Fines in the Code: Computationally Mapping Local Monetary Sanctions

Local monetary fines are among the most direct points of contact between government and residents, yet their structure remains largely unstudied at scale due to the inaccessibility of municipal code text. This paper makes two contributions. First, we develop and validate an LLM-based extraction pipeline for identifying penalty architecture variables, including escalation schedules, per-day violation language, and maximum fine ceilings, from municipal and county codes drawn from LOCUS v1.0, a corpus of 2.2 million ordinance chunks spanning all 50 states. Validating against 30 hand-coded jurisdictions, we find that binary structural variables extract reliably (67--90\% agreement) while specific dollar amounts are harder to extract consistently (60--67\%), with performance improving through prompt refinement targeting general penalty sections. Second, applying the pipeline to 853 jurisdictions across 11 states and merging with Census income data, we find that nominal fine amounts are strikingly uniform across income tiers, as the median first violation is \$100 regardless of jurisdiction wealth, yet fine burden as a share of monthly income is 2.2 times higher in low-income jurisdictions than high-income ones, consistent with fines functioning as a regressive tax.