SemIf
Semantic ifs on a home 3090: open models doing structured branches without pretending to be TypeSafe.
Overview
SemIf-OpenJev (github.com/TheoLeeCJ/SemIf-OpenJev, openjev.com) is Theo Lee CJ's independent decision server formerly marketed as OpenJev on the site. It reads shared state once, asks many typed questions, and returns option probabilities from open models in a single forward pass. The project explicitly disclaims affiliation with TypeSafe: same interface ideas, different weights and training. WebGPU browser demos and PyTorch/MPS paths ship in-repo for replayable evidence.
Problem: Teams want System One shaped APIs and calibrated option probabilities without routing every gate through a closed hosted model.
Built for: Researchers and hackers self-hosting open weights who need Choice, Score, and Noul style batches on hardware they control.
First indexed on Jev Directory: 2026-09-17
Creator and team
- Name
- Theo Lee CJ
- Handle
- @TheoLeeCJ
How Jev is used
- Role in the product flow
- Self-hosted System One compatible decision server over open model logits
- Primitives
- ChoiceScoreNoul
- State in
- Shared text or structured state plus runtime-defined question schemas (criteria maps, rubric scores, boolean nouls) documented in SemIf harnesses.
- Decision out
- Per-question probability vectors read directly from logits without generating answer prose or JSON repair loops.
- Pack state and typed questions into one model forward pass
- Read option logits for each Choice, Score, or Noul head
- Return calibrated-style scores (operators should run local calibration per README)
- Downstream code thresholds probabilities like any System One client
SemIf is the open alternative when you want semantic ifs without TypeSafe inference bills. Agent code still writes thresholds; the server answers which branch wins, how risky a state feels, or whether evidence supports a claim. Because probabilities come from logits, you skip the usual chat completion detour. Treat scores as research baselines: README adds temperature calibration work and warns that home GPU numbers are not production SLAs. Pair SemIf with hosted classifier.dev when you want a managed fast tier beside your own hardware experiments.
Sourced performance claims
- README positions SemIf as reproducing the interface pattern, not TypeSafe's undisclosed model or training.Source: github.com/TheoLeeCJ/SemIf-OpenJev README
Features and stack
Features
- openjev.com browser demos plus local PyTorch and MPS scoring
- Independent disclaimer and renamed SemIf branding in README
- Batch many Choice, Score, and Noul questions on one state
- Community bridges for additional open weight families
Stack
- Python
- PyTorch
- WebGPU demo
- Open weights
- Self-hosted inference
Pricing: Open source software; you pay for GPUs, electricity, and any cloud you rent. Not a TypeSafe substitute for compliance-sensitive hosted gates.
Links
FAQ
- Is SemIf the same as TypeSafe Jev?
- No. README states SemIf is independent research that mimics the API shape with open models, not TypeSafe's closed service.
- Can I trust the default probabilities out of the box?
- README documents calibration passes and workload-specific temperature tuning. Run the bundled eval harnesses on your hardware before hard gates.
- Where did the OpenJev name go?
- The project rebranded to SemIf while keeping openjev.com as the demo home. This listing uses catalog slug theoleecj-semif.
Related learn guides
Original Jev guidance that pairs with this product pattern.
- System One model
The model family behind Jev: parallel typed questions, one forward pass, probabilities you can threshold.
- Jev vs LLM classification
When to gate with System One probabilities instead of asking a chat model to label things.
Related products
Hand-picked neighbors with rich profiles or overlapping tags.