Somewhere in a lot of production agent setups there’s a field called confidence. A classifier emits it, or the agent is asked to include it, or a framework attaches it to a tool choice. Then teams build on it. Auto-approve above 0.9. Send to a person below 0.6. Chart the average over time and treat a drop as a warning.
All of that assumes the field answers “should I believe this decision?” I think it usually answers a smaller question, and which smaller question depends on where the number came from. This piece takes the number apart.
Four things that get called confidence
They share a field because they share a shape, a value between zero and one. They are not the same object.
| Kind | Where it comes from | What it tells you | What it can't tell you |
|------------------|----------------------------------------------------------|--------------------------------------------------|----------------------------------------------------------|
| **Stated** | The model, asked to rate its own certainty in its output | How the model reports its certainty, under that prompt | Whether that report tracks outcomes for this task and protocol |
| **Token-level** | Probabilities assigned during generation, where exposed | How probable the generated tokens were under the model | Whether that signal maps reliably to decision correctness here |
| **Stability** | Reruns and small meaning-preserving variations | Whether the decision holds up when asked again | Whether a decision that holds up was right |
| **Track record** | Past outcomes at this decision point, joined to a signal | How often this kind of signal has held up here | Anything about a decision point with no recorded outcomes |The first three are readings taken from a single decision, and each is useful inside its boundary. A stated confidence is generated text, produced by the same process that produced the answer. It can carry real information in some tasks, and how much depends heavily on how it was elicited, so the prompt and protocol are part of the measurement. Token probabilities describe how likely the generated text was under the model. Aggregated, they can say something about whether an answer is right, but a probability over tokens is not automatically a calibrated probability that the decision was correct, and not every provider exposes them anyway. Stability I covered in the triage piece: it catches fragile decisions and says nothing about a system that is consistently wrong in the same direction. Margin, how close a call was between alternatives, which I used in the decisions piece, is estimated the same way, by rerunning and probing, so it has the same boundary.
None of those three touches an outcome. The fourth does, and it’s the only entry that isn’t a property of one decision. A track record is a relationship: when this signal read this high, at this decision point, how often did the outcome hold up? That is what calibration means, and classical ML has practiced it for a long time wherever labeled data exists. Most agent decisions in production don’t come with labels, which is most of the problem.
So here is the title’s claim, stated narrowly. A confidence value becomes interpretable as a claim about correctness only when you know how that signal relates to outcomes in a relevant population. That relationship belongs to a decision point and a population, and it can’t be validated from a single instance. Read alone, 0.92 is a number with no unit.
The same 0.9, two meanings
A composite, to show the shape rather than report a case.
A support team runs an agent with two decision points that both emit a confidence field. One decides whether a refund request is within policy. The other decides whether a ticket goes to a human specialist. Someone wrote a single rule for the pipeline: above 0.9, the agent acts without review.
The refund path has been live long enough to have a history. A reviewer has looked at some of its decisions and corrected some, and those corrections sit in a spreadsheet next to the scores. When the refund path says 0.9, the team has at least a rough sense of what that has meant.
The escalation path was added after a prompt change and has no such history. It emits 0.9 just as often. Same field, same threshold, same dashboard column, and nothing marks it as different. The rule can’t tell the two apart because the number is identical, and one of them rests on a record while the other rests on hope.
Then the refund path moves to a newer model version. The field still looks normal. Whether its scores relate to outcomes the way they used to is now an open question, since the threshold stayed put while what it admits may have shifted. Nothing errors. The only thing that could answer it is the outcomes, and nobody is reading them, because the number looks fine.
The investigator has a confidence too
A second confidence gets blended in with the first: the one attached to the explanation. When a team reconstructs why a decision happened, the reconstruction carries its own certainty, and that depends on where the evidence came from. Something captured independently at the time is a different grade of evidence from something the application reported about itself, and both differ from something inferred afterward. An agent’s own account of why it chose what it chose is generated text as well. It can read as a convincing narrative without being a faithful record of what drove the choice.
Here’s where it goes wrong in practice. A reviewer opens a decision and sees a high confidence score next to a fluent explanation, and experiences two confirmations. They aren’t independent if the explanation was built from the same self-report as the score. Two readings that share a source are one reading, counted twice.
A record of a decision therefore has to keep several questions apart: how sure the agent was, how sure we are about our account of the agent, and what eventually happened, when anything did.
What has to exist
The research question on this channel is always what object would have to exist for the thing to work. Here: for a confidence reading to be interpretable, what has to travel with it?
Five things, at minimum. The kind of confidence it is, from the table above. How it was derived, including whether it was measured or estimated. Which decision point, model and configuration produced it. The track record of that kind of signal at that decision point. And how many outcomes that track record rests on, which is the part most likely to be missing, since a record built on a handful of cases deserves less trust than one built on thousands. A reading carrying all five can be interrogated. A bare number can only be believed or ignored.
Two cautions, because this is easy to over-claim.
Outcomes are evidence, not truth. A reviewer’s correction records that someone with context disagreed with the decision. Reviewers can be inconsistent, overly cautious, or applying a preference nobody wrote down. And the absence of a correction can mean the decision was fine or that nobody looked. So a track record is itself an estimate, and it has to be reported with its coverage: how many of the decisions at this point have any recorded outcome at all.
And “unknown” has to be a legitimate value. For a new decision point, or one whose model just changed, the honest status of the track record is that there isn’t one yet. A record that says so tells the next person where the uncertainty actually sits. A confidence field that keeps emitting 0.9 regardless is describing a signal, and the team is reading it as a fact about the world.
The research conclusion
Confidence isn’t an intrinsic property of a decision. It is a reading produced by some signal, and whether that reading deserves belief depends on how the signal has behaved against relevant outcomes, a history that in most agent systems nobody kept. Until the history exists, a column called confidence is better labeled with whatever it actually is: a stated number, a token statistic, a stability reading. Each one is a hypothesis about correctness that hasn’t yet been tested against outcomes.
I’m building Nalyqor on the belief that a decision record has to hold the reading, what kind of reading it is, and what it has and hasn’t been checked against.

