Part 5 of 12 · 2 min read
Say uncertainty only when it's calibrated, ask when it's close
A confident-sounding sentence is only honest if the probability behind it has actually been checked against reality.
Think of a weather forecaster who says "80 percent chance of rain" and is actually right about 80 percent of the time they say that, versus one who says it because it sounds appropriately hedged.
In plain words: wording like "I'm fairly sure" is only meaningful if the confidence behind it is calibrated. If not, it's just decoration. When two readings of a request are genuinely close, asking a focused question beats guessing either way.
A worked example. With three possible readings of a request and the agent always acting on its top-scoring guess, it's wrong 13.3 percent of the time and asks 0 percent. Asking whenever the top two readings are within 0.1 of each other cuts wrong actions to 4.9 percent, at the cost of asking 22 percent of the time. Widening that margin to 0.3 brings wrong actions down to 0.3 percent, but now 65 percent of requests trigger a question. The margin is a direct dial between how often the agent guesses wrong and how often it interrupts to ask.
The key lines. From code/mini-ui/uncertainty.py:
# from code/mini-ui/uncertainty.py
def phrase(p, sources=0):
"""Map a (calibrated) probability to wording. Without calibration these words mislead."""
base = ("I could not confirm this" if p < 0.4 else "I think this is likely" if p < 0.7 else "I am fairly sure" if p < 0.9 else "This is well supported")
Read it as: the comment says outright that this mapping only means what it says if p is calibrated. The function will happily produce confident-sounding text from an uncalibrated number. That's precisely the failure mode it's warning against.
Calibration, provenance
Calibration is how well a stated confidence matches how often the agent is actually right at that confidence level. Provenance is showing where an output came from — which sources, tools or agents produced it.
Example: phrase()'s docstring warning above is itself the calibration caveat. The function can't verify its own input is calibrated, so provenance (showing the sources behind a claim) is often the more honest substitute for a confidence adjective.