Spark
Testnet

Trace

Trace verifies whether a coding agent actually completed its task instead of simply trusting its claim. A developer defines and locks the task’s acceptance criteria in Monad, then submits a specific GitHub commit for verification. An isolated worker clones that commit and runs the real build, tests, linting, type checks, static analysis, deployment probes, and any required Monad transaction checks. A criterion passes only when Trace can produce affirmative execution evidence; a check that never ran cannot be treated as successful. Trace also detects common forms of fabricated completion, including hardcoded transaction hashes, stubbed results, disabled tests, and meaningless assertions. Each verification attempt produces an off-chain evidence report, while its hash, commit reference, verdict, and attempt history are recorded through the TraceRegistry contract on Monad. Anyone can then open the public receipt to inspect what was requested, what was tested, what passed, what failed, and why. We built Trace as a Next.js application with wallet authentication, task creation, live verification progress, attempt history, and public receipts; an isolated Docker worker that executes submitted repositories under strict resource and network controls; and the TraceRegistry contract deployed on Monad testnet. The central lesson was that running checks is not enough; the system must also prove that those checks ran and preserve the evidence. Monad makes it practical to create a tamper-resistant receipt for every attempt. At the same time, Trace’s evidence-first model prevents the verification pipeline from treating assumptions, missing checks, or confident agent output as proof. Running the complete workflow against Trace itself also exposed several genuine implementation issues before release.

讨论

登录后即可评论此项目。

还没有评论,快来分享你的看法吧。