SKC Mechanism

The SKC (Srinivasan-Karger-Chen) mechanism, from peer-reviewed Harvard research published at ACM EC 2025.

Core Innovation

The SKC mechanism solves the fundamental problem of information elicitation without verification — how to discover truth when ground truth cannot be directly observed. It does this by using a carefully chosen reference agent as a proxy for ground truth, with mathematical guarantees that honest reporting is optimal.

How It Works

1. Market created with question + parameters
2. Agents report sequentially — each posts a bond, observes all previous reports
3. Random stop — after each report: hash % WAD < α?
4. Resolution — last report becomes reference truth (qFinal)
5. Scoring — cross-entropy delta determines payouts

Why It Works

Every agent could be the last one. The last agent has observed all previous reports plus their own private information, making them the most informed participant. Their report naturally aggregates all distributed information in the system.

Theorem (Srinivasan, Karger, Chen): When the reference agent observes k independent informational substitutes, the strategic advantage of deviating from truthful reporting diminishes exponentially: |Δ| ≤ (1-δ)^k.

This constitutes a Perfect Bayesian Equilibrium — no agent benefits from lying, regardless of what other agents do.

Payout Types

Scored agents (first n-k):

payout = max(0, bond + b × [S(qFinal, priceAfter) - S(qFinal, priceBefore)])

Last k agents:

payout = bond + R

The last k agents always profit, which incentivizes late participation even when consensus is mature.

Handling Strategic Behavior

The mechanism addresses three potential manipulation vectors:

1. Misleading the reference agent — Solved by the reference agent having k independent informational substitutes that "wash out" any single agent's influence.

2. Hedging toward the prior — Solved by market termination structure ensuring most agents are observed.

3. Uninformative equilibria — Solved by flat rewards only for final k agents; earlier agents must differentiate to earn scoring rewards.

Comparison to Related Approaches