{"resourceId":"whro-norfolk-harvard-triage-scores-20260918","versions":[{"version":"external-0bff7af54df0cb44783b499d8946d671367f8f9ec28396dcee8c51244fe498c0","resource":{"id":"whro-norfolk-harvard-triage-scores-20260918","title":"WHRO: Norfolk prosecutors partner with Harvard on AI triage-score plea pilot","organization":"WHRO Public Media","sector":"Courts / prosecution","geography":"United States (Norfolk, Virginia; Philadelphia partner site)","publishedAt":"September 18, 2026","publicationDate":"2026-09-18","eventDate":"2026-09-15","sourceName":"WHRO","sourceLabel":"Local public-media report","sourceUrl":"https://www.whro.org/local-government/2026-09-18/260918-harvard-nfk-saunders","evidenceClass":"independent-reporting","outcomeClass":"emerging","topics":["knowledge-work","governance-procurement","operating-model","developers-agents"],"finding":"On 2026-09-18 WHRO reports Norfolk City Council accepted a $38,251 Harvard Access to Justice Lab grant to fund a part-time data analyst in the Commonwealth's Attorney's Office for a 24-month 'Triage Scores' study, with Philadelphia Legal Assistance as the other initial partner site. Triage scores are described as AI-based alternatives to conventional risk scores that weigh effects of decisions such as plea offers on reoffending. Commonwealth's Attorney Fatehi states there is no outside body independently overseeing how recommendations are used and that his office will remain the safeguard via individualized assessment rather than mechanical adherence. Findings are to be published after the study concludes.","sledRelevance":"Concrete U.S. prosecutor implementation signal for courts/public-safety stream: plea-decision AI pilots need evaluation design and oversight architecture before claiming public-safety benefit.","evidence":"Opened WHRO HTTPS article via WebFetch and curl; confirmed grant amount, 24-month study, partner sites, Fatehi oversight quotes, and absence of measured outcomes.","architectureImplications":"Keep triage outputs advisory in CMS; log recommendations and human overrides; separate research feature pipelines from automated plea generation.","governanceImplications":"Charter independent or external review despite office self-safeguard claims; pre-register metrics; engage defense stakeholders early.","securityPrivacyImplications":"Research extracts of case data need DUAs, minimization, and bars on model training reuse of confidential files.","caveats":"No outcome results yet; partner naming varies in quotes (Legal Assistance vs Public Defender Service); Crimson used secondarily for professor names.","streamIds":["public-safety"],"roles":{"sales":"Sales framing: Norfolk's Harvard-backed triage-score pilot is an emerging prosecutor decision-support use case, not a finished outcome study. Position advisory work around study design, data-sharing governance, and external review—not claims that triage scores reduce reoffending. Stakeholders are the elected prosecutor, city finance (grant acceptance), defense bar, and research partners. Bounded engagement is standing up evaluation governance for a plea-decision pilot; this article cannot support guaranteed recidivism improvements.","engineering":"Engineering framing: triage scores differ architecturally from classic risk scores by targeting counterfactual effects of decisions (e.g., offer plea vs not). Integration should keep model outputs advisory, log features/recommendations/overrides, and isolate case data used for research from production charging automation. Prerequisites: data-use agreements with researchers, CJIS/privacy review, and explicit non-automation policy. Constraint called out in-source: no outside independent oversight body for recommendation use—office self-safeguard only.","delivery":"Delivery framing: deliver the part-time analyst workflow, data compilation, and outcome tracking for a 24-month study while locking human final authority on pleas. Dependencies: council grant acceptance, Harvard A2J methods, defense-side partner site, and publication plan. Ownership with Commonwealth's Attorney leadership. Acceptance criteria: documented individualized review, override logging, and no mechanical plea offers from scores. Risks: automation bias, fairness challenges, and credibility gaps if oversight stays entirely internal."},"roleDetails":{"sales":{"customerProblem":"Prosecutors want better decision aids for pleas but lack evaluated alternatives to conventional risk scores and face oversight expectations.","stakeholders":["Commonwealth's/District Attorney","City council / grant admin","Public defender / defense bar","Research university partner","Victim/witness coordinators"],"discoveryQuestions":["Will recommendations be advisory-only with logged overrides?","What independent review exists beyond the office itself?","How is training/test data minimized and retained?","What fairness metrics and stop criteria exist before wider use?"],"valueHypothesis":"Bounded value is learning whether counterfactual triage aids improve plea decisions under human review—not proven recidivism reduction yet.","potentialEngagement":"Governance design for a 90-day pilot start: DPAs, advisory-only controls, override taxonomy, and external ethics/advisory charter.","unsupportedClaims":["Measured reduction in reoffending from Norfolk triage scores","That office self-safeguard equals independent oversight","Guaranteed plea fairness improvements"]},"engineering":{"fit":"Research/pilot decision-support alongside case management—not autonomous charging or detention engines.","architectureIntegration":"Batch/feature store for study variables; score API returns advisory triage output; CMS stores recommendation, human decision, and rationale; no auto-filing of pleas.","prerequisites":["Executed research DUA","Privacy/CJIS review","Override and audit logging","Clear separation of study vs production systems"],"constraints":["24-month study before published findings","No external oversight body named in reporting","Plea context differs from pretrial RCT illustrations"],"security":"Limit PII to need-to-know analyst role; prohibit vendor training on confidential case files; encrypt research extracts in transit/at rest.","proofOfValue":["Override rate and reasons tracked","No automated plea offers in audit","Pre-registered evaluation metrics with researchers"]},"delivery":{"work":"Stand up analyst role, data pipelines for case/outcome features, prosecutor training on advisory use, and quarterly governance reviews.","dependencies":["Harvard A2J Lab methods team","City grant disbursement","Defense partner site coordination","CMS reporting fields"],"ownership":"Commonwealth's Attorney as operational owner of plea decisions; research partner owns model development/evaluation.","skillsAdoption":"Train prosecutors that scores estimate decision effects under uncertainty and never replace individualized assessment.","governanceCheckpoints":["Kickoff ethics/review charter","Midpoint fairness audit","Go/no-go before any expansion beyond study"],"acceptanceCriteria":["100% recommendations human-reviewed","Overrides logged","Public findings plan documented"],"risks":["Internal-only safeguard loses public trust","Feature leakage of protected attributes","Premature operationalization before results"]}},"retrievedAt":"2026-09-19T03:15:05Z","enrichedAt":"2026-09-19T03:16:30Z","enrichmentBasis":"retrieved source","accessibilityWorkforceImplications":"Part-time analyst capacity and prosecutor training on advisory use are prerequisites; under-resourced offices cannot copy the grant structure blindly.","procurementImplications":"Grant-funded research pilots still need privacy review, audit logging, and exit/data-destruction terms with university partners.","operatingModelImplications":"Plea AI must preserve individualized human assessment; publish evaluation plans before expanding beyond study sites.","sourceVerification":{"openedUrl":"https://www.whro.org/local-government/2026-09-18/260918-harvard-nfk-saunders","referenceExcerpt":"there is no outside body independently overseeing how the tool's recommendations get used","promptVersion":"sled-research-v4.0","model":null,"basis":"agent-reported inspection"}}}]}