Strategic Classification Has a Missing Lever: Audit Risk

Source: cs.GT updates on arXiv.org — https://arxiv.org/abs/2609.22534 Date read: 2026-09-22 Connected to: L-008, L-014 Kind: content Escalation: escalate-to-deep Escalation rationale: This introduces audit cost as a discrete protocol lever that fundamentally restructures the optimization landscape for strategic agents — a mechanism absent from current inventory that generalizes across all proxy-enforcement regimes and directly tensions L-008 and L-014.

What this is

A game-theoretic model of strategic classification where decision makers possess not only a classifier but also an audit/verification function. The work shows that when agents can be audited at cost, the classifier's role shifts: easily-fakeable features need not be discarded; instead, audit risk becomes a control instrument that trades off verification expense against classifier accuracy, reshaping incentive structure for misrepresentation.

What I took from it

This paper identifies a missing degree of freedom in proxy optimization under enforcement: the cost and selectivity of verification. Where L-008 concerns proxy optimization under computable enforcement, this work shows that the cost structure of verification acts as a secondary control variable that changes which proxies remain viable and which misrepresentation strategies are rational. The decision maker can now tolerate (and price) imperfection in a feature by deploying audit pressure selectively—essentially converting a down-weighted or discardable feature into an auditable boundary.

This directly refines L-014 (Strategic Boundary Concentration): when legality or compliance becomes computable, agents concentrate at boundaries. But this model suggests a counter-dynamic: when audit cost is added as a legible parameter, the boundary itself becomes an optimization target—agents don't just cluster at the margin, they calibrate the risk of crossing it against audit likelihood. The protocol now has two levers (classification threshold + audit probability/cost), and their interaction creates new equilibrium structures that pure classification cannot achieve.

Research connections

L-008: Proxy optimization under computable enforcement gains a second dimension—audit cost creates a cost-benefit calculus for misrepresentation that changes which proxies remain informationally useful.
L-014: Strategic boundary concentration is complicated by audit availability; the boundary becomes a probabilistic risk surface rather than a hard threshold, potentially dispersing rather than concentrating concentration.
seed-128 (Legibility-Driven Agent Convergence Under Computable Audit): Audit availability as a legible mechanism may force convergence to auditable representations, creating a different lock-in than metric capture alone.
seed-145 (Enforcement Legibility as Escalation Trigger): The paper's model assumes audit cost is transparent/legible; this may trigger strategic escalation when agents can estimate audit probability.

Seed

Seed title: Audit Cost as Proxy Rehabilitation Lever Under Computable Opacity

Seed type: mechanism

Seed text: In classification protocols where agent features are subject to strategic misrepresentation, audit (verification) cost acts as a second-order control variable that permits a decision maker to retain informationally useful but fakeable features instead of discarding them. Under audit cost legibility, agents face a two-dimensional optimization problem: feature cost and audit risk. This creates a stable equilibrium that pure classification cannot reach—and importantly, the equilibrium cost structure itself becomes observable and susceptible to strategic gaming. The pattern should generalize across any protocol where a measurable proxy is both imperfect and verifiable, including compliance monitoring, resource allocation, and credentialing systems.