{"schema_version":1,"experiment_id":"eoa_inverse_innovation_exp04_retrieval_first_paired20_20260802","cell_id":"negative_space_design__tech_ethics_ai_governance","arm":"RETRIEVAL_FIRST","round_index":0,"hypotheses":[{"hypothesis_id":"H1","title":"Protected abstention window for model release reviews","problem":"Release reviews become anchored on launch arguments before reviewers independently assess uncertain harms.","affected_stakeholder":"Internal model-risk reviewers","workflow_boundary":"Between submission of a model card and the first release recommendation","failure_mode":"Early advocacy crowds out independent risk discovery and produces premature consensus.","unit_of_analysis":"Release-review decision","causal_lever":"Withhold sponsor recommendations during a timed, independently documented risk-assessment interval.","archetype_mapping":"Facilitation silence creates protected procedural absence; sponsor views return only after reviewers record independent judgments.","expected_value":"Potentially more independent hazard findings and less anchoring without removing evidence.","falsifiable_claim":"Compared with standard reviews, a sponsor-silent first phase increases distinct independently raised material risks per review by at least 20% without increasing median review time by more than 15%.","diversity_rationale":"Targets meeting power and temporal sequencing rather than interface content, document density, incident status, or deployment scope.","mechanism_slugs":["facilitation_silence","blank_or_rest_frame"],"search_questions":["Do AI release reviews already use independent silent assessment before deliberation?","Does delayed disclosure of recommendations reduce anchoring in expert committees?","When does imposed silence suppress rather than broaden dissent?"]},{"hypothesis_id":"H2","title":"Governance dashboard that distinguishes benign silence","problem":"Dashboards make no incidents, missing telemetry, access denial, and pipeline failure look alike.","affected_stakeholder":"AI oversight and audit teams","workflow_boundary":"Opening an operational governance dashboard when a metric panel has no records","failure_mode":"Ambiguous emptiness is misread as evidence that no harm occurred.","unit_of_analysis":"Empty dashboard panel viewed during oversight","causal_lever":"Classify each empty panel by cause and expose its provenance, coverage, and next action.","archetype_mapping":"Empty-state design makes absence diagnostic while preserving context and a path to recover data.","expected_value":"Potentially fewer false assurances and faster escalation of monitoring gaps.","falsifiable_claim":"In scenario testing, typed empty states reduce erroneous 'no incident' conclusions by at least 30% relative to a generic zero-or-blank display.","diversity_rationale":"Targets interpretation of absent operational data at the interface boundary, unlike deliberation, policy editing, controls, or pilot capacity.","mechanism_slugs":["empty_state_design"],"search_questions":["How do current AI-governance dashboards encode missing versus zero data?","Are typed empty states used in safety, compliance, or observability systems?","Which provenance cues best prevent absence-of-evidence errors?"]},{"hypothesis_id":"H3","title":"Recoverable cuts in AI policy packets","problem":"Governance committees receive policy packets whose repeated principles and background obscure actionable evidence and unresolved exceptions.","affected_stakeholder":"Board-level AI governance committee members","workflow_boundary":"Preparation and reading of the pre-meeting decision packet","failure_mode":"Members skim crowded packets and miss decision-critical caveats.","unit_of_analysis":"Committee decision packet","causal_lever":"Move redundant or premature material into a recoverable appendix while retaining claims, evidence, uncertainties, and dissent in the main packet.","archetype_mapping":"Editorial cuts create absence around decision-critical content, with context-preservation and recoverability guardrails.","expected_value":"Potentially better recall of caveats and more evidence-based discussion at equal reading time.","falsifiable_claim":"Readers of edited packets identify at least 15% more seeded decision-critical caveats than readers of full packets, with no increase in unsupported conclusions.","diversity_rationale":"Targets document curation and individual comprehension, distinct from group silence, data-state diagnosis, interaction controls, and organizational deployment pacing.","mechanism_slugs":["editorial_cut","whitespace"],"search_questions":["What concise AI-governance packet formats already exist?","How does layered disclosure affect recall of risk evidence and caveats?","Which policy context becomes misleading when moved out of the primary packet?"]},{"hypothesis_id":"H4","title":"Risk-rehearsal focus mode for human override","problem":"Operators rehearsing high-risk AI decisions face dashboards crowded with analytics and administrative controls that compete with override cues.","affected_stakeholder":"Frontline operators accountable for AI-assisted decisions","workflow_boundary":"Timed simulation from alert presentation through accept, reject, or override action","failure_mode":"Secondary controls delay or displace the safety-critical action.","unit_of_analysis":"Operator response to one simulated high-risk alert","causal_lever":"Temporarily hide noncritical chrome while keeping evidence, uncertainty, override, and emergency controls visible and instantly restoring all controls on demand.","archetype_mapping":"Focus mode clears mapped attention competitors while exempting safety context and guaranteeing reintroduction.","expected_value":"Potentially faster, more accurate intervention without concealing model limitations.","falsifiable_claim":"In randomized simulations, focus mode reduces median correct-override time by at least 10% without increasing evidence-omission errors or control-recovery failures.","diversity_rationale":"Targets momentary human-computer action under urgency, rather than meetings, empty data, policy reading, or portfolio commitments.","mechanism_slugs":["focus_mode_or_control_hiding","sparse_layout"],"search_questions":["Do safety-critical AI interfaces already use temporary control hiding?","Which controls and evidence must never be hidden in high-risk decisions?","Does focus mode improve override performance across novice and expert operators?"]},{"hypothesis_id":"H5","title":"Reserved nondeployment capacity in AI portfolios","problem":"Organizations commit all pilot capacity to live use, leaving no protected room to investigate emerging harms before systems become entrenched.","affected_stakeholder":"People exposed to experimental organizational AI systems","workflow_boundary":"Annual allocation of AI pilot capacity before individual deployments are approved","failure_mode":"Portfolio saturation converts every learning opportunity into a live deployment and accelerates lock-in.","unit_of_analysis":"Organization-year AI pilot portfolio","causal_lever":"Reserve a bounded share of pilot capacity for nondeployment activities such as red-teaming, shadow evaluation, and affected-community inquiry, with explicit release criteria.","archetype_mapping":"An architectural void treats intentionally unoccupied deployment capacity as a protected program element that supports safer surrounding commitments.","expected_value":"Potentially earlier harm discovery and cheaper revision before operational dependence forms.","falsifiable_claim":"Organizations reserving at least 15% of pilot capacity for nondeployment inquiry detect more material predeployment issues per proposed system than matched organizations without a reserve, without reducing the number of validated pilots by more than 10%.","diversity_rationale":"Targets portfolio-level resource allocation and path dependence, structurally separate from single reviews, screens, documents, and operator actions.","mechanism_slugs":["architectural_void"],"search_questions":["Do AI portfolio-governance frameworks reserve capacity for nondeployment learning?","Does protected evaluation capacity produce earlier or cheaper design changes?","What reserve boundary prevents symbolic underuse or later infill?"]}]}