Inference Practice Test

Public agencies increasingly rely on algorithmic tools to allocate resources, enforce rules, and prioritize inspections. Two competing transparency models dominate the debate. The first is maximal disclosure: publish the source code, training data documentation, and decision logic so that any member of the public can scrutinize the system. The second is constrained auditability: allow qualified auditors access to the system and its inputs under nondisclosure agreements or statutory confidentiality, then publish detailed assessment reports, error analyses, and mitigation steps, but not the full technical artifacts. Proponents of disclosure argue that democratic legitimacy demands that the public see how the state is being automated. Yet practitioners note important trade-offs. In a fare-evasion detection system used by one transit authority, full publication of model details led to rapid adaptation by fare evaders, who learned which stations and times were deprioritized; enforcement efficacy fell until the model was retrained and thresholds were altered. By contrast, an air-quality inspection triage tool subjected to third-party auditing, with synthetic test cases and input perturbation studies performed under confidentiality, surfaced biases against small manufacturers without revealing enough about the model to enable easy gaming. In both instances, courts permitted protection of certain proprietary features, provided that impact assessments were thorough and made public. Small municipalities face capacity constraints: they lack in-house data scientists and cannot sustain repeated model retraining if strategies are widely exposed. Hybrid regimes have emerged in response: agencies publish high-level model cards, decision policies, and error bounds, while enabling sandbox audits in which civil-society groups probe behavior using curated datasets. This approach aims to balance two values—explainability and enforceability—without committing to either complete secrecy or complete openness. Importantly, nothing in the hybrid model precludes eventual disclosure where risks of gaming are low; rather, the model is tailored to the risk profile of the domain, the stakes of errors, and the adaptability of regulated parties.

Cities confronting worsening traffic have experimented with two high-profile interventions: capping the number of ride-hail vehicles permitted to operate and charging drivers to enter congested zones during peak hours. Both aim to reduce gridlock and emissions, but they differ in mechanism and in the ways their burdens and benefits are distributed. In City B, a cap on ride-hail licenses led to fewer vehicles cruising the core at midday and a brief dip in average speeds. Yet the cap also increased wait times and fares in outlying neighborhoods where transit options are sparse, especially late at night, as platforms concentrated scarce vehicles where demand was most profitable. Some drivers left the market altogether, citing reduced flexibility.

Congestion pricing in City A produced a more consistent reduction in travel times for buses and delivery trucks, and bus schedule adherence improved within the charged zone. Because the toll applies to most vehicles, critics argued that it imposed new costs on low-income commuters who must drive. Policymakers responded with exemptions and discounts for certain categories of drivers and earmarked a share of toll revenue for transit upgrades in bus-dependent districts outside the zone, including more frequent service and bus lanes. These features did not eliminate all distributional concerns, but they gave City A tools to adjust who paid and who benefited as data accumulated.

Evidence from both cities complicates easy narratives about causation. Ride-hail vehicles contributed to peak congestion but were not its sole driver; construction zones and delivery growth mattered as well. Caps did not directly generate revenue to improve alternatives, and, once issued, licenses became administratively and politically difficult to reallocate to address newly observed inequities. Congestion pricing required significant upfront investment in tolling infrastructure and analytics, and it placed ongoing demands on an agency capable of calibrating rates, processing exemptions, and steering revenues.

Advocates for caps emphasize their relative simplicity and immediate visibility. Proponents of pricing stress its flexibility and the linkage it creates between road use and funding for alternatives. Neither approach automatically achieves equity: each requires design choices informed by data and constrained by administrative capacity. Nonetheless, the pilots suggest that some policy tools are more amenable than others to iterative adjustment in response to distributional impacts.