Done Is Not Done Until It Is Verified

An agent saying "done" is a claim. It is not the end of the job.

The packet should say, up front, what evidence would make the claim believable. A live URL and the status code. A note about what changed. A capture at phone width if the change is visual. When the result comes back, iBrains refuses it if the evidence field is empty. Confidence is not a substitute for the field.

That sounds strict until you remember the alternative. You open five chats and try to reconstruct whether the page was really fetched, whether the check really passed, and whether anyone actually looked. Evidence is how you stop doing that reconstruction by hand. It is also how a second agent, or you, can review the work without scrolling back through a transcript.

Some results still need a person. Publishing, merging, spending money, anything legal, anything a brain's rules mark as gated: those do not finish on an agent's say-so. They land in one approval queue for that brain. You see the packet, the result, and the evidence in the same place. Approve or reject is recorded on the audit trail. The agent does not get to skip the step because the paragraph sounded finished.

The queue is per brain on purpose. A family decision should not sit in the same list as a work merge. When you switch brains, the queue switches with you. Nothing from the other brain is in the list, even if you know the packet id.

iBrains does not pretend to have watched the live web for you in this first version. The evidence is what the agent, or the person carrying a relay link, wrote down. You can read it, and you can refuse the packet. Later, an evidence worker can fetch pages and take its own captures. The rule in front of that is already here: no evidence, no accepted result.

This is a quieter definition of done. The work is finished when the evidence is there and, if a gate was set, when you said yes.