A compensation, attribution, and metering layer for authored AI Skills — "Carta for AI work artifacts." Creators keep a durable economic claim each time others use their Skill, instead of handing the value over once under work-for-hire.
This is a research/spike repository: design documents plus executable evidence, not a product. Every claim below is labeled measured, modeled, or hypothesis.
The manifesto is live — and it is a working protocol endpoint:
neverhandedover.com (also at
skillassetprotocol.com; source in site/).
On 2026-07-12, one wallet paid per model call and per hosted-skill invocation over x402 (Base Sepolia, real facilitator, real USDC), landing both asset classes in one attributed session ledger:
claude/plan $0.041 · skill $0.25 → creator $0.24375 / treasury $0.00625
On-chain balances reconciled to the cent (Wielder 20 → 19.668 USDC; sellers received exactly
0.332). Measured x402 payment overhead across 48 settled calls (2026-07-15, two model providers):
p50 731 ms / p95 1206 ms per paid call — facilitator verify/settle is nearly all of it
(p50 729 ms); the 402 roundtrip + EIP-3009 signature add ~2 ms.
Details and txHashes: spikes/pi-wielder/README.md.
All four proofs run with no API keys, no network payments, and no wallet. Where a
package-lock.json exists, npm ci is the reproducible choice.
| Where | Commands | Proves |
|---|---|---|
spikes/pi-wielder |
npm install && npm run e2e |
20 checks: 402-first on every leg, no skill-content leak, replay rejection, exact split match |
spikes/clone-economics |
npm run e2e (no install) |
97 checks: deterministic clone-distillation harness, byte-identical reruns |
prototype |
node spike-fork-economics.mjs (no install) |
64 invariants: fork/royalty economics on the settlement engine |
phase0 |
npm install && npm test |
18 tests: Story provenance registration against injected fakes |
CONTEXT.md— the ubiquitous language: Skill, Creator, Wielder, Beneficiary, Collar, Invocation, Derivative, Royalty claim.docs/adr/— 8 decision records, including 0007 (the closed-mode compensation layer is the terminal product) and 0008 (the Wielder is a wallet, not a harness).spikes/— pi-wielder (one wallet, two asset classes, unified ledger) and clone-economics (how cheaply can N paid outputs be distilled into a clone?).prototype/— the settlement engine (pure logic) plus fork-economics and CMA-latency spikes.phase0/— Story Protocol provenance: register a Skill as an IP Asset and declare Derivatives (Aeneid testnet only)..claude/skills/and.agents/skills/— the bundled example skill (optimizing-claude-code-prompts) is intentionally public. The protocol's claim is that a Wielder never receives the skill at runtime — output only — not that the skill is secret. The spikes host and meter this exact skill behind the Collar.
The repo's discipline is to label every number:
- Measured (real network, n=48 settled calls + a live pi session): the runs above — real 402 → sign → settle, real USDC per call, splits credited by the settlement engine, wallet reconciled on-chain to the cent; two failure modes documented (pay-then-fail, settled-but-rejected).
- Measured (n=3, one model): hosted-agent cold start — first answer token p50 ~2.5 s; pay-then-run-async reads as usable on top of the ~0.8 s testnet payment gate.
- Measured (N=6, small fixtures): the clone attack failed on fidelity — all 6 held-out cases failed critical gates — but modeled break-even is 8 invocations if a clone ever passes. Cost is no defense; fidelity was. High-N behavior is unknown.
- Modeled (deterministic arithmetic, not observed behavior): education-mode flow-through is dominated by free re-authoring — every school-paying royalty rate loses at parity, so Education mode is deferred.
- Validated arithmetically: settlement splits, multi-level derivative flow-through, and payment gating ("no credential, no run") via the engine's invariants.
Unvalidated: that employers will buy this. Design-partner interviews remain the open step.
Apache-2.0 — see LICENSE. Copyright 2026 Antony Zaki (Aznatkoiny).