{"resourceId":"lumbroso-cybernetic-epistemic-agentic-delegation-20261001","versions":[{"version":"external-a1a8cc3036f7e13952d73f1a33a88c5b51110933fea6d495d96cfee6736750a7","resource":{"id":"lumbroso-cybernetic-epistemic-agentic-delegation-20261001","title":"Cybernetic vs epistemic oversight: approval-only HITL is rubber-stampable; require checkable decision conditions (ORRCF)","organization":"University of Pennsylvania (Lumbroso)","sector":"Higher education CS education / trustworthy agentic GenAI governance","geography":"United States (University of Pennsylvania; computing-education framing with cross-domain oversight examples)","publishedAt":"Submitted to arXiv 1 October 2026 (cs.AI)","publicationDate":"2026-10-01","eventDate":null,"sourceName":"arXiv","sourceLabel":"Academic preprint (position / conceptual framework)","sourceUrl":"https://arxiv.org/pdf/2610.00961","evidenceClass":"academic-research","outcomeClass":"cautionary","topics":["governance-procurement","developers-agents","operating-model","knowledge-work"],"finding":"Argues computing education's shift from writing code to supervising agents exposes a vocabulary gap: cybernetic language coordinates action, epistemic language coordinates checkable understanding. Explanation-shaped outputs can do cybernetic work (secure approval) without enabling verification. Oversight that only checks whether an output was approved is satisfiable by rubber-stamping. Proposes that consequential delegation choices carry the condition under which they would have gone otherwise, scored by a reconstruction test and recorded via ORRCF (Options, Recommendation, Rationale, Confidence, Falsifier). Illustrative episodes plus motivating opaque discretionary-risk example. Why now: after 2026-10-02 engagement-gate and actionable-feedback findings, HE GenAI tutoring/advising need oversight artifacts that are checkable, not merely approved.","sledRelevance":"Gives student-success and academic-technology teams language to reject rubber-stamp HITL in GenAI tutors, coding agents, and advising bots—complementing staged disclosure (2610.00870) and verification-by-task (2610.01158).","evidence":"Conceptual argument with illustrative episodes and secondary econometric motivation; not a student-outcome RCT; preprint.","architectureImplications":"Require agent decision records with falsifiers; support reconstruction tests on oversight logs; separate approval UI from reason-capture UI.","governanceImplications":"Do not accept 'human in the loop' as sufficient without checkable reasons; mandate ORRCF-like fields for consequential GenAI tutoring/advising actions.","securityPrivacyImplications":"Decision records may contain sensitive student context; purpose-limit access; preserve integrity of rationale logs against silent edits.","caveats":"Position paper; episodes not controlled experiments; transfer beyond CS coding agents is interpretive; preprint.","streamIds":["student-success"],"roles":{"sales":"For provosts buying GenAI tutors and advising agents with 'human oversight' checkboxes, Lumbroso shows approval-only HITL is rubber-stampable unless decisions carry checkable conditions. Sell governance design for ORRCF-style decision records and reconstruction tests—not another chatbot seat. Bound claims to this conceptual framework and refuse guaranteed bias elimination.","engineering":"Engineering should implement decision-record schemas (options, recommendation, rationale, confidence, falsifier), immutable audit logs, and reconstruction-test harnesses for consequential agent actions. Do not equate approve/deny UI with epistemic oversight. Constraints include conceptual evidence base and CS-education framing. Proof-of-value is that a second reader can predict agent behavior under perturbation from the record alone. Protect rationale logs as sensitive education/operations records.","delivery":"Delivery should map which tutoring/advising agent actions are consequential, require ORRCF fields before go-live, and sample reconstruction tests weekly. Ownership sits with academic-technology governance and student-success ops jointly. Train staff to reject approval-only workflows. Acceptance means documented consequential-action inventory and exportable reason logs. Risks include checkbox theatre and rationale logs that are unreadable in practice."},"roleDetails":{"sales":{"customerProblem":"Campuses treat human-approval checkboxes as sufficient GenAI oversight while agents produce explanation-shaped outputs optimized for acceptance.","stakeholders":["provost/AI governance","academic technology","student success","faculty senate","risk/compliance"],"discoveryQuestions":["Which agent actions only require click-to-approve?","Can a third party reconstruct why a recommendation was made?","Do logs include falsifiers?"],"valueHypothesis":"Requiring checkable decision conditions for consequential GenAI actions reduces rubber-stamp oversight risk without claiming automatic equity or accuracy.","potentialEngagement":"Governance workshop plus pilot ORRCF logging on one tutoring or advising agent workflow.","unsupportedClaims":["Guaranteed bias removal","RCT learning gains","All HITL is worthless"]},"engineering":{"fit":"HE GenAI tutoring, coding agents, and advising bots that currently rely on human approve/deny gates.","architectureIntegration":"Agent action → ORRCF decision record → approval UI that displays rationale/falsifier → immutable audit store → reconstruction-test jobs.","prerequisites":["consequential-action inventory","logging pipeline","role-based access"],"constraints":["conceptual evidence","CS-education origin","rationale quality depends on authors"],"security":"Rationale logs may contain student PII; integrity-protect records; least-privilege access.","proofOfValue":["Reconstruction test pass rate","No consequential action without falsifier field","Audit export works"]},"delivery":{"work":"Inventory consequential actions, configure ORRCF fields, train approvers, run weekly reconstruction samples, revise prompts/policies.","dependencies":["AI governance committee","academic technology","advising/teaching leads"],"ownership":"Academic-technology governance lead with student-success co-owner for advising/tutoring agents.","skillsAdoption":"Approvers read falsifiers before approving; engineers stop shipping approval-only gates for consequential actions.","governanceCheckpoints":["consequential-action inventory signoff","ORRCF schema review","monthly reconstruction audit"],"acceptanceCriteria":["ORRCF required on listed actions","Reconstruction sample documented","Exportable audit trail"],"risks":["Checkbox theatre","Unreadably long rationales","Scope creep to all trivial actions"]}},"retrievedAt":"2026-10-04T03:18:48Z","enrichedAt":"2026-10-04T03:21:30Z","enrichmentBasis":"retrieved source","accessibilityWorkforceImplications":"Rubber-stamp oversight can silently approve biased agent actions affecting who gets help; checkable reasons support equity review of automated tutoring/advising decisions.","procurementImplications":"Write contracts requiring exportable decision rationales with falsifiers for consequential agent actions, and rights to audit reconstruction quality—not only human-approval checkboxes.","operatingModelImplications":"Train faculty/advisors to reject approval-only agent workflows; sample reconstruction tests in weekly GenAI ops reviews.","sourceVerification":{"openedUrl":"https://arxiv.org/pdf/2610.00961","referenceExcerpt":"Oversight that checks only whether an output was approved is satisfiable by rubber-stamping and cannot see this failure","promptVersion":"sled-research-v4.0","model":null,"basis":"agent-reported inspection"}}}]}