AI Engineering from Scratch（英文原版）

Given a multi-agent scenario that needs incentive alignment (open network, heterogeneous operators, tokenized rewards, or reputation-based routing), design the economy layer.

Produce:

1. Identity layer. W3C DIDs for portable identity, or platform-internal IDs if the system is closed. Justify by openness of the network.
2. Credit attribution. Equal split, last-contributor-takes-all, contribution-weighted, Shapley (exact or sampled), or none (pay-per-call). Recommend Shapley sampling when coalitions matter; equal split for simple pay-per-call.
3. Payment mechanism. Second-price auction for task assignment (truthful under monotone aggregation), first-price for speed, posted-price for simplicity. Escrow if payoffs depend on quality verification.
4. Reputation rule. Exponential decay constant, slashing policy, minimum floor, maximum ceiling. Reputation reads cheaply (O(1) for routing) and writes after verification.
5. Verification. Who verifies contribution quality? A separate agent, human review, on-chain oracles, cross-agent attestation? Without verification, credit attribution is guesswork.
6. Sybil mitigation. What stops one operator spinning up N fake agents? Reputation cost-to-forge, proof-of-humanity attestation, stake requirement, or capped reputation per DID.
7. Legal and jurisdictional check. Token-denominated payments touch financial regulation in most jurisdictions. If this applies, flag it and recommend legal review.

Hard rejects:

• Any design without verification of contribution quality. Credit will accrue to fastest-but-wrongest agents.
• Reputation without decay. Stale reputation rewards agents who did good work years ago but are now broken.
• Shapley exact computation for N > 6. Computation time grows as N!; sample instead.
• Second-price auctions where the aggregation function is not monotone. Truthfulness does not hold.
• Token distribution without a regulatory check. Many jurisdictions treat this as securities activity.

Refusal rules:

• If the system is fully internal (one company, one operator), recommend simpler allocation (managers assign, metrics are internal). Economic mechanisms are overkill.
• If there is no way to verify contribution quality, recommend adding verification before economy design. Without it, the economy is ornamental.
• If the user wants a tokenized system but has no legal team, flag the risk and recommend starting with reputation (non-token).

Output: a two-page brief. Start with a one-sentence summary ("Reputation-only system with DIDs, Shapley-sampled credit on 3-agent pipelines, second-price auction for slot assignment, slashing on verification failure."), then the seven sections above. End with a 30-day pilot plan: warmup phase, verification pipeline setup, reputation-weighted rollout, audit schedule.