OpenJev (Verdict): Non-Autoregressive Decision Engine (151M)

OpenJev (Verdict) is an open-source, post-trained foundational decision model built for structured software workflows, inspired by TypeSafe AI's Jev and Reinforcement Learning for Calibrated Decisions (RLCD). It provides calibrated semantic judgments (discrete choices, ordinal scores, and binary probabilities) in a single forward pass without conversational text generation.

• GitHub Repository: https://github.com/Heman10x-NGU/Verdict-open-jev
• Base Architecture: ModernBERT-base (knowledgator/gliclass-modern-base-v2.0, 151,378,177 parameters)
• Logit Capacity: 25 candidate slots (24 substantive options + 1 explicit abstention slot)

---

Key Features

1. Non-Autoregressive Single Pass: Evaluates all candidate options simultaneously in a single forward pass (< 35ms latency) without token generation loops.
2. Proper Scoring Calibration: Trained with composite Cross-Entropy + Brier Score loss: $$\mathcal{L}{\text{total}} = \mathcal{L}{\text{CE}} + 1.0 \times \mathcal{L}_{\text{Brier}}$$ followed by post-hoc L-BFGS temperature scaling ($T = 1.0716$).
3. Explicit Abstention Route: Dedicated __insufficient_evidence__ candidate slot ensures calibrated rejection on out-of-distribution or insufficient context queries.
4. Edge and In-Browser WebGPU: Runs locally in browsers via WebGPU/WASM and on servers via PyTorch/ONNX Runtime.

---

Quickstart

Python SDK

git clone https://github.com/Heman10x-NGU/Verdict-open-jev.git
cd Verdict-open-jev
pip install -e .
python scripts/download_artifacts.py

from rlcd import DecisionEngine, Choice, Option

engine = DecisionEngine()
query = Choice(
    question="What is the primary customer inquiry?",
    options=[
        Option(id="card_lost", description="Reporting a lost or stolen card"),
        Option(id="dispute_charge", description="Disputing an unrecognized charge"),
        Option(id="pin_reset", description="Requesting a PIN reminder or reset"),
    ]
)
result = engine.evaluate(
    context="I lost my wallet yesterday and need to stop my debit card immediately.",
    queries=[query]
)

print(f"Selected: {result.results[0].selected_option_id}")
print(f"Confidence: {result.results[0].confidence:.4f}")
print(f"Abstention probability: {result.results[0].p_abstain:.4f}")

Direct ONNX Runtime Loading

import onnxruntime as ort
from huggingface_hub import hf_hub_download

model_path = hf_hub_download(repo_id="heman10x/rlcd-modernbert-151m", filename="model.onnx")
session = ort.InferenceSession(model_path, providers=["CPUExecutionProvider"])

---

Empirical Benchmark Results

All metrics reflect evaluation on the held-out test split (1,000 cases, 5 candidates) and out-of-scope challenge sets:

---

Citation and Upstream Credits

• Inspired by TypeSafe AI's Jev architecture and RLCD (Reinforcement Learning for Calibrated Decisions).
• Base encoder backbone: ModernBERT (knowledgator/gliclass-modern-base-v2.0).
• Evaluation benchmarks: PolyAI Banking77 and CLINC150 Out-of-Scope datasets.