Unit 06.04: Why a fluent guess is the expensive failure
Not all wrong outcomes cost the same, and threshold decisions should follow the costs rather than the accuracy number.
Four outcomes, wildly asymmetric costs
A system can refuse correctly, refuse unnecessarily, answer correctly, or answer fluently and wrongly. Three of those are cheap. One is not.
The table below lays out all four with their consequences.
outcome consequence cost
refuses correctly user asks a colleague low
refuses when it could have answered user is mildly annoyed low
answers correctly user proceeds none
answers fluently and wrongly user acts on it and finds out later high
The costs are not symmetric. An unnecessary refusal is visible and recoverable
-- the user asks someone else. A fluent wrong answer is invisible: it is acted
on, and the error surfaces later, somewhere else, attributed to something else.
Tune the threshold towards refusal, and record that you did.
An unnecessary refusal is visible and recoverable: the user is mildly annoyed and asks a colleague. The loss is one interaction. A fluent wrong answer is invisible - it is acted on, and the error surfaces later, somewhere else, usually attributed to something other than the assistant.
That asymmetry has a direct consequence for the retrieval threshold. Raising it produces more refusals and fewer confident errors, which trades a cheap failure for an expensive one. Make that trade knowingly and record that you made it.
The mistake this prevents
The mistake is optimising for answer rate. "The assistant answers 94% of questions" is a metric that rewards guessing, and it will be reported to you as progress. Pair it with the fluent-error rate from your eval set, or it means nothing.
Takeaway
Tune the threshold towards refusal, deliberately, and write down that you did. The costs of the four outcomes are not symmetric, so a threshold chosen to maximise accuracy is a threshold chosen for the wrong objective.
