# ALPR Error Layers and Outcome Measurement “ALPR accuracy” is not one auditable number. A program can be accurate at one layer and fail at another: image capture, OCR, jurisdiction/state recognition, hot-list freshness, alert de-duplication, physical-plate verification, occupant correlation, or outcome attribution. A credible rate must identify the layer, denominator, vendor/model/build, threshold, environment, time period, and human workflow. ## The Sherwood failure: OCR and verification Sherwood's February 11, 2026 record is direct Tier-1 evidence of a one-character Flock OCR mismatch. The OCR-generated result returned stolen when queried, officers conducted an armed stop and handcuffed both adult occupants, and the physical plate was inspected only afterward; a small child was also in the vehicle ([[Sherwood Incident and CAD Records]], incident report p. 4; [[Sherwood Body-Camera and Dispatch Media]], approximately 17:28–17:35). The event is not proof of a fleet-wide Flock error rate. It is proof of one consequential misread and a workflow that did not intercept it before the stop. Sherwood's successor policy, effective March 24, 2026, now separates physical-plate verification from database-status confirmation and says an unverified hit cannot be the sole basis for a stop ([[Sherwood ALPR Policy 11.08.00]], p. 3). Because it postdates the event, it is a prospective control, not evidence of the rule in force on February 11. ## The LAPD audit: program governance, not Flock OCR precision For August 1 through September 30, 2025, the LAPD Office of Inspector General reported **210,568,103 reads, 50,183 alerts, and 5,911 unique plates** across a mixed ALPR environment. It reported **337 stolen vehicles recovered, 24 stolen plates, 74 arrests from 68 stops, and nine pursuits**. It also found incomplete or inconsistent outcome records and 161 alerts where the plate match was accurate but the vehicle was no longer stolen (primary public record, [LAPD OIG ALPR audit](../../web%20archive/2026-07-20/lapdpolicecom.lacity.org/lapd-oig-alpr-audit-2026-07.md)). The 161 are chiefly a **hot-list/status-maintenance** problem, not camera OCR errors. Dividing them by recoveries or alerts and calling the result a “Flock false-positive rate” would mix vendors, denominators, error layers, and outcomes. The audit instead demonstrates why incomplete agreements, uncertain vendor access, conflicting retention terms, duplicate events, and missing outcome tracking make program performance difficult to validate. ## A defensible measurement frame | Layer | Required denominator and evidence | |---|---| | Capture | Eligible passes versus usable images, stratified by site, light, weather, speed, and angle | | OCR | Human-verified plate characters versus system result, including hidden/low-confidence candidates | | Hot list | Current and stale records, update latency, list source, and duplicate alerts | | Verification | Alerts received versus plate/state visually checked before enforcement and status independently rechecked | | Association | Correct vehicle versus correct occupant/person/investigation association | | Outcome | Stops, releases, recoveries, arrests, prosecutions, complaints, pursuits, and force, with duplicate attribution removed | Flock's public LPR policy says low-confidence reads may be hidden and that users should confirm computer translation; those are vendor representations of controls, not proof of Arkansas tenant settings or officer compliance (vendor primary/self-description, [Flock LPR policy](../../web%20archive/2026-07-20/flocksafety.com/lpr-policy.md)). ## Caveats - The Sherwood incident supports a specific error-and-response sequence, not a prevalence estimate. - LAPD's mixed program cannot yield a Flock-only false-positive, success, or arrest-causation rate. - Vendor case outcomes and marketing success rates are not controlled performance studies. - No current independent per-SKU Flock confusion matrix was located for OCR, vehicle attributes, FreeForm retrieval, people search, convoy/association, Raven event classes, or Alpha plate reading. ## Open questions - Do Arkansas agencies retain raw images, confidence values, rejected candidates, corrected reads, hot-list update records, and alert-to-outcome links needed to calculate each layer? - Are complaints, stops, handcuffings, pursuits, and releases linked back to the originating alert without subject PII entering public summaries? - What firmware/model/threshold was involved in the Sherwood read, and was it reviewed by the vendor as a defect or training event?