{"schema_version":1,"experiment_id":"eoa_inverse_innovation_exp04_retrieval_first_paired20_20260802","cell_id":"negative_space_design__tech_ethics_ai_governance","arm":"PROPOSAL_FIRST","candidate_id":"nsd_ai_governance_deliberation_void_v0","hypothesis_id":null,"version":0,"title":"Protected Deliberation Void for AI Deployment Reviews","problem":"In an AI deployment review, the sponsor recommendation, composite risk rating, checklist results, presentation, and live commentary can occupy the decision surface before each reviewer has formed an independent account of unresolved harms. The problem is not necessarily missing evidence: necessary evidence may be present but crowded by signals that imply closure. Without a bounded interval in which those competing signals are absent, reviewers may converge before distinguishing no detected concern from a concern that was never independently examined.","actors":["AI governance board members","Model or product sponsors","Risk, compliance, privacy, security, and legal reviewers","Representatives of people subject to the system","Meeting facilitator or governance chair","People whose access, treatment, work, or opportunities could be affected by the deployment"],"observable_state":"At the start of deliberation, the review interface simultaneously displays the sponsor's recommendation, an aggregate status or score, detailed evidence, checklist completions, and other reviewers' comments while the facilitator continues speaking. The record contains no protected point at which each reviewer must interpret the decision question independently, and blank risk fields do not reliably distinguish 'examined and no concern found' from 'not examined' or 'evidence unavailable.'","consequence":"Material uncertainty or dissent can remain perceptually buried even though it is technically available, producing a decision record that overstates how thoroughly harms were examined. A board could then authorize deployment without recognizing an unresolved impact, or be unable to reconstruct whether silence meant agreement, uncertainty, lack of evidence, or reluctance to speak.","affected_objective":"Independent, evidence-grounded identification of unresolved harms and accountable interpretation of uncertainty before an AI deployment decision.","intervention":"Introduce a two-pass protected deliberation void. After reviewers receive the complete evidence packet, the first deliberation screen shows only the decision question, affected actors, non-negotiable safety constraints, and explicitly labeled fields for 'concern found,' 'examined—no concern found,' and 'not assessable from available evidence.' For a short bounded interval, the interface withholds the sponsor recommendation, composite score, peer comments, notifications, and verbal framing; primary evidence and urgent safety information remain recoverable with one action. The facilitator reads a neutral prompt, holds silence while each reviewer records an independent note, and intervenes only for confusion, distress, or an urgent safety issue. Once every participating reviewer submits a note or the time boundary is reached, the omitted signals return for collective deliberation. The interface preserves the initial notes alongside later revisions so the board can see what emerged before and after social and summary signals appeared.","structural_mapping":[{"archetype_element":"Crowded field","domain_realization":"A governance screen and meeting opening filled simultaneously with evidence, composite ratings, sponsor advocacy, peer comments, and facilitator narration."},{"archetype_element":"Omission candidate","domain_realization":"The sponsor recommendation, composite score, peer comments, notifications, and nonessential facilitator speech are temporarily withheld during independent interpretation."},{"archetype_element":"Protected empty space","domain_realization":"A bounded first-pass interval and sparse decision canvas are reserved for independent reading, reflection, and note formation."},{"archetype_element":"Positive form relationship","domain_realization":"The absence frames the decision question, affected actors, red-line constraints, and unresolved evidence states as the forms requiring attention."},{"archetype_element":"Absence boundary","domain_realization":"Interface permissions and meeting rules prevent omitted signals from returning until a note is submitted, time expires, or a safety override is invoked."},{"archetype_element":"Meaning-of-absence check","domain_realization":"Distinct labels separate examined-with-no-concern, not-yet-examined, unavailable evidence, permission failure, and technical loading failure."},{"archetype_element":"Reintroduction trigger","domain_realization":"Recommendations, scores, comments, and ordinary discussion return after independent notes are captured or the bounded interval ends."},{"archetype_element":"Context preservation and recoverability","domain_realization":"Primary evidence, accessibility support, decision criteria, and urgent safety information remain available; every temporarily hidden item is restored without loss."},{"archetype_element":"Clarity or effect test","domain_realization":"A shadow comparison tests whether reviewers correctly interpret evidence gaps and produce independently grounded concerns without losing comprehension."}],"mechanism_mapping":[{"mechanism_slug":"editorial_cut","role":"Temporarily removes conclusion-like and redundant signals from the first-pass surface while preserving the complete packet and an audit trail of every cut item.","counterfactual_removal":"Without the editorial cut, the first pass remains crowded by sponsor and aggregate signals, so the intervention no longer creates meaningful absence around the decision question."},{"mechanism_slug":"focus_mode_or_control_hiding","role":"Clears comments, notifications, scoring chrome, and secondary controls while keeping primary evidence and safety-critical controls immediately recoverable.","counterfactual_removal":"Without focus mode, attention competitors remain continuously visible, and subtraction becomes either cosmetic or an unsafe permanent deletion."},{"mechanism_slug":"blank_or_rest_frame","role":"Marks the transition from evidence intake to judgment with an unmistakably intentional quiet screen and a defined exit into collective deliberation.","counterfactual_removal":"Without the rest frame, the process moves directly from presentation to discussion, leaving no perceptible boundary for consolidation."},{"mechanism_slug":"facilitation_silence","role":"Requires the person with the greatest procedural power to refrain from supplying interpretations while reviewers form their own accounts, with explicit reentry triggers for confusion, distress, or safety.","counterfactual_removal":"Without facilitation silence, verbal framing can refill the protected interval even if the interface is sparse."},{"mechanism_slug":"empty_state_design","role":"Makes blank risk fields diagnostically meaningful by distinguishing no concern, no examination, unavailable evidence, permission blockage, and system failure.","counterfactual_removal":"Without designed empty states, blankness is ambiguous and may falsely communicate either safety or malfunction."}],"causal_chain":["Simultaneous recommendations, scores, comments, controls, and speech compete with the unresolved deployment question for reviewer attention.","The process temporarily omits conclusion-like and socially anchoring signals while preserving necessary evidence and safety context.","The resulting sparse surface perceptually isolates affected actors, decision criteria, red-line constraints, and evidence gaps.","A bounded silence prevents the facilitator from immediately refilling that space with an interpretation.","Each reviewer records what the evidence supports, what concern remains, or why the issue cannot yet be assessed.","Explicit empty-state labels prevent nonresponse or missing evidence from being mistaken for an examined absence of concern.","The omitted signals then return for adversarial comparison, revision, and the board's authorized collective decision.","The record retains pre- and post-reintroduction judgments, making convergence, correction, and unresolved disagreement inspectable."],"baseline":"The ordinary baseline is a single-pass review: the sponsor presents the case and recommendation, a dashboard displays aggregate ratings and detailed evidence together, the chair opens continuous discussion immediately, and reviewers record a final position after hearing others. Blank fields are handled generically or informally.","nearest_rivals":["Strongest rival: require an independent written assessment before discussion while leaving the complete dashboard, sponsor recommendation, and aggregate score visible throughout.","Use a structured risk checklist with mandatory justification fields but no protected omission or silence.","Blind reviewers to the sponsor's identity or to one another's identities while retaining the ordinary information-dense review surface.","Assign a devil's advocate or red team to articulate objections during the normal meeting."],"remaining_contrastive_claim":"The remaining testable claim is whether protected, recoverable absence contributes something beyond an independent-writing requirement: holding the evidence, decision criteria, and note requirement constant, does temporarily withholding conclusion-like signals and facilitator speech produce more differentiated, correctly grounded accounts of unresolved impacts without increasing missed safety context or misunderstanding?","authority_safety":{"decision_authority":"The authorized AI governance board retains all live approval, rejection, escalation, and deployment authority. A governance chair or research lead may administer only an approved shadow evaluation; individual reviewers and the prototype cannot authorize deployment.","authorized_first_step":"With the governance chair's and data owner's approval, build a non-production prototype and run a consented shadow review of sanitized archived cases whose disposition cannot be changed by the exercise.","excluded_actions":["Do not hide primary evidence, uncertainty disclosures, accessibility support, legal obligations, or urgent safety information.","Do not use the prototype for a live deployment decision during the first test.","Do not treat reviewer silence or an empty field as consent or absence of harm.","Do not replace participation by affected people with interface-generated categories.","Do not expose identifiable case data or attribute sensitive reviewer notes beyond the approved evaluation protocol.","Do not let a sponsor or facilitator bypass the protected interval except through a logged safety or accessibility override."],"halt_rollback":"Halt the session and restore the ordinary full-information interface immediately if a reviewer cannot retrieve necessary evidence, interprets the sparse state as a system failure, reports coercive or distressing silence, or misses a safety-critical fact attributable to the intervention. Preserve the archived case unchanged and discard prototype-derived decision outputs from operational records."},"negative_tests":{"strongest_counterevidence":"In a counterbalanced shadow comparison, the protected-absence condition yields no additional correctly grounded concerns or evidence-gap distinctions beyond mandatory independent writing, while producing more comprehension errors, missed safety facts, or perceived pressure.","problem_falsifier":"The proposed problem is falsified in the target workflow if reviewers already document genuinely independent assessments before seeing recommendations or peer views, blank states are consistently interpreted correctly, and changing the visibility or timing of conclusion-like signals does not change what material concerns are noticed or recorded.","intervention_falsifier":"The intervention is falsified if temporarily omitting the selected signals causes reviewers to misunderstand the case or overlook necessary context, and the resulting concern records are no more independently differentiated or evidence-grounded than those produced by the full-information independent-writing rival.","risks":["Temporary omission could be used strategically to conceal inconvenient evidence or dissent rather than to protect attention.","Reviewers may mistake the sparse screen for missing data, a permission failure, or a technical fault.","Silence may amplify anxiety, exclusion, or power pressure, especially for participants unfamiliar with the procedure.","Hiding controls may create accessibility or discoverability barriers.","Pre-reintroduction notes may be overinterpreted as more authentic even when later evidence-based revisions are better.","The added interval may slow urgent decisions or become ritualized without improving interpretation.","Stored independent notes could expose sensitive dissent or create retaliation risk."]},"next_evidence_step":"Run a bounded, counterbalanced usability test with six to eight authorized reviewers and two sanitized archived deployment cases. Each reviewer uses the ordinary full-information independent-writing workflow for one case and the protected-deliberation-void prototype for the other. Before any collective discussion, record whether the reviewer correctly distinguishes examined-no-concern from unavailable or unexamined evidence, identifies case-grounded unresolved impacts, retrieves necessary context, and understands why signals are absent. Also record time, accessibility failures, perceived pressure, and safety-critical omissions. Do not infer live effectiveness from this test; use it only to decide whether a larger shadow evaluation is justified.","prior_art_status":"UNSEARCHED","revision_record":{"parent_version":null,"progress_targets_addressed":[],"conceptual_changes":[],"operational_changes":[],"evidence_changes":[],"claim_changes":[]}}