Report card π
After every change, Moss sits the same exam. 96 messages in English and Malay, plus 26 questions for its second brain.
The average of the three scores.
Did the right tools reach the AI?
Did it actually do what you asked?
Did it find the right saved note?
Real exam answers
Tap a message.
The bug the exam caught
Sometimes the AI promised "Added!" without calling the tool. In a normal chat, you'd never notice.
A golden ruleOnly a tool call saves anything. Never just say it.
A safety net in codeA reply that claims "added" with no tool call is sent back once to do it for real.
Clearer routing"grab rm18" is spending, not a setting.
Tuning the second brain
Moss only uses a saved note if it matches the question closely enough. Slide to pick that cut-off.
Speed & cost
Which kinds of tools a message needs.
The middle of all exam replies.
Out of all of them. Smaller prompts cost less.
Never left a needed tool out.
The exam