{"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","candidate_id":"CAND-H5-RESERVED-NONDEPLOYMENT-CAPACITY","hypothesis_id":"H5","version":0,"title":"Protected nondeployment reserve in annual AI pilot portfolios","problem":"During annual AI-pilot allocation, organizations can commit all available pilot capacity to live-use candidates, leaving no protected capacity to investigate emerging harms before operational commitments and dependencies form.","actors":["AI portfolio governing body","AI project teams proposing pilots","Independent risk, assurance, audit, or ethics personnel","Affected-community representatives or researchers","People exposed to experimental organizational AI systems"],"observable_state":"For each organization-year, records show total AI-pilot capacity, the share committed to live-use work before individual approvals, the share protected for nondeployment inquiry, attempted reallocations, release decisions, proposed systems examined, material issues detected before deployment, resulting design or approval changes, and the number of pilots ultimately validated. The target failure state is effectively saturated live-use capacity with no enforceable nondeployment reserve.","consequence":"Potential harms receive less dedicated predeployment investigation, so material issues may be discovered only after people are exposed and after revision has become more difficult; reserving capacity may instead enable earlier discovery, but could also displace useful pilots.","affected_objective":"Increase material predeployment issue detection for people exposed to experimental AI while limiting loss of validated organizational pilots.","intervention":"At the start of the annual allocation cycle and before individual pilots are approved, reserve at least 15% of total AI-pilot capacity exclusively for nondeployment harm inquiry, including red-teaming, shadow evaluation, and affected-community inquiry. Record the reserve as unavailable to live-use commitments. Permit release or reallocation only through documented criteria: the governing body confirms that eligible inquiry demand has been exhausted, independent assurance concurs, the reasons and destination are recorded, and affected inquiries retain enough capacity to finish. Unused capacity remains protected until that decision rather than being informally absorbed by deployment work.","structural_mapping":[{"archetype_element":"Omission Candidate","domain_realization":"Live deployment is deliberately omitted from the permitted uses of the reserved capacity."},{"archetype_element":"Protected Empty Space","domain_realization":"At least 15% of the organization-year pilot portfolio is kept unoccupied by live-use commitments."},{"archetype_element":"Positive Form Relationship","domain_realization":"The protected reserve supports safer decisions about the surrounding 85% or less of capacity that may become live pilots."},{"archetype_element":"Absence Boundary","domain_realization":"A portfolio rule, separate accounting, and approval-controlled release prevent ordinary project demand from consuming the reserve."},{"archetype_element":"Meaning-of-Absence Check","domain_realization":"Unused capacity is recorded as deliberate harm-inquiry capacity rather than mistaken for delay, inefficiency, or unclaimed project budget."},{"archetype_element":"Reintroduction Trigger","domain_realization":"Capacity may return to general allocation only after eligible inquiry demand is exhausted and the explicit release criteria are satisfied."},{"archetype_element":"Accessibility and Recoverability Guardrail","domain_realization":"The reserve cannot remove mandatory evidence, safety work, accessibility support, community participation, or routes for raising urgent concerns."},{"archetype_element":"Clarity or Effect Test","domain_realization":"Portfolio records test whether the reserve increases material issues detected per proposed system without reducing validated pilots by more than the specified tolerance."}],"mechanism_mapping":[{"mechanism_slug":"architectural_void","role":"Treats deliberately unoccupied deployment capacity as a planned portfolio element, sized and bounded in advance so it can support inquiry before surrounding commitments harden.","counterfactual_removal":"Without the protected void, live-use proposals can consume the full portfolio, leaving nondeployment inquiry to compete for residual capacity and eliminating the proposed causal lever."}],"causal_chain":["The portfolio governing body calculates total annual AI-pilot capacity before selecting individual deployments.","At least 15% is bounded against live-use commitments and assigned only to eligible nondeployment inquiry.","Independent investigators can schedule red-teaming, shadow evaluation, or affected-community inquiry without first displacing an approved deployment project.","More proposed systems receive inquiry while revision remains feasible and before people are exposed through live use.","Inquiry produces documented material issues and feeds them into design, rejection, or approval conditions.","Earlier changes reduce avoidable lock-in while the remaining portfolio continues through ordinary validation.","The intervention succeeds only if issue detection per proposed system rises and the number of validated pilots falls by no more than 10% relative to a credible comparison."],"baseline":"Annual portfolio allocation without a minimum protected share: assurance may still occur through project-specific reviews, risk-tiered approval, audits, or red teams, but its capacity competes with live-use commitments and can be reduced to residual availability.","nearest_rivals":["UNFPA ring-fenced evaluation funding, which closely matches protected annual evaluation resources and formal reallocation but is not an AI-pilot harm-inquiry rule.","Percentage-based innovation portfolio allocation described in public-sector innovation practice, which protects bounded learning categories but directs them toward innovation rather than AI-harm inquiry.","Independent AI assurance and three-lines governance, which separates challenge from development and supports assessment, audit, and red teaming but does not establish a stated organization-year pilot-capacity reserve.","Anthropic’s Responsible Scaling Policy, which institutionalizes evaluations, external review, red teaming, risk reporting, and possible pauses but is capability-threshold governance rather than a fixed multi-pilot annual reserve.","The patented cloud reserve-pool and release-governance approach in U.S. Patent Application 20260057320, which uses planning-period reserves and controlled release for cloud cost or runtime governance rather than AI-harm inquiry."],"remaining_contrastive_claim":"The remaining distinction is an AI-specific organization-year rule that reserves a stated minimum share of total pilot capacity exclusively for nondeployment harm inquiry before individual pilots are approved, protects that capacity from live-use commitments, and permits release only under explicit criteria. This is only a bounded contrast against the supplied prior-art search, not a claim of world novelty.","authority_safety":{"decision_authority":"The organizational body already authorized to allocate the annual AI-pilot portfolio, with concurrence from an independent risk, assurance, audit, or ethics function for release decisions.","authorized_first_step":"Conduct a records-only, counterfactual shadow allocation using one completed organization-year of proposed AI pilots; no live approval, budget, staffing, or deployment decision changes during this first step.","excluded_actions":["Deploying an unapproved system","Treating reserved inquiry as permission to expose people experimentally","Diverting legally or operationally mandatory safety, privacy, security, accessibility, or incident-response resources","Concealing uncertainty, dissent, adverse findings, or the fact that capacity was released","Using affected-community inquiry without appropriate consent, compensation, privacy protection, and safeguarding","Automatically blocking urgent remediation because the annual reserve is exhausted","Allowing project teams alone to release protected capacity"],"halt_rollback":"Halt the first step if required records cannot be used lawfully, issue materiality cannot be assessed independently, or the reconstruction would expose sensitive personal or security information. Because the first step is records-only, rollback consists of terminating analysis, deleting or returning derived working data under the organization’s existing data-handling rules, and making no portfolio change."},"negative_tests":{"strongest_counterevidence":"Existing independent AI assurance, risk-tiered approval, and recurring evaluation programs may already obtain equivalent or better predeployment coverage without a fixed portfolio percentage; the adjacent UNFPA and innovation-portfolio precedents also show that ring-fencing and percentage allocation are established structures, sharply limiting any novelty claim.","problem_falsifier":"The problem is falsified if saturated organization-years already provide timely, adequately resourced nondeployment inquiry for proposed systems, or if lack of protected capacity is not associated with missed or later material findings after accounting for proposal mix and risk tier.","intervention_falsifier":"The intervention is falsified if a credible matched or within-organization comparison shows that a reserve of at least 15% does not increase independently adjudicated material predeployment issues detected per proposed system, or reduces the number of validated pilots by more than 10%; it is also falsified as a protected-void mechanism if the reserved share is routinely diverted without satisfying its release rules.","risks":["The 15% threshold may be arbitrary, excessive, or insufficient for a particular portfolio.","Teams may relabel ordinary development or compliance work as nondeployment inquiry to satisfy the rule symbolically.","A fixed reserve may idle despite urgent useful pilots or may encourage spending merely to exhaust the allocation.","Issue counts may reward low-value findings unless materiality is defined and adjudicated independently.","Organizations with stronger safety cultures may both create reserves and detect more issues, confounding observational comparisons.","Affected-community inquiry can become extractive or coercive if participation safeguards are weak.","Release authority may become a bottleneck or may approve routine erosion of the reserve.","Focusing on annual pilots may omit continuous deployment, procurement, or model-update pathways."]},"next_evidence_step":"For one organization and one completed annual portfolio, reconstruct total proposed-pilot capacity and apply a preregistered 15% shadow reserve before proposal selection. Using only existing records, identify which eligible red-teaming, shadow-evaluation, or affected-community inquiries could have occupied the reserve; have an independent reviewer adjudicate issue materiality; then compare estimated material predeployment issues per proposed system and validated-pilot count with the actual allocation. Treat the result as feasibility and measurement evidence only, not causal confirmation, and do not alter deployments.","prior_art_status":"SEARCHED_BOUNDED","revision_record":{"parent_version":null,"progress_targets_addressed":["Convert selected hypothesis H5 into a complete, testable candidate.","Preserve the portfolio-saturation problem and protected-capacity causal lever.","Incorporate the independent prior-art criticism without expanding the residual distinction.","Specify authority limits, safety exclusions, falsifiers, and a bounded first evidence step."],"conceptual_changes":["Defined the reserve as a deliberately protected absence within the organization-year AI-pilot portfolio.","Limited the residual distinction to the specific combination recorded by the bounded prior-art search.","Separated the proposed efficacy claim from any claim of world novelty."],"operational_changes":["Specified a minimum 15% reserve, exclusive eligible uses, accounting boundary, and explicit release conditions.","Defined observable portfolio, issue-detection, and validated-pilot measures.","Limited the first evidence step to a reversible records-only shadow allocation."],"evidence_changes":["Named the closest supplied analogues and their remaining differences.","Carried forward the bounded eight-query, eight-source prior-art status.","Added comparison and measurement requirements without introducing external evidence."],"claim_changes":["Retained the hypothesis threshold of increased material issues per proposed system with no more than a 10% reduction in validated pilots.","Qualified novelty as a bounded remaining contrastive claim rather than world novelty.","Added explicit problem and intervention falsifiers."]}}