The claim and the check are separate
The agent's "done" is recorded as a claim. What ran, what changed, and what passed is recorded separately, so you read a result rather than a summary.
Excellent runs beside the agent, checks the work against evidence you can read, and hands back a receipt instead of a confident summary.
curl -fsSL https://excellent.so/install.sh | shThe agent produces a plausible patch faster than you can prove whether it works, so you read the whole diff anyway.
You rerun the same commands by hand after every attempt and still miss the case you did not think to check.
When an approach finally works, nothing records why. The next task starts from scratch.
The agent's "done" is recorded as a claim. What ran, what changed, and what passed is recorded separately, so you read a result rather than a summary.
Returned work names the unmet check and the observed output, so the next attempt starts from evidence instead of a re-prompt.
Attempts, artifacts and receipts are written locally. There is no account to create and nothing to upload before you can see a result.
The prerequisites and the two commands, with the output each one prints.
Wiring Excellent to the agent you already run.
One real task, start to finish, including what a returned result looks like.
The actual record a checked run produces, field by field.
Run one real task and see what held up
curl -fsSL https://excellent.so/install.sh | sh