Why you
can trust
a bot
The usual complaint about AI is “it said it did the work and it did not”. VibeBot has five defences against exactly that, and all of them are visible to you rather than buried in promises.
Five
defences
A measure
Every turn a bot takes has an outcome: work done, a question asked, or just talk. The quality page shows this in numbers, so you can see where the swarm is stalling.
Acceptance
The auditor — a separate model — does not just tick a box: it sends the work back until it is actually done. You see the result after its “good enough”.
Claims against the log
“I did it” is checked against the bot's own action log. If they disagree, the bot gets the discrepancy back and finishes the job instead of closing it.
A second attempt
A temporary tool failure does not kill the task: the bot reads the error, waits and tries again. Only genuinely broken things stay broken.
An exam
Before a release the swarm sits 26 machine-checked tasks. Fail and the version does not ship. That guard is aimed at us, not at you.
Version history
Projects show a before/after for every change and roll back with one button. A bot's mistake costs you a click rather than an evening.
You decide what
gets asked
Ask about everything
For someone else's project or a first run: the bot checks every step that changes state.
Ask about what matters
The normal mode: routine runs itself, while publishing, paying and deleting wait for your yes.
Autonomous
For overnight work: the bot does not stop on questions and shows you in the morning what it did and where it stopped.