Evidence, not a verdict from the model

Why self-review runs out#

Ask an agent whether its own change is good and you get a competent review of the diff: the code compiles, the naming is consistent, the edge case is handled. All true, and none of it answers the question you asked. The diff being correct and the feature being usable are different claims, and only one of them can be checked by reading code.

The questions that matter early are coarser than code review anyway: which step does nobody get through, which sentence does nobody understand, which button does nobody find. Those surface in one run of somebody actually using the thing — in minutes rather than weeks, and before the effort is sunk.

What running it looks like#

Evaluation here is an explicit skill, not an always-on review pass. You invoke `$eval` when you want evidence, and the simulation runs locally: simulated users go through the real build — the one the agent just produced — rather than through a description of it.

What comes back is not a score. Screenshots, traces, commands and reports land in the artifacts panel, anchored to the host and the run that produced them, previewable beside the session that made it. The working rule is blunt: a conclusion you cannot trace back to a run is treated as no conclusion.

Where the evidence goes#

The point of gathering it is that it changes what happens next: the evidence comes back as the next batch of tasks. That is what makes it a loop rather than a report — the thing that failed becomes the thing an agent works on, without a human retyping the finding as a ticket.

Evidence also stays with the project it belongs to, so a second round can be compared against the first instead of replacing it. Round two answering "did the fix work" requires round one to still be there, which sounds obvious and is the part most setups lose.

One boundary worth naming: simulated users are not a substitute for real ones, and nothing here claims they predict what your customers will do. They are a fast, repeatable way to find the failures that are obvious in hindsight — the ones you would be embarrassed to have shipped, and would otherwise have shipped.