# 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. ## Arkansas's own statutory numbers The ASP and Springdale productions released on July 24 supplied Arkansas outcome data at scale. LRPD's later six-month report supplies a much larger scan denominator but maps the statutory categories onto a different vendor-defined outcome construct. All three are agency-compiled § 12-12-1805 figures, not outside estimates. **Arkansas State Police, May 2025 through June 2026** ([[ASP Statewide ALPR Data Reports 2025-2026]]): | Measure | Value | |---|---:| | Plates scanned | 61,798,979 | | Plates run against lists | 37,554 | | Confirmed matches | 54 | | Misidentifications | 919 | | Arrests and prosecutions | 5 | Misidentifications exceed confirmed matches by roughly **seventeen to one**, and one arrest corresponds to approximately **12.4 million scans**. Seven of the fourteen months recorded zero confirmed matches. **Springdale Police Department, January 2025 through June 2026** ([[Springdale Semiannual LPR Reports 2025-2026]]): the Axon fleet system reports 6,804 misidentifications across 9.6 million reads over all three reporting periods, while the Flock stationary system reports **one** across 23.8 million reads over the two periods in which Flock data exists. Compared across the same two periods, Axon reports 5,872 across 8.8 million. **Little Rock Police Department, January through June 2026** ([[LRPD January-June 2026 ALPR Practice and Usage Report]]): LRPD reports **103,375,512 scans** and **350,211 alerts**. Its dashboard then reports **91 "successes"**, defined as an arrest, lead, warrant, or stop associated with a Flock hit, plus 86 cases cleared, 54 stolen vehicles recovered, and $856,299 in recovered property (`6 Month Report Jan-June 2026.pdf`, pp. 1, 3). That is one recorded success per approximately 3,848 alerts, but it is an outcome-attribution ratio, not an error rate. **North Little Rock Police Department, January through June 2026** ([[North Little Rock January-June 2026 ALPR Practice and Usage Report]]): NLRPD reports **49,569,171 scans** and **241,872 alerts** across Flock and SkyCop. Its Flock dashboard defines success with the same arrest/lead/warrant/stop taxonomy but displays **0 successes**, **0 cases cleared**, and no data in the remaining panels (`January 01-June 30 2026.docx`, § II and Figure 2). That zero does not establish zero false alerts, arrests, or prosecutions because the required fields are not separately displayed ([[T040 - NLRPD Required Outcome Categories vs Zero-Data Success Dashboard]]). These figures belong at the **outcome-attribution** layer, and each carries a caveat that this page insists on rather than sets aside. ASP's series includes two reports silently revised the day before production and a month recording confirmed matches against zero query activity ([[T024 - ASP Revised ALPR Reports and the May 2025 Query-Read Anomaly]]). Springdale's near-zero Flock misidentification count is more plausibly an uncaptured metric than a measured one ([[T025 - Springdale Flock Zero-Misidentification Reporting]]), and its counting method changed mid-series. LRPD says its figure reflects non-correlating matches and arrest-and-prosecution outcomes, but its visible dashboard supplies neither as a separate field; "success" is broader than both ([[T034 - LRPD Required Outcome Categories vs Flock Success Dashboard]]). The August 13 ASP release makes the first caveat concrete. Its earlier May-September report records 3,580,132 May scans, 45 confirmed matches, and 77 misidentifications, while the revised July-production file records 1,796,900, 10, and 36. Its August row changes from 20 confirmed matches and zero arrests/prosecutions to one confirmed match and one arrest/prosecution ([[ASP ALPR Data Report May-September 2025]], earlier report pp. 1-5; [[ASP Statewide ALPR Data Reports 2025-2026]], revised tables). The revised series remains the later agency-produced compilation used in the table above, but its calculation history cannot be audited from either release. The correct reading is therefore not "Arkansas ALPR is 17:1 wrong," nor that LRPD converts one in 3,848 alerts into an arrest. The statutory outcome fields are compiled inconsistently between agencies and vendors: ASP separates misidentifications and arrest/prosecution; Springdale reports a near-zero Flock misidentification field beside a much larger Axon one; LRPD substitutes a broad success taxonomy. A statewide error or enforcement-conversion rate remains unavailable, and these reports show why. ### Jonesboro: the first case-level outcome record Everything above measures at the program level. Jonesboro's Real Time Crime Center produced the first Arkansas record that measures at the **pull** level: 5,201 entries from December 2023 to June 2026, each one an officer asking the RTCC to check cameras for a case ([[Jonesboro RTCC Camera Pull Log 2023-2026]]). | Measure | Result | Denominator | |---|---|---:| | Evidence found | 3,956 yes / 781 no (+21 other, 9 nothing found) | 4,767 | | Arrest made | 136 yes / 1,527 no | 1,663 | | FOIA-driven | 694 yes / 4,083 no | 4,777 | | Traffic vs criminal | 1,339 / 284 | — | This sits at the **outcome-attribution** layer and it behaves differently from the statutory series. Where ASP reports 54 confirmed matches against 919 misidentifications, Jonesboro reports evidence found in roughly 83% of pulls. The two are not in conflict: ASP is measuring automated plate alerts against hot lists, Jonesboro is measuring an analyst deliberately retrieving footage for a known case. **They are different acts with different base rates**, and conflating them would reproduce exactly the error this page exists to prevent. Two caveats travel with these figures. The log was rebuilt three times across the period, so fields appear and disappear and no column spans all 31 tabs. And "evidence found" is the analyst's own characterisation at logging time, with no stated standard, no review step, and no relationship established to whether the evidence proved probative. What the dataset does establish, and nothing else in the corpus does, is the **ratio between retrieval and consequence**: a camera pull that finds something is ordinary, an arrest is not. That gap is the measurement territory the rest of this page argues is missing. ### Jonesboro: parallel vendor logs without comparable denominators The completion package adds monthly Flock and Axon workbooks, but their labels do not create a controlled head-to-head measure. Flock reports scans, official matches, custom matches, and arrests across March 2025–June 2026. Axon supplies only six months and labels one column `False` without defining whether it means false alerts, non-correlating matches, rejected reads, or something else. The Flock sheet also notes a missing February 2026 custom-match dataset ([[Jonesboro Flock and Axon ALPR Activity Logs 2025-2026]]). The correct inference is limited: Jonesboro retained program-activity figures from both vendors, and the schemas are materially different. A comparative false-positive or arrest-conversion claim would require common definitions, complete periods, duplicate-alert handling, and a documented link between a scan, alert, verification, and outcome. ### Jacksonville: unfinished statutory form and unmatched dashboard scopes Jacksonville's supplement contains a statutory-report template with 6,478,446 scans and 2,640 NCIC confirmed matches entered, but both reporting-term boxes, the compilation/public dates, and the supervisor signature fields are blank. It is an unfinished record, not a completed six-month report ([[Jacksonville Flock Compliance Analytics and Outcomes]], compliance DOCX sections 1, 2, 4, and 5). The same release has a Flock dashboard showing 1,535,124 selected-period scans, 7,756,225 year-to-date scans, and 15,604 hot-list alerts, plus a nine-success dashboard and a fifteen-row historical outcome CSV. Their visible periods and filters do not align. The records therefore add documentary counts at several layers without supplying a single error or conversion denominator. ## 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?