SemIf

What it does

SemIf, launched as OpenJev and renamed two days later, reproduces the interface of Jev with open models: you send a state, a question, and a list of options, and it returns a probability for each option from a single forward pass. It doesn’t generate an answer and parse it. Instead it reads the model’s scores for each option directly. The author is careful to say it reproduces the interface pattern, not Jev’s undisclosed model or training, and that the project isn’t affiliated with TypeSafe.

The main configuration runs Qwen3.5-4B on a single consumer GPU. There’s also a browser demo that runs smaller quantized models entirely in your browser with WebGPU.

Setup

The documented path needs Python 3.10+, CUDA, and a GPU that can hold a 4B model in BF16:

pip install -e '.[test]'
CUDA_VISIBLE_DEVICES=0 semif-score --mode direct \
  --model Qwen/Qwen3.5-4B --revision 851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a \
  --input examples/decisions.jsonl --output results.jsonl

We had no GPU, so we installed the CPU build of PyTorch. All 11 unit tests passed, but the model loader refused to run without exactly one CUDA device, as documented. To exercise the actual scoring code, we loaded the project’s pinned Qwen3-0.6B revision on CPU ourselves and called its direct.score function, bypassing only that device check. This isn’t a supported configuration, but it confirmed the mechanism works end to end.

Using it

• The output is built for auditing. Each result includes the option IDs, raw option logits, probabilities that sum to one, the input token count, the exact model and revision, a prompt version, and a SHA-256 hash of the prompt. You can trace any decision back to exactly what the model saw.
• It’s honest about what the numbers mean. Every result carries a probability_status field that reads “conditional option score; uncalibrated as decision confidence.” That’s the key difference from Jev, which is trained specifically so its confidence is calibrated. If you swap SemIf in for Jev, re-derive your thresholds on your own data first. Our confidence thresholds guide explains how, and jevcal can help.
• The README is unusually rigorous. Fixtures, runners, raw timings, row-level predictions, prompts, and known failures are all committed. The author reports that reading option scores directly for 21 yes/no criteria took about a second on an RTX 3090, versus about five seconds for the same model generating a compact JSON array. The README also compares SemIf with TypeSafe’s published evaluation results on a subset of rows the author could align, and explains the limits of that comparison.
• Shared-state modes are marked experimental. Modes that reuse one long state across many questions are much faster in the author’s measurements, but the README notes they changed a handful of answers compared with fresh scoring.

Permissions and data

• Everything runs locally. The command-line tool loads model weights from Hugging Face and runs on your own GPU, so your data never goes to TypeSafe or any other API.
• Model downloads are pinned. Remote models must be requested by a full 40-character commit revision, and trust_remote_code is turned off, so a changed model repository can’t silently swap in different weights or execute code on your machine.
• Check each model’s license. SemIf’s code is MIT licensed, but the models it uses keep their own terms. THIRD_PARTY.md lists each one with its pinned revision.
• The browser demo downloads large files. The models it offers range from roughly 0.6 GB to 3 GB. Per the project, inference runs in your browser; we didn’t audit the hosted demo’s network traffic.

Maintenance

Nine commits over three days, a few open issues, and a clearly scoped “phase 1” label. The rename from OpenJev to SemIf came with prominent non-affiliation notices; the project didn’t state a reason. Link to the new repository name, since the old one only works through GitHub’s redirect.

Verdict

SemIf is a notably careful open reproduction: pinned everything, committed evidence, and plain language about what it doesn’t do. It’s a strong choice if you want typed decisions on your own hardware, or if you want to understand how System One models work. Just don’t treat its scores as calibrated confidence without measuring, and budget for a GPU if you want to run it outside the browser.

For other open approaches, see Jevlike, NanoJev, and the Jev-compatible openjev-sglang server, or read our comparison of Jev alternatives.

See how it compares with other tools in Best Jev tools, tested hands-on.

Review updated Sep 19, 2026. Numbers quoted from the project are its author's own; we don't publish our own measurements of Jev.