{"schema_version":1,"research_id":"eoa_inverse_innovation_exp06_external_evaluation_20260803","source_assessment_id":"bounded_rivalry_governance__mathematics:P4:v0","cell_id":"bounded_rivalry_governance__mathematics","search_queries":["HPC shared facility allocation scarcity queue walltime overestimate utilization research paper","auction based allocation HPC compute resources internal credits second price","site:docs.nersc.gov queue policy reservations walltime allocation official","site:operations.access-ci.org allocations policy resource provider official","Tycoon distributed market-based resource allocation virtual currency computing auction paper","Bellagio distributed resource allocation auction virtual currency computing paper","HPC internal economy credits auction scheduler research","high performance computing user requested walltime overestimate utilization paper","NERSC official policy reservations dedicated system access queue long jobs walltime","site:docs.nersc.gov jobs policy qos walltime limits reservations","site:slurm.schedmd.com multifactor fairshare official priority documentation","NSF ACCESS allocation policy demand exceeds capacity official","official guidance bid rigging screens patterns not proof collusion procurement data","OECD bid rigging screening indicators do not prove collusion official","DOJ procurement collusion red flags bid rotation cover bids official","competition authority algorithmic screening bid rigging false positives official","site:slurm.schedmd.com priority_multifactor fair-share official","site:slurm.schedmd.com resource_limits fairshare account QOS official scheduling","Slurm official advance reservations documentation priority multifactor fairshare","\"Job scheduling with adjusted runtime estimates on production supercomputers\" full text","\"Job scheduling with adjusted runtime estimates\" DOI","site:osti.gov HPC user requested walltime inaccurate production supercomputers","site:osti.gov/biblio \"Towards Understanding HPC Users and Systems\"","\"Towards Understanding HPC Users and Systems: A NERSC Case Study\""],"sources":[{"source_id":"S1","title":"Allocations Policy","publisher":"ACCESS Resource Allocations Marketplace and Platform Services / National Science Foundation","url":"https://access-ci.atlassian.net/wiki/spaces/ACCESSdocumentation/pages/942669827/Allocations%2BPolicy","source_class":"OFFICIAL_GUIDANCE","publication_date":"not stated","accessed_at":"2026-08-03","claims_supported":["ACCESS expressly curates allocations when demand exceeds available cyberinfrastructure supply.","ACCESS identifies efficiency, equity, transparency, scalable review, and reduced barriers as allocation objectives.","The RAMPS team administers the process, with policies approved by NSF, identifying credible operators and authorizers."]},{"source_id":"S2","title":"Queues and Charges","publisher":"National Energy Research Scientific Computing Center, Lawrence Berkeley National Laboratory","url":"https://docs.nersc.gov/jobs/policy/","source_class":"OFFICIAL_GUIDANCE","publication_date":"not stated","accessed_at":"2026-08-03","claims_supported":["A production scientific-computing facility already differentiates service through priority, walltime, node-count, submission, run, and accounting rules.","NERSC charges projects in node-hours and uses QOS multipliers, including premium and preemptible service.","NERSC charges resources made unavailable to other jobs, demonstrating that reservation and operational spillovers are observable and administratively attributable."]},{"source_id":"S3","title":"Reservations","publisher":"National Energy Research Scientific Computing Center, Lawrence Berkeley National Laboratory","url":"https://docs.nersc.gov/jobs/reservations/","source_class":"OFFICIAL_GUIDANCE","publication_date":"not stated","accessed_at":"2026-08-03","claims_supported":["NERSC supports approved dedicated reservations for needs not accommodated by the regular batch system.","Reservations can exceed ordinary walltime limits, terminate jobs at the reservation end, and be ended early through operations.","This demonstrates technical and organizational authority to control indivisible time windows, eligibility, cancellation, and release."]},{"source_id":"S4","title":"Slurm Workload Manager — Multifactor Priority Plugin","publisher":"SchedMD","url":"https://slurm.schedmd.com/priority_multifactor.html","source_class":"OFFICIAL_PRODUCT_DOCUMENTATION","publication_date":"not stated","accessed_at":"2026-08-03","claims_supported":["Established schedulers already combine age, fair-share, job size, QOS, site, and resource factors to prioritize jobs.","Slurm fair-share uses allocated and consumed resources and can account for CPUs, nodes, memory, licenses, and generic resources.","Administrative weights and accounting databases provide an implementable baseline and integration path, but also make the proposed auction incremental rather than a replacement for an absent scheduling capability."]},{"source_id":"S5","title":"Tycoon: An implementation of a distributed, market-based resource allocation system","publisher":"SAGE Publications / Multiagent and Grid Systems","url":"https://journals.sagepub.com/doi/abs/10.3233/MGS-2005-1303","source_class":"PRIMARY_RESEARCH","publication_date":"2005-08-01","accessed_at":"2026-08-03","claims_supported":["Tycoon implemented and experimentally evaluated market-based allocation of shared computing resources.","Its stated purpose includes letting users differentiate the value of their jobs where proportional shares cannot.","This substantially overlaps the candidate's opportunity-cost bidding rationale, although Tycoon allocates proportional shares continuously rather than bounded, indivisible reservations."]},{"source_id":"S6","title":"Resource Allocation in Federated Distributed Computing Infrastructures","publisher":"University of California, San Diego / OASIS Workshop","url":"https://cseweb.ucsd.edu/~snoeren/papers/share-oasis04.html","source_class":"PRIMARY_RESEARCH","publication_date":"2004-10-01","accessed_at":"2026-08-03","claims_supported":["Bellagio uses a market-based combinatorial auction to allocate finite computing resources among self-interested users.","Its bidding language represents preferences over differentiated resource bundles and seeks truthful valuation and economic efficiency.","Bellagio is the closest located analogue to internal-credit bidding for compute resources, though it does not establish this candidate's complete readiness, bond, standby, monitoring, and appeal package."]},{"source_id":"S7","title":"Towards understanding HPC users and systems: A NERSC case study","publisher":"University of California eScholarship / Journal of Parallel and Distributed Computing","url":"https://escholarship.org/uc/item/9g63d08m","source_class":"PRIMARY_RESEARCH","publication_date":"2018-01-01","accessed_at":"2026-08-03","claims_supported":["A primary study of multiple NERSC systems documents heterogeneous jobs, queues, and performance over production workloads.","The authors identify workload change and heterogeneity as challenges to classical HPC schedulers and call for informed scheduling research.","The evidence supports the importance of empirical workload characterization but does not establish urgency exaggeration or reciprocal priority trading at a mathematics institute."]},{"source_id":"S8","title":"Ex officio Cartel Investigation","publisher":"Organisation for Economic Co-operation and Development","url":"https://www.oecd.org/content/dam/oecd/en/publications/reports/2013/10/ex-officio-cartel-investigation_4ebcb223/4089581e-en.pdf","source_class":"OFFICIAL_GUIDANCE","publication_date":"2013-10-30","accessed_at":"2026-08-03","claims_supported":["Cross-round bid data can be screened for correlation and other patterns associated with bid rigging.","Screens can generate false positives and false negatives, can be evaded, require adequate data and expertise, and do not prove collusion.","Screening should initiate an investigation rather than itself determine liability or punishment."]}],"problem_evidence":{"support":"MODERATE","rationale":"Scarcity and consequential allocation visibly exist: ACCESS expressly curates requests when demand exceeds supply, while NERSC operates priority, accounting, reservation, cancellation, and long-walltime rules for shared scientific computing. Production-workload research also documents heterogeneous jobs and scheduling challenges. However, no direct source establishes the candidate's institute-specific informal queue, urgency exaggeration, sham projects, reciprocal priority deals, or the frequency with which truly indivisible mathematical windows collide.","source_ids":["S1","S2","S3","S7"]},"stakeholder_evidence":{"support":"MODERATE","rationale":"ACCESS RAMPS, NSF, NERSC, and comparable facility operators are identifiable authorizers or adopters with expressed needs for efficient, equitable, transparent allocation under scarcity. NERSC already controls reservations and resource charges. No source expresses demand for an internal auction specifically, and no mathematics institute has agreed to supply data or run the shadow exercise.","source_ids":["S1","S2","S3"]},"prior_art":{"proximity":"SUBSTANTIAL_COLLISION","closest_analogues":[{"name":"Bellagio combinatorial auction for distributed computing resources","similarity":"Allocates finite compute resources among self-interested users through bids representing differentiated preferences; this collides with the candidate's core market-allocation concept.","remaining_difference":"The located description does not combine an indivisible uninterrupted-window auction with stable-team nontransferable endowments, readiness gating, separately calculated externality charges, a refundable credit bond, standby transfer, anomaly-only collusion referral, appeals, and mandatory retirement review.","source_ids":["S6"]},{"name":"Tycoon market-based shared-compute allocator","similarity":"Implements and experimentally evaluates market allocation intended to let users express differing job values rather than rely on static proportional shares.","remaining_difference":"Tycoon uses proportional-share, low-latency allocation rather than a seasonal sealed second-price auction for indivisible reservations with the candidate's governance bundle.","source_ids":["S5"]},{"name":"Slurm multifactor fair-share scheduling","similarity":"Uses past consumption, allocations, job size, QOS, site policy, and resource accounting to prioritize contending jobs without ranking scientific merit.","remaining_difference":"It applies administrator-weighted scheduling factors rather than asking teams to sacrifice finite future-access credits through a sealed second-price bid.","source_ids":["S4"]},{"name":"NERSC QOS, project accounting, and dedicated reservations","similarity":"Production practice already differentiates priority and cost, charges unavailable resources, approves special reservations, permits long walltimes, and manages early release.","remaining_difference":"NERSC's published process does not use the proposed internal-credit auction to select between simultaneous qualified requests for one reservation.","source_ids":["S2","S3"]},{"name":"ACCESS allocations marketplace and merit-review process","similarity":"A transparent, governed process curates scarce cyberinfrastructure across disciplines and scales justification with requested resources.","remaining_difference":"ACCESS relies on request and merit-review procedures rather than repeated sealed bids against team-specific nontransferable credit budgets.","source_ids":["S1"]}],"distinctive_claim_remaining":"In a preregistered shadow test of repeated, indivisible compute-window conflicts, the complete package—frozen eligibility, nontransferable budgeted credits, sealed second-price clearing, separately computed spillover charges, refundable credit bonds, standby transfer, anomaly-only collusion referral, appeals, and scheduled review—will be independently reproducible and will reduce unused-window time plus unrecovered operational burden relative to first-ready, round-robin, and lottery allocation, without allocating an unready job, producing greater concentration than a preregistered tolerance, repeatedly accusing innocent correlated bidders, or exceeding a preset administrative-time ceiling. Failure on any criterion falsifies progression; the auction alone is not distinctive.","confidence":"HIGH"},"implementation_evidence":{"support":"MODERATE","rationale":"Existing facility reservation, QOS, accounting, and scheduler controls make eligibility, window assignment, resource metering, priority accounting, cancellation, and rollback technically credible. Tycoon and Bellagio demonstrate that market-based compute allocation can be implemented. OECD guidance supports anomaly-to-investigation separation while warning that screens are data- and expertise-intensive and cannot prove collusion. Missing evidence concerns team identity governance, confidential job descriptions, credit calibration, second-price behavior in a small repeated field, facility-specific integration, appeals staffing, and whether operational metadata suffice to distinguish coordination from innocent correlated demand.","source_ids":["S2","S3","S4","S5","S6","S8"]},"scores":{"meaningful_impact":{"score":3,"rationale":"Improved allocation could matter for rare, high-cost uninterrupted computations, but the prevalence and consequence of conflicting windows at a mathematics institute are unmeasured.","source_ids":["S1","S2","S3","S7"]},"stakeholder_pull":{"score":3,"rationale":"Credible operators explicitly need efficient and equitable scarcity management, but none requests this auction design or offers a pilot partnership.","source_ids":["S1","S2","S3"]},"incremental_advantage":{"score":2,"rationale":"No comparative evidence shows superiority to Slurm fair-share, QOS pricing, ordinary reservations, first-ready scheduling, rotation, lottery, or peer review.","source_ids":["S2","S3","S4","S5","S6"]},"distinctiveness_plausibility":{"score":2,"rationale":"Bellagio and Tycoon substantially collide with market-based compute allocation; distinctiveness survives only in the specific governance bundle and mathematics-window application.","source_ids":["S5","S6"]},"technical_implementability":{"score":4,"rationale":"Reservation, accounting, priority, resource-metering, auction, and audit components all have credible precedents, though integration and strategic robustness remain untested.","source_ids":["S2","S3","S4","S5","S6"]},"adoption_authority_feasibility":{"score":3,"rationale":"Facility operators demonstrably control allocation policies and reservations, but a live auction would require local approval, privacy review, appeals capacity, and user-governance legitimacy.","source_ids":["S1","S2","S3"]},"evidence_readiness":{"score":2,"rationale":"A safe scripted test is well bounded, but there are no institute-specific logs, stakeholders, calibrated parameters, behavioral results, or verified cost estimates.","source_ids":["S5","S6","S7","S8"]},"safety_net_benefit":{"score":4,"rationale":"A no-stakes shadow design, frozen rules, independent reproduction, standby transfer, appeal, anomaly-only referral, and rollback sharply limit first-step harm.","source_ids":["S3","S8"]},"scalability":{"score":3,"rationale":"Software clearing and accounting should scale, but identity administration, appeals, confidential eligibility review, and collusion analysis may become costly; small repeated fields also create strategic risks.","source_ids":["S4","S5","S6","S8"]}},"score_confidence":"MODERATE","costs":{"first_evidence":{"band_2026_usd":"10K_TO_50K","scope":"Design, preregister, execute, independently reproduce, and analyze a 12-round no-stakes shadow exercise with scripted teams, three comparator rules, 24 innocent/adversarial scenarios, and a short governance review.","confidence":"MODERATE","assumptions":["No production queue, hardware acquisition, or paid compute is used.","Approximately 120-300 combined hours from an HPC administrator, mechanism-design analyst, research coordinator, and independent reviewer are valued at 2026 fully loaded resource-equivalent rates.","Existing scheduler logs or synthetic job cards can be transformed without purchasing proprietary data."],"source_ids":["S4","S5","S6","S8"]},"initial_deployment_startup":{"band_2026_usd":"50K_TO_250K","scope":"Build and security-test the auction service; integrate identity, scheduler, accounting, sealed-record, surcharge, bond, standby, and appeal workflows; draft frozen policies and privacy controls.","confidence":"LOW","assumptions":["One facility and one reservation class are in scope.","Approximately 500-1,500 engineering, policy, security, and administrator hours are required.","Existing Slurm-compatible reservation and accounting infrastructure is available; no new HPC hardware is included."],"source_ids":["S2","S3","S4","S8"]},"operational_launch":{"band_2026_usd":"50K_TO_250K","scope":"Run one tightly controlled allocation season for one window class, including training, help desk, independent clearing checks, manual appeal coverage, monitoring, incident response, and post-season evaluation.","confidence":"LOW","assumptions":["Launch follows successful shadow evidence and separate authorization.","Approximately 600-1,800 combined staff hours cover the first season and heightened oversight.","Compute-window opportunity cost and foregone research output are excluded because facility-specific utilization data are unavailable."],"source_ids":["S1","S2","S3","S8"]},"annual_recurring":{"band_2026_usd":"50K_TO_250K","scope":"Maintain software and policies, administer rounds, review eligibility and appeals, monitor anomalies, audit identities and resource charges, recalibrate parameters, and conduct annual impact review.","confidence":"LOW","assumptions":["Recurring work is roughly 0.5-1.5 fully loaded FTE plus modest software, security, and independent-review support.","The number of contested windows and investigations remains low to moderate.","No monetary prize, credit market, new hardware, or general surveillance system is operated."],"source_ids":["S2","S3","S4","S8"]}},"verified_pipeline_gates":{"externally_supported_problem":{"status":"YES","reason":"Official operators document demand exceeding supply, priority queues, charged scarcity, and special reservations; primary research documents heterogeneous production workloads and scheduling challenges. The specific manipulation story remains unverified but is not required to establish the broader scarcity problem.","source_ids":["S1","S2","S3","S7"]},"externally_credible_adopter_or_authorizer":{"status":"YES","reason":"ACCESS RAMPS and NSF approve allocation policy, and NERSC controls QOS, accounting, and dedicated reservations. These are credible authorizer analogues, although none has expressed interest in this auction.","source_ids":["S1","S2","S3"]},"distinct_testable_incremental_claim":{"status":"YES","reason":"After substantial prior-art collision, a contrastive claim remains for the complete governance package and can be compared against first-ready, round-robin, and lottery rules using preregistered operational, concentration, integrity, and staff-time outcomes.","source_ids":["S4","S5","S6","S8"]},"bounded_next_evidence_step":{"status":"YES","reason":"A 12-round, no-stakes scripted shadow exercise has fixed comparators, observables, abuse cases, independent reproduction, and explicit stop conditions without affecting live access.","source_ids":["S3","S5","S6","S8"]},"no_unresolved_safety_or_authority_stop":{"status":"YES","reason":"The next step can be confined to synthetic or voluntarily deidentified job cards, with no live scheduling, real credits, employment effects, research ranking, private-communication inspection, or accusation from a screen. Live deployment would require a new authority review.","source_ids":["S1","S3","S8"]},"credible_cost_scope_and_range":{"status":"UNCERTAIN","reason":"Resource-equivalent bands are transparently scoped and broad, but no source supplies facility-specific labor rates, integration effort, dispute volume, opportunity cost, or annual round count.","source_ids":["S2","S3","S4","S8"]}},"next_evidence_step":"With a willing facility operator, preregister and run 12 no-stakes allocation rounds using deidentified or synthetic job cards. Include at least 24 labeled scenarios covering ordinary demand, innocent correlated deadlines, unstable runtime estimates, early release, avoidable nonuse, overrun, sham identities, cover bids, bid rotation, and side-transfer attempts. Give simulated teams equal nontransferable budgets and independently compute the sealed second-price outcomes. Replay the identical requests under first-ready, equal round-robin, and qualified lottery rules. Measure clearing reproducibility, eligibility violations, consistency between bids and preregistered ordinal priority, unused-window minutes, recovered spillover burden, standby success, winner concentration, innocent flags, detected scripted coordination, appeal reversals, and total staff time. Do not progress if either implementation disagrees on any clearing result; any unready or unsafe card wins; auction unused time or unrecovered burden is not lower than every comparator; concentration exceeds round-robin by more than 0.10 HHI without a preregistered justification; any innocent scenario produces an unresolved allegation; fewer than 80% of scripted coordination scenarios are referred; or administration exceeds 160 staff-hours.","blocking_evidence":["No facility-specific evidence establishes how often two qualified mathematical jobs require the same indivisible window or whether existing scheduling rules materially fail.","No empirical comparison shows that credit bids track a team's relative time priority rather than team size, risk tolerance, patience, or strategic budget hoarding.","No shadow results establish incremental performance against first-ready, round-robin, lottery, peer review, or Slurm fair-share/QOS baselines.","Stable-team identity rules for new collaborations, project splitting, departures, and cross-institution teams remain unspecified and unvalidated.","Privacy, retention, access, and appeal rules for confidential eligibility descriptions and operational metadata require local authorization.","The proposed collusion screens lack calibrated false-positive and false-negative rates for a small repeated research community.","Facility-specific engineering effort, staffing rates, opportunity costs, round volume, and recurring investigation load are unavailable."],"research_disposition":"PARTNERED_RESEARCH_PROGRAM","world_novelty_boundary":"This evaluation establishes neither world novelty nor absence of earlier implementations. Bellagio, Tycoon, Slurm fair-share, NERSC QOS/reservations, and ACCESS allocation governance define a substantial prior-art boundary. Only the contrastive package for repeated indivisible mathematical-compute windows remains to be differentiated empirically. Patentability, freedom to operate, market size, realized impact, and comprehensive global prior art remain unmeasured.","arm":"COMPLETE_PROPOSAL_PORTFOLIO","candidate_version":0,"controller_recommendation":{"action":"STOP_EMPIRICAL_RESEARCH_NEEDED","repairable":false,"material_progress_observed":true,"progress_targets":["Secure one facility partner authorized to provide synthetic or deidentified workload structure and staff-time observations.","Measure the prevalence and operational cost of genuinely overlapping indivisible-window requests under the current policy.","Complete the preregistered 12-round shadow comparison against first-ready, round-robin, and qualified lottery baselines.","Demonstrate independent clearing reproducibility and zero allocation of unready or unsafe jobs.","Estimate bid-priority consistency, unused-window time, recovered spillovers, concentration, appeals, and administrative burden with uncertainty intervals.","Calibrate anti-collusion screens on labeled innocent and coordinated scenarios, preserving the rule that flags only trigger independent inquiry.","Validate identity, confidentiality, retention, appeal, and rollback rules with the facility's authorized governance and security reviewers.","Replace resource-equivalent cost assumptions with partner-specific engineering, staffing, review, and recurring-operation estimates."],"reason":"Bounded web research verified the broader scarcity problem, credible authorizers, implementation precedents, and substantial prior-art collision. It cannot determine whether the complete auction package improves allocation, whether bids reflect priority, whether small-field collusion screens are tolerable, or whether administration is proportionate. Those questions require partnered shadow testing and facility data; therefore the required terminal recommendation is an empirical-research stop, with repairable set false."},"proposal_index":4}