A Skill Provenance and Royalty Marketplace for AI Agents
AI agent runtimes increasingly generate reusable skills, but those skills are trapped inside the runtime that produced them — no portable record of authorship, no way to verify what a consumer is about to run, and no mechanism to compensate authors when their work is reused. This paper specifies the manifest format, the on-chain settlement contract and its conservation invariants, the royalty model, the threat model, and the off-chain architecture, and argues that a tokenless, per-invocation design is the right primitive for an emerging skill economy.
The skill-economy thesis and why the value stays trapped today.
Hermes, OpenClaude, MCP, and existing registries — what they do and don't provide.
Skill manifests, DID identity, content addressing, and canonical signing.
AtriumRegistry design, conservation invariants, and the pull-payment pattern.
Per-invocation pricing, the royalty cascade, and the tokenless thesis.
Threat model, mitigations, and open issues.
Base rationale, indexer + interface architecture, and governance.
Atrium vs npm, PyPI, Hugging Face Hub, the GPT Store, and agentskills.io.
Encrypted bodies, ZK benchmark proofs, and cross-chain reputation.
The whitepaper source (Markdown + diagrams) lives in the repository; the PDF is built reproducibly in CI.