Confidence ≠ Authority

TypeSafe AI provides a useful example of the first. The founding team includes Diogo Almeida, who co-invented RLHF, the training method behind ChatGPT and GPT-4, and their diagnosis of where that method leads is sharp: today's models are trained to be "a helpful, articulate, pleasant assistant," which is the right goal when a person is on the other end of it, and the wrong one for software meant to run unattended, because it produces AI that "requires humans in the loop instead of running in the background." Their answer is a new model called Jev, which replaces a free-text answer with a calibrated confidence score that software can consume directly.

It's an interesting architectural bet. Nothing here is a criticism of it. But it's worth being careful about what a calibrated confidence score is actually answering, because the vocabulary around this launch — auditable, runs unattended, inspect it, constrain it — sits close enough to the vocabulary of institutional authority that the two get read as the same claim. They aren't, and the reason isn't a matter of degree.

A system about to act needs answers to two different questions. The first is epistemic: what should happen, and how confident are we? The second is institutional: are we permitted to make it happen?

Confidence concerns the quality of a judgment. Authority concerns the legitimacy of an action.

The first is checked against outcomes: was the belief right, over time, across cases. The second is checked against a rule: did whoever's asking have the standing to do this, granted by whoever had the standing to grant it. Nowhere on TypeSafe's site is Jev framed as answering the second question, and it shouldn't be read as one that does. It's a bet on the model itself being a dependable primitive, a better, checkable answer to "what should happen."

The two questions vary independently, which is the whole point. A new hire on their first day, no track record anyone could vouch for, can still be fully authorized to approve a $50 expense report, because the org chart already settled it; nobody needed to be confident in their judgment first. Run it the other way: the best-calibrated forecaster alive doesn't get to move money out of an account that isn't theirs just because they're certain it's the right call. The money was never theirs to move, and no amount of being right about the outcome changes that.

A model can be highly confident and still have no authority to act. An actor can have authority without being confident. Confidence and authority answer different questions.

Something falls out of that, and it's checkable, not just asserted: if the two are genuinely independent, an authority verdict has to come out the same way regardless of what proposed the action: a fixed rule, a calibrated model, a different calibrated model, a person typing into a form. If replacing the proposer changes the authorization verdict, the system has allowed epistemic judgment and institutional authority to collapse into one another.

Run that test on Jev. Swap the checkpoint behind it (a retrain, a new version, same architecture, different weights) and the confidence number moves, because it was never a fact about the world. It was a fact about that model, at that moment, on that input. Ask twice, six months apart, against two different versions, and there's no reason to expect the same number back, because nothing ever claimed there would be.

Now do the same swap on an action governed by a sealed authority artifact: reserve, settle, release, whatever the shape of it. Replace what proposed the action with a rule, a model, a different model, or a person: the authorization verdict does not change merely because the proposer changed, because it was never asking any of them anything. It checked a sealed, versioned artifact against a set of facts and recorded exactly which version, which facts, which verdict. Pull that record up two years later and the authorization decision can be independently reconstructed from the recorded artifact, facts, and verdict, because nothing in the chain was ever a live inference to begin with.

None of this makes the two projects competitors. A calibrated decision engine can sit behind an authority check: it can propose an action, along with whatever confidence or uncertainty the model attaches to it, and can() can independently determine whether that action is authorized. The proposer can be a model, a rules engine, or a person. The authority layer does not need to trust the source of the proposal in order to evaluate its authority.

AI can provide the judgment; authority determines whether that judgment may become institutional action.

What's genuinely worth naming is that "calibrated" and "authorized" are close enough as words that they get used as if they were the same claim, and once that happens in a pitch deck or a procurement conversation, it's hard to unwind. A calibrated model answering what should happen is not the same claim as an authority artifact proving what was permitted. As AI systems become increasingly capable of producing machine-consumable judgments, the authority layer becomes more important, not less. You need both. Not one instead of the other, both, doing different jobs, neither one able to cover for the other.

This note is an independent analysis. It is not affiliated with, endorsed by, or reviewed by TypeSafe AI, and no claim here should be read as a statement of their intent beyond what is quoted directly from their own published materials.