Chunk 03 · the building blocks
Three question types. Pick by the shape of the answer.
Every question you send is one of these. The temptation is to choose by how you would phrase the question out loud — choose instead by what your code needs to do with the result.
Choice
Which one of these options?
Reach for it when
Exactly one option from a closed set should win — a department, an intent, a document type, a next handler.
Wrong tool when
More than one label can be true at once, or nothing in the set might fit (add an explicit no-match option if so).
"department": {
"type": "choice",
"instructions": "Which team should handle this",
"criteria": {
"billing": "Payment or subscription issues",
"technical": "Bugs or integration problems",
"sales": "Pricing or account questions"
}
}"department": {
"type": "choice",
"choice": "technical",
"confidence": 0.78,
"probabilities": {
"technical": 0.85,
"billing": 0.15,
"sales": 0.0
}
}Chunk 06 · preview
Confidence is the shape of the distribution
For a Choice or Score, confidence summarises how concentrated the probabilities are. A winner always exists — confidence tells you whether the win was decisive. Toggle the three shapes below and watch the same "winning" answer become worth very different amounts of trust.
A near-tie between two acceptable options is harmless. A near-tie between approve and deny is your escalation signal. Set the threshold against your own data and the cost of being wrong — there is no universal number.
Choice · 3 options
Confidence
0.85
90%
escalate
6%
auto_approve
4%
decline
Three traps
Where people misread the numbers
A Noul of 0.5 is not 'medium'
It means the model gives yes and no roughly equal probability. 'Medium intensity' is a Score, with levels you describe.
Normalise before you combine
A Score of 2 on a 4-level rubric and a Score of 2 on a 3-level rubric are not the same signal. Divide by (levels − 1) before applying weights.
Typed is not true
The contract is guaranteed; the judgment is not. Calibration holds across groups of predictions, never as a promise about any single answer.