Every decision model on the market can be made to abstain, and every one of them can be made conformal, because conformal prediction is a wrapper you can put around any classifier. That is exactly the point. As of October 2026, across every hosted and open-weight decision model we can find in the public record, ours is the only one where the conformal shortlist and the right to say nothing are built into the model itself rather than bolted on afterwards.
That sentence is checkable, so here is the check.
What the public record shows
The hosted flagship in this category does not offer conformal sets. What exists instead is a third-party project that wraps its API in split conformal prediction from the outside, and that project's own README is the best argument for building it in: across 860 answers, the hosted model returned only 94 distinct top probabilities, and with 460 calibration queries the tightest achievable bound floored near 2%, with a 1% risk target unreachable. A wrapper can only work with the numbers the API lets out, and rounded probabilities starve it.
The open alternatives, our own laya included, ship calibrated probabilities and leave the thresholding to your application code. The community lists that track this space describe exactly one model with conformal answer sets and an explicit abstain flag in the response itself, and it is vegaml.
What built in actually means
One forward pass returns four things no wrapper can retrofit faithfully: a typed answer, a probability whose temperature was set per decision rather than per model, a conformal shortlist built against a risk level you chose, and two flags. abstain fires when nothing clears a floor fitted on held-out data. unbound fires when the record looks like nothing the model knows, which is the question an out-of-distribution input should raise and a confidence score never does.
The distinction is not academic. A wrapper sees a finished probability and has to take it at face value. The engine sees the state the decision ended in, and the calibration is fitted to that state. The coverage guarantee itself still comes from calibration data, the way conformal prediction always works. What is different is that nothing between the model and the guarantee is rounding, truncating or summarising the evidence.
Why this is hard to copy
Because it is the part a competitor cannot add with a patch. Typed outputs are a schema. Calibrated probabilities are a fine-tune. Prediction sets with a stated risk level, honoured by an engine that was trained to express its own uncertainty, are an architecture, and the way we got there is written out in full in the RLCD method page.
It is also the part your auditor cares about. “The model was 90% sure” defends nothing. “The correct answer is in the returned set nine times in ten, at a risk level we chose, measured on rows nothing was fitted on, and it said nothing at all on the cases that cleared no bar” is a sentence a compliance team can sign.
Embedding Studio is that engine behind an API: fit it to cases your team already decided, set the precision the automatic share must hold, and download the calibration as a file you keep. $5 of credit, no card, and the first call needs no project at all.
One honest footnote: “only” is a claim about the public record on the date above, and we will keep it current. If another model ships conformal sets and abstention natively, this page will say so. The habit of saying so is worth as much as the architecture.