AI saying “done” does not make the work done.
Bellfly checks the actual result against what you asked for. When something is wrong, it identifies the problem, gets it repaired and checks again before bringing the result back for your review.
Find the gap. Fix it. Check again.
A report compares three options, but one conclusion has no source. “Report written” is not enough to call that finished.
A result comes back
The agent returns the report together with the sources it used.
Bellfly checks it
It compares the output with the required questions, citations and format.
A specific problem is found
The unsupported conclusion is kept visible, with the passage that needs attention.
The agent repairs it
It adds a valid source, narrows the statement or records that the question is unanswered.
Bellfly checks the change
It checks the revised passage and anything the repair may have affected.
You decide whether to accept
The result arrives with the checks, repairs and remaining questions together.
An unanswered question stays visible.
An unavailable check, missing source or incomplete review is not a pass. A passing check tells you what was checked; it does not promise that an AI result can never be wrong.