Jev vs SemIf / OpenJev: Buy the Decision, or Read the Logits Yourself
SemIf — previously called OpenJev — is the most interesting item in this list, because it is not a competing model. It is a technique, packaged: take an open model you already have, declare your options, and read their logits in a single forward pass without sampling an answer token.
That is a genuinely different trade. You are not buying a decision model; you are getting decision-shaped behaviour out of a general model you control, with the quality ceiling of whatever model you point it at.
Background
What SemIf Is
The project describes itself as semantic ifs from open models, running on a consumer GPU at home, and is explicit that it is independent of TypeSafe and does not reproduce hosted Jev’s training or accuracy. It implements runtime-defined criteria and option descriptions, and reads declared option probabilities from one forward pass rather than generating an answer.
The engineering argument is sound and worth understanding even if you never run it: if the options are known before the forward pass, the model has already scored them, and sampling tokens to say which one it picked is a step you can skip. That is the same insight the hosted decision models are built on.
The difference is everything around it. A packaged technique inherits the quality, the calibration and the failure modes of the base model you point it at, and it is on you to work out what those are for your task.
Source for the factual claims above: github.com/TheoLeeCJ/SemIf, checked 2026-09-25. We link it so you can check it rather than take our word for it.
Side by side
What Actually Differs
Based on the project’s own description, checked on 2026-09-25.
| Jev (via jev-ai.org) | SemIf / OpenJev | |
|---|---|---|
| What it is | A model trained for calibrated decisions | A method for reading decisions out of a general open model |
| Where it runs | Hosted; one HTTPS call | Your hardware, your model, your serving stack |
| Quality ceiling | The vendor's model, whatever it is | Whatever base model you point it at |
| Calibration | Trained for it; vendor publishes claims | Raw logits from a model not trained for it; yours to evaluate |
| Data residency | Leaves your network | Stays put, if you host it |
| Marginal cost | Per call, input tokens only | GPU time, whether or not you are using it |
| Auditability | Request logs and usage on your account | The full pipeline, because you own it |
How to choose
Neither One Wins Every Time
Choose the hosted model when
- You want calibrated probabilities without having to establish, yourself, whether the numbers your pipeline produces mean anything.
- You do not want a GPU in the critical path of a support ticket being routed.
- You want a documented error contract, rate limits, spend caps and an account with a history.
- The team that would own the self-hosted version has other things to do this quarter.
Choose SemIf when
- Data residency is a hard requirement. Nothing about a hosted API answers that.
- You already serve an open model and adding decision-shaped reads to it costs you almost nothing extra.
- You want to inspect and modify the whole pipeline — the project emphasises committed fixtures, runners and row-level outputs, which is the right instinct.
- You are researching how decision models work, rather than shipping one. It is an unusually legible implementation of the idea.
We sell one of these two, so read the right-hand column with that in mind — and then run both on two hundred of your own labelled examples, which settles it better than any page on the internet can.
Questions
Jev vs SemIf / OpenJev: Common Questions
Is SemIf affiliated with TypeSafe or with Jev?
No, and the project says so itself. It is an independent implementation of a similar interface on open models, and does not run or reproduce the hosted model.
Will reading logits give me the same calibration?
Not automatically. A base model’s logits are not the same thing as probabilities trained to be well calibrated, and the gap is exactly what a decision model is trained to close. If you go this route, measure calibration on your own labelled data before you build thresholds on the numbers.
Which base model should I use with it?
That is the whole question, and it is yours to answer. The method’s quality ceiling is the base model’s understanding of your task, so evaluate candidates on your data the way you would evaluate any classifier.
Could I use the hosted model to bootstrap the self-hosted one?
Yes, and it is a sensible path: use a hosted decision model to label a set cheaply, use the labels to select and validate a base model, then move the volume in-house once you can show the quality holds.
More
Other Comparisons
The Cheapest Way to Settle It
Load a scenario, paste in your own text, and see what the distribution says. Editing costs nothing and needs no account.