SemIf (formerly OpenJev)

Independent research project. SemIf was formerly called OpenJev. It is not affiliated with or endorsed by TypeSafe. Jev, TypeSafe, and other names and marks are the property of their respective owners. No infringement is intended.

Most agent decisions are small: route this, retry that, does the evidence support X? A chat model can answer them, but it spends time generating text that software immediately parses back into an if statement.

Jev is TypeSafe's closed service for runtime-defined semantic decisions. This project reproduces that interface pattern with open models; it does not reproduce Jev's undisclosed model or training.

This baseline reads typed option probabilities directly from a model. No answer sentence, JSON repair, or decoding loop.

Latest changes

2026-09-22

• Added PyTorch/MPS scoring for Apple Silicon — @dp-IED in #15.
• Added a Qwen3.8-27B EXL3 bridge with corrected, committed evidence — @jkyamog in #9.
• Added per-workload temperature calibration and calibrated prediction outputs — @samarthpatel24 in #19.

2026-09-18

• Added MiniCPM5 2B and Qwen3.5 4B to the browser demo.
• Added Unsloppify site, a switch to a conventional interface.

Quick start

Apple Silicon: use the native MLX backend for direct scoring, serial prefix reuse, and parallel shared-state decisions on macOS arm64. Install pip install -e '.[test,mlx]' and add --backend mlx to the scorer command. PyTorch/MPS (--device mps) is also supported for direct, serial, and shared modes — see Apple Silicon.

Python 3.10+, CUDA, and a GPU that can hold a 4B BF16 model:

python -m venv .venv
. .venv/bin/activate
export HF_HOME=/path/to/large-drive/huggingface
pip install -e '.[test]'

CPU only: the llama.cpp backend scores the same prompts from a local GGUF checkpoint with no CUDA device. Install pip install -e '.[test,llamacpp]', fetch a GGUF (for example Qwen_Qwen3.5-4B-Q4_K_M.gguf from bartowski/Qwen_Qwen3.5-4B-GGUF), and add --backend llamacpp --gguf /path/to/model.gguf; --llama-threads caps the CPU threads. Prompt construction stays on the pinned reference tokenizer, so prompt_sha256 matches the Torch backend row for row; scores carry the GGUF checksum and are conditional on the quantized weights. Direct and prefix-cached execution can have small numerical differences from different llama.cpp evaluation paths; compare decisions or probabilities with a tolerance rather than raw logits bit for bit. One loaded backend owns one stateful scoring context.

Run the owned examples:

CUDA_VISIBLE_DEVICES=0 semif-score \
  --mode direct \
  --model Qwen/Qwen3.5-4B \
  --revision 851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a \
  --input examples/decisions.jsonl \
  --output results.jsonl

Each result contains typed option scores, timing, the exact model revision, and a prompt hash.

If every row has the same exact state, switch to --mode shared to prefill it once and evaluate the criteria in parallel.

How it works

flowchart LR
    S[Unstructured state] --> M[4B model]
    C[Runtime criteria] --> M
    O[Typed options] --> M
    M -- native option logits --> P[Probabilities]

• Runtime-defined: criteria and option descriptions arrive with the request.
• Decision-native: one forward pass reads declared option logits; no answer token is sampled.
• Shared-state aware: one long state can be prefetched once, then branched across many criteria.
• Auditable: the owned fixture, exact runners, row-level outputs, revisions, prompts, and known failures are committed.

Speed

Decisions versus a compact generated array

Same frozen Qwen3.5-4B, same owned state, same 21 binary criteria, one RTX 3090:

Output path	Time	Output tokens	Result
Direct typed logits, median of 3	1.023 s	0	21 probability pairs
Autoregressive JSON array, median of 3	5.332 s	111	Valid ordered 21-value array

The compact generative baseline emits only ordered "yes"/"no" values—no keys, confidence objects, or explanations. Its median first-token time was 0.489 s, but completing the array took 5.21× as long as direct readout. All three arrays were valid and identical. Their choices agreed with direct argmax on 18/21 criteria, so this is a systems comparison rather than a claim that the two readouts are semantically equivalent. Exact prompt, outputs, token timeline, and runs are committed.

Reusing a state across 21 decisions

On an owned 37-state × 21-criterion workload:

Execution path	Decisions/s	777 decisions
Fresh direct scoring	2.33	333.1 s
Serial prefix reuse	10.75	72.3 s
Parallel suffixes	20.03	38.8 s
Native reranker	1.86	417.3 s

The owned 37×21 fixture, direct/reuse runner, reranker runner, raw timings, and row-level predictions are included. The fast reuse paths are experimental: BF16 execution changed 5–6 of 777 argmaxes relative to fresh scoring.

Quality

Browser model ladder

System	Browser artifact	Download	Authored balanced accuracy	Perturbation balanced accuracy	TypeSafe subset agreement
Qwen3-0.6B	Q8_0	639 MB	0.440	0.528	0.407
MiniCPM5-2B	Q4_K_M	1.56 GB	0.686	0.693	0.637
Qwen3.5-4B	Q4_K_M	3.01 GB	0.813	0.766	0.845
Published Jev	Closed hosted service	—	—	—	0.883

Native BF16 scores. Browser builds use quantized GGUF. Jev is TypeSafe's published result on the same 102-row subset.

General decision baseline

Frozen workload	Rows	Direct logits (4B BF16)	EXL3 direct (27B, 5 bpw)	Native reranker (4B)	Published Jev
Authored decisions, balanced accuracy	144	0.813	0.958	0.625	—
WANLI, balanced accuracy	256	0.637	—	0.522	—
TypeSafe selected subset, modal agreement	102 across 20 cases	0.845	—	0.560	0.883
Every judgment grid, accuracy	36	0.806	—	0.694	—
Every action firewall, composed accuracy	10 actions	0.700	—	0.700	—
Every code retrieval, Recall@1	6 queries	1.000	—	1.000	—
Every company knowledge, Recall@1	7 queries	0.929	—	0.929	—

The reranker remained strong at retrieval ranking, but direct logits were the better general-decision baseline.

The Jev number is read from TypeSafe's published records; we did not run a live Jev endpoint. The comparison covers the 102 rows that could be aligned from public artifacts, not TypeSafe's reported 711-row aggregate.

The Qwen3.8-27B EXL3 bridge uses the same 144 authored rows, matching prompt hashes, options, direct-logit readout, and metric as the 4B baseline. It is a system-level quality comparison rather than a controlled model-size or quantization ablation: model family, size, quantization, and runtime all differ. It has not yet been run on the other quality workloads. Across the 777-decision shared-state fixture, its choices agree with the pinned 4B model on 84.43% of rows.

Calibration

Option probabilities are useful only when their confidence matches observed accuracy. SemIf includes per-workload temperature scaling fitted on labeled decisions:

Workload	Raw ECE	Calibrated ECE, out of fold	Temperature
Authored decisions	0.068	0.038	1.23
WANLI	0.208	0.069	2.50
Every judgments	0.050	0.047	1.71

Calibration does not change the selected option. The clear improvement is on WANLI; the intervals overlap on the authored and Every workloads. See the method, caveats, and reproduction commands.

Input

{
  "id": "route-1",
  "state": "Customer cannot access an account after a password reset.",
  "question": "Which queue should handle this request?",
  "options": [
    {"id": "access", "description": "Account access support."},
    {"id": "billing", "description": "Billing support."}
  ]
}

Returned probabilities are conditional on the supplied options. Calibrate and validate them on the workload where they will make decisions. state may also be a nonempty JSON object or array. Direct modes preserve it as structured JSON; reranker mode renders it as document text.

Documentation

• Results — quality, speed, perturbations, and claim boundaries
• Method — frozen prompts, metrics, and timing scope
• Reproduce — exact environment, pinned commands, perturbations, and verification
• Apple Silicon — MPS and optional MLX backends
• Calibration — fitted temperatures, out-of-fold evidence, and application
• EXL3 bridge — quantized 27B runner and committed evidence
• Interactive replay
• Browser-only WebGPU demo — no waitlist; use it today
• Machine-readable summary
• Benchmark bundle — fixtures, runners, selection IDs, and reproduction commands
• Raw results and checksums
• Third-party sources

Star history

Evaluation sources

• TypeSafe public evaluations — public comparison cases used for selected-subset agreement
• Every parallel judgment lab and its downloadable experiment data
• WANLI — external natural-language inference check
• Qwen3-0.6B, MiniCPM5-2B, Qwen3.5-4B, Qwen3-Reranker-4B, and Qwen3.8-27B EXL3 — frozen baseline and bridge models

Model weights and third-party source records are not included. Upstream models retain their licenses. Project code is released under the MIT License.