AI usage
Leash was built with Claude Code (Claude Sonnet 5.5) working alongside the author, over three days.
What the AI did
- Read QoreChain's documentation and the organisers' sample repository, and summarised the precompile interfaces, window rules and spending-rule behaviour.
- Proposed the design (the "classical key can only stop" rule) and wrote the first versions of
LeashGuard.sol, its Foundry tests, the policy compiler, the attack suite and the testnet pipeline. - Ran the tests, the local drill and the testnet run, and wrote this documentation.
What the author decided
- Which track and which idea to build, the scope for three days, and every action on accounts and funds. The author signed in to the faucet; the AI claimed tokens for a throwaway testnet address only.
- Registration for the fest, and publishing.
Where the AI was wrong, and how it was caught
- A test used
vm.expectRevertand then called a helper that made its own external call, so three tests failed for the wrong reason. Found by running the tests; fixed by computing the signatures first. - An edit to the testnet script dropped the last lines, so a command ran silently with exit code 0. Found because the output was empty; fixed and rerun.
-
The first full testnet run stopped on a transient network error after deploying and arming the guard. The drill was made resumable so the guard was not redeployed.
-
While building v2, one of the contract mutation checks (leaving an approval open after a call) survived the first test suite, because the mock router consumed the whole approval. Found by running the mutation; fixed by adding a router that pulls less than approved. A test token with no code made several happy-path tests fail for the wrong reason, found by reading the revert.
- The first live testnet run of the swap drill showed a FAIL row: the router itself refused an under-delivering swap before the guard's own rate-floor check could, so the row did not test what its name said. Fixed by having the simulated attacker set the minimum output to zero.
What the AI did for v2
- Designed LeashGuardV2 (routes, measured value flows, rate floors), the adapters, the plain-English flow grammar, the monitor rules and the MCP execution tools, and ran them: 25 more contract tests, six mutation checks, a local drill, an MCP end-to-end test and a live testnet run.
- Research on QoreChain's bridge, AMM, payment features and agent harness formats was delegated to a Claude Opus sub-agent working read-only from public documentation; its findings (no live bridge route, native AMM not callable from a contract) shaped what is and is not claimed.
- The author decided to extend the project, and every action on accounts and funds.
How the work is verified
The AI's claims are not trusted. They are checked by 21 contract tests (including a fuzz test), 10 compiler tests, a local end-to-end drill, a mutation check that breaks the contract three ways and confirms the drill fails each time, and a run on the real testnet with a genuine ML-DSA-87 signature.
Demo video
- Voiceover: generated locally with Kokoro-82M (Apache-2.0, ONNX build from
onnx-community/Kokoro-82M-v1.0-ONNX, voiceaf_heart). The script is the author's design, written with the AI. No real person's voice is imitated. - Music and sound effects: original, synthesized locally with numpy (pad, bass, arpeggio, ticks, whooshes, stamps, chimes). No samples and no third-party music.
- Motion graphics: built as a HyperFrames project (HTML rendered to video). Every terminal output on screen is copied by script from the real command output, unchanged.
- How it was made: the video and the agent research were delegated to Claude Opus sub-agents working under my instructions. I checked the finished file myself (format, loudness, a spread of frames) before publishing. Neither the voice nor the music has been verified by ear by the AI.
No model at runtime
The plain-English front end is a small deterministic grammar, not a language model. A model may help a person draft the sentence, but the compiler decides: anything it cannot parse exactly is refused.