{"resourceId":"riscanvi-explainability-governance-analysis-2026","versions":[{"version":"external-776c65dbe01ae06d562ff96b74df85f55386aec0f4e34ec9dc8c12ad43a4944d","resource":{"id":"riscanvi-explainability-governance-analysis-2026","title":"RisCanvi analysis examines who can understand and challenge corrections risk scores","organization":"Oier Mentxaka and colleagues","sector":"Corrections and courts","geography":"Catalonia, Spain; limited U.S. transferability","publishedAt":"April 25, 2026","publicationDate":"2026-04-25","eventDate":null,"sourceName":"AI & SOCIETY, Springer Nature","sourceLabel":"Academic Open Forum analysis synthesizing prior technical and institutional evidence","sourceUrl":"https://link.springer.com/article/10.1007/s00146-026-03039-0","evidenceClass":"academic-research","outcomeClass":"cautionary","topics":["infrastructure","data-security","governance-procurement","accessibility-workforce","operating-model"],"finding":"The authors argue that restricted access to score explanations weakens meaningful challenge and oversight.","sledRelevance":"Historical international evidence adds a prison risk-assessment governance case; Spanish procedures and European legal claims do not establish U.S. obligations.","evidence":"Integrative analysis of prior audits, reporting and scholarship, with a proposed governance framework. It is not a new predictive-performance experiment or a deployed framework evaluation.","architectureImplications":"Interpretation: Assess database-integrated models through reproducible input, transformation and version records.","governanceImplications":"Interpretation: Evaluate whether affected people can actually initiate a correction and receive an answer.","securityPrivacyImplications":"Interpretation: Separate an individual's explanatory record from third-party sensitive information.","caveats":"Underlying audit datasets were not independently inspected. No new participant sample, comparator or measured benefit of the proposal is established. Do not infer that logistic regression is inherently unexplainable.","streamIds":["public-safety"],"roles":{"sales":"Interpretation: Corrections leaders, parole authorities, defense representatives and oversight bodies need to know whether a scoring workflow can be explained and corrected. Ask what information reaches the person assessed, who can dispute an input and whether the agency can reconstruct an older result. A bounded engagement could examine one existing review pathway using approved records. The value hypothesis is identifying obstacles to accountable operation, not increasing detention efficiency or promising lower recidivism. Applicability is to analogous decision-support processes. The paper does not validate a U.S. deployment, establish a local legal duty or show that a new explanatory interface alone improves justice outcomes.","engineering":"Interpretation: Fit is an assurance layer around an authorized existing system, conditional on access to the relevant scoring artifacts. Prerequisites include data lineage, exact model versions and permission to inspect transformations. Test whether a corrected input produces a reproducible changed result and whether the explanation matches the actual computation. Avoid asking a language model to invent reasons for an opaque score. Keep explanations within case permissions and verify redaction of third-party information. A proof of value should test reconstruction and user comprehension separately from predictive performance. Moving a database model to cloud infrastructure cannot by itself resolve limited access or unclear decision authority.","delivery":"Interpretation: A corrections records and review lead should work with legal counsel, analysts, front-line staff and appropriately represented affected people. Map the route from assessment to human decision, rehearse a disputed-record case and establish responsibility for correction. Dependencies include archival scoring records, authorized participation and a process owner able to act on findings. Proposed acceptance: every sampled test decision is reconstructible, each seeded input error reaches resolution, and users can identify a working review channel. Set response-time targets before the exercise. Risks include inaccessible notices, disclosures that reveal others' data and explanations that sound persuasive while leaving the underlying decision impossible to challenge."},"retrievedAt":"2026-09-14T03:01:18Z","enrichedAt":"2026-09-14T03:03:45Z","enrichmentBasis":"retrieved source","accessibilityWorkforceImplications":"Interpretation: Test comprehension with low-literacy and language-access needs, using approved participation safeguards.","procurementImplications":"Interpretation: Make usable audit access and data export conditions of any assurance engagement.","operatingModelImplications":"Interpretation: Assign a corrections process owner who can resolve disputes independently of the model supplier.","updateExplanation":"New-to-archive historical source adds institutional explainability analysis distinct from the archived COMPAS performance reanalysis. No fresh operational change claimed.","sourceVerification":{"openedUrl":"https://link.springer.com/article/10.1007/s00146-026-03039-0","referenceExcerpt":"our purpose here is not to assess predictive performance","promptVersion":"sled-research-v3.2","model":null,"basis":"agent-reported inspection"}}}]}