Evidence, not a second opinion
Checks are things that run: git state, tests, commands. A claim with no observation behind it is reported as unchecked rather than assumed.
Excellent records what each agent run did, checks it against evidence rather than a model's opinion, and tests a change to the prompt, model or tools on held-out work before it becomes the default.
curl -fsSL https://excellent.so/install.sh | shPrompt changes, model swaps and retry rules happen ad hoc, so nobody can say whether the setup improved or just changed.
A tool that asks a model whether the model's work is good is not an independent check, and you know it.
A passing check in one repository does not tell you the same setup holds for the next class of task.
Checks are things that run: git state, tests, commands. A claim with no observation behind it is reported as unchecked rather than assumed.
A receipt binds the work item, attempt, versions, checks and result into a record that can be read and verified after the session is gone.
A candidate instruction, model or tool change runs against held-out tasks, and the comparison decides whether it is promoted.
The rules behind verified, returned and unchecked, including what Excellent will not conclude.
The receipt schema and how one is checked offline.
What has to be true before a candidate setup replaces the default, and how it is reverted.
Excellent next to the other tools in this category, with dated prices and sourced claims.
Run one real task and see what held up
curl -fsSL https://excellent.so/install.sh | sh