, symmetric or branch-wide governance.
- The Threshold Decision Record should document authority, score semantics, the native decision rule, alternatives, error allocation, human departures and revision conditions.
- A model revision, operating-point revision or changed deployment context can require a new record even when the displayed threshold number does not change.
Related research
SSRN manuscriptsPricing Contractual Vagueness
Contracts cannot anticipate every disruption. Flexible terms can protect a party from extreme losses, yet leave both sides uncertain about how a court will interpret them. This paper models the balance between wider protection and disputed meaning in force majeure and material-adverse-change clauses. It shows that familiar expected-utility reasoning can explain the main results, then examines what a model of attitudes toward uncertain probabilities adds. Within the type of clause studied, some openness can be optimal. Proposals for measurable triggers and bounded renegotiation remain conditional on practical costs the model does not measure.
Linking Decisions, Duties and Relief in Public Benefits Administration
When a benefits agency discovers an error, correcting a database is only part of the response. Programme law determines who should receive relief, how much is due, the relevant date and what must actually happen. Using Rhode Island SNAP and a Medicaid home-care test, with bounded comparisons to Social Security and Michigan MiDAS, this paper connects those duties to the records and officials needed to fulfil them. It distinguishes an apparent discrepancy from an established loss, and a posted benefit from money a recipient can use. It does not claim that historical records prove households received restoration or that present systems perform better.
Components of Automated Decision-Making in Benefit Allocation and Their Legal Status
An automated system helped decide how many hours of home care Arkansas Medicaid recipients could receive. This paper asks which parts of that system counted as legal rules, when the agency adopted them and what was made public. It separates the instruction linking care hours to a group, the logic placing people in groups and the table assigning hours. Using official filings, agency accounts and court records, it shows why those parts need separate legal histories. Some conclusions rely on retrospective evidence; the earlier system's full specifications and its legal effect on particular recipients remain unresolved.
Human Endorsement in Automated Government
Does a human signature make an automated government decision legitimate? This paper examines what the records actually establish about human authorisation in England's Ofqual grading arrangements and Australia's Robodebt scheme. It distinguishes work on calculations, participation in a proposal and endorsement of a decision process. Some evidence supports specific knowledge and actions; other claims about endorsement, deference and reliance remain unresolved. The contribution is a careful account of those evidential limits, rather than a general verdict on all human oversight.