COMSCI/ECON 206 · Team FP8

Auctions with Risk Reserves for Scarce AI Research

When portfolio managers compete for four premium AI research slots, when does one verified-risk reserve justify the private value it displaces?

20 portfolio managers 4 AI research slots 3 allocation mechanisms Browser-only artifact

The asset-manager decision

During earnings season, 20 risk-neutral portfolio managers compete for four identical premium AI research slots. Each PM privately knows the value of access. The risk team freezes a public risk score before bidding.

1

FCFS

The first four requests in an exogenous random arrival order receive slots. Payment is zero.

Non-price baseline
2

Pure first-price auction

All four slots go to the highest sealed bids. Each winner pays their own bid.

Private-value benchmark
3

Hybrid

The highest-risk request with r ≥ 60 receives one free slot and exits. Three remaining slots are auctioned. If no request qualifies, all four return to auction. Ties in risk scores or bids are broken by fixed PM ID.

Verified-risk priority

Fixed baseline

  • N=20, K=4, M=100
  • vᵢ ∈ [0,100]
  • rᵢ = 100EᵢSᵢ, frozen before bidding
  • Assumed synthetic bid bᵢ = 0.8vᵢ
  • Reserve size R=1; threshold r*=60

Shared research question

Can a hybrid mechanism that combines auction-based allocation with reserved capacity for objectively high-risk research requests improve the allocation of scarce AI research capacity relative to a pure first-price auction?

Two different bid rules: b=0.8v is an assumed multi-slot simulation rule, not an equilibrium result. The exact b(v)=0.75v BNE applies only to the separate four-bidder, single-slot behavioral benchmark.
Decision rule: For hybrid H and pure first-price auction P, let ΔC = C(H) − C(P) and L = V(P) − V(H). The hybrid improves the stated firm index exactly when λΔC > L.

Three-mechanism lab

Compare FCFS, pure first-price auction, and hybrid allocation on the same 20 synthetic requests. The page recalculates r=100ES and the assumed simulation bid b=0.8v. Lower fixed PM ID breaks exact ties in arrivals, risk scores, and bids. This browser example is separate from the proposal's reported 1,000-batch run.

K=4 · R=1 · r*=60

PM IDArrivalPrivate value vExposure ESeverity S
Awaiting actual peer-play observations

Fixed-payoff framing round

This is a separate one-slot first-price auction with four symmetric bidders. It tests whether people bid like the rational first-price benchmark, especially whether loss framing moves bids away from it. It does not directly test Hybrid versus Pure. Under risk neutrality and independent uniform private values, the benchmark is b(v)=3v/4.

Experimental validity: Gain and Loss treatments hold private value, opponent-value distribution, winning rule, payment rule, and numerical payoff constant. Only the wording changes: same economic payoff + different framing.
Exploratory peer play

Structured four-person round

  1. Assign all four participants the same frame.
  2. Give each person a private value and collect sealed bids before revealing any value.
  3. Enter values and bids after the reveal, then settle the round.
  4. Compare each bid with 0.75v and record one observation that could change the model or interface.
PM IDPrivate value vSubmitted bid b

Aggregate actual peer-play evidence

Repeat the four-person round under assigned Gain and Loss frames, then append each completed round or enter observations below. Use anonymous record IDs rather than names. The summary reports descriptive exploratory evidence, not population-level evidence.

Anonymous record IDAssigned framePrivate value vSubmitted bid b

Evidence and interpretation

The completed computational evidence uses 1,000 independent matched batches with seed 20260926, N=20, K=4, R=1, r*=60, λ=1, and assumed bids b=0.8v.

RuleVCWPayments PPM utility U
Pure first-price auction353.08101.35454.43282.4670.62
Hybrid326.26137.95464.20226.9899.28
FCFS199.45100.43299.880.00199.45
+36.59Hybrid risk-coverage gain versus pure first-price auction
−26.82Hybrid private-value difference versus pure first-price auction
+9.78Mean ΔW at λ=1; Monte Carlo SE 0.89
Scope: Mean ΔW is −8.52 at λ=0.5, and the ratio-of-means break-even weight is 0.733. Only 43.8% of primary batches strictly improve W at λ=1. Average advantage ≠ state-by-state dominance. Under negative value–risk rank alignment, mean ΔW is −3.24 at λ=1. Risk coverage is not avoided financial loss.

Behavioral decision rule

Actual peer-play evidence has not yet been entered in this public artifact. Use the aggregate log in Peer play to report sample size, Gain/Loss mean b/v, distance from the 0.75 benchmark, and overbid/underbid counts.

ObservationResponse
Loss-frame bids are systematically higher with fixed payoffsQualify: test whether the change affects payments or allocation.
Bids remain close to 0.75v in both framesRetain: keep the single-slot benchmark as a working reference with its assumptions stated.
Both frames depart materially from the benchmarkRevise / investigate: reconsider the assumed bid rule or decision interface.
Interdisciplinary integration: The behavioral task estimates a human bidding pattern relative to 0.75v. That pattern can then replace or supplement the assumed 0.8v rule in a labeled GitHub sensitivity rerun of Pure first-price versus Hybrid. The mechanism conclusion is revised only after the rerun.
Privacy boundary: This page makes no external network requests and stores no server-side research dataset. Do not enter personal names, portfolio positions, firm data, or other sensitive information.

Methods and sources

Three connected lenses

  • Economics defines incentives and the value–risk trade-off.
  • Computation compares matched allocations and sensitivity conditions.
  • Behavioral science probes whether fixed-payoff framing changes bids enough to affect the comparison.

Baseline limits

  • Synthetic private values and public fixed risk scores
  • No common financial values or future rounds
  • No strategic score manipulation or endogenous arrival
  • The multi-slot 0.8v bid rule is assumed, not solved as equilibrium
  • The exact 0.75v BNE belongs only to the separate four-bidder single-slot benchmark
  • Lower fixed PM ID breaks exact ties in arrivals, risk scores, and bids

Proposal sources

  1. Harsanyi (1968), Games with Incomplete Information Played by Bayesian Players, II. DOI
  2. Pathak et al. (2023), Fair Allocation of Vaccines, Ventilators and Antiviral Treatments. arXiv
  3. Sheng et al. (2026), Generative AI and Asset Management. DOI
  4. Sivan et al. (2026), Quota Marketplace: Dynamic Pricing for Efficient Allocation of ML Training Resources. arXiv
  5. Stokely et al. (2009), Using a Market Economy to Provision Compute Resources Across Planet-wide Clusters. Google Research
  6. Vickrey (1961), Counterspeculation, Auctions, and Competitive Sealed Tenders. DOI