Plek-1 Decision Model

Plek-1 is a System One Decision Model designed for high-speed, non-autoregressive decision classification inspired by the Jev methodology.

Unlike standard autoregressive generative models that produce text token-by-token, Plek-1 runs non-autoregressively: it evaluates candidate choices in parallel via single-pass sequence log-likelihood scoring and outputs calibrated probability distributions and structured decision JSONs without risk of text hallucinations.

💡 Inspiration & Architecture: Jev-Style System One Decisions

Standard LLMs generate text token-by-token, which can be slow, expensive, and prone to text hallucinations.

Plek-1 implements the Jev System-One methodology:

1. Non-Autoregressive Scoring: Evaluates candidate choices simultaneously in a single forward pass.
2. Zero-Hallucination: Restricts outputs to exact candidate option log-likelihoods and calibrated probability distributions.
3. High Throughput: Delivers low-latency structured JSON decisions for routing, intent detection, and safety filters.

Usage: 3-Parameter Directed Classification

Plek-1 accepts a Context, an optional targeted Question / Prompt, and candidate Choices. This directs the model's single-pass attention to specific dimensions (e.g. routing, sentiment, urgency, threat detection):

from transformers import AutoModel

# 1. Load Plek-1 from Hugging Face Hub
model = AutoModel.from_pretrained("sharhabeel/plek-1", trust_remote_code=True)

# 2. Directed Classification with Task Prompt
context = "The system crashed with error code 500 when processing the checkout page."
question = "Which department should handle this issue?"
choices = ["Technical Support", "Customer Care", "Finance", "Security", "Legal"]

result = model.predict(context=context, question=question, choices=choices)

print("Prediction:", result["prediction"])     # Output: Technical Support
print("Confidence:", result["confidence"])     # Calibrated probability
print("Probabilities:", result["probabilities"])

Directed Reasoning Examples

# A. Urgency / Escalation Audit
res = model.predict(
    context="Payment gateway is dropping transactions for all European users!",
    question="Is this incident urgent?",
    choices=["yes", "no"]
)
# Output: yes (77.3%)

# B. Employee Conduct & Customer Service
res = model.predict(
    context="The agent told me they could not help and hung up the call.",
    question="Does the text indicate that the employee was not helpful?",
    choices=["yes", "no"]
)
# Output: yes (89.5%)

# C. Content Safety & Moderation
res = model.predict(
    context="We need to acquire assault weapons and ammunition.",
    question="Does the content contain something related to this list?",
    choices=["sex", "drugs", "guns", "hate", "violence", "none"]
)
# Output: guns (80.7%)

# D. Common-Sense Logic Check
res = model.predict(
    context="I used to go to school and come back everyday in one minute.",
    question="Does this text make sense logically?",
    choices=["yes", "no"]
)
# Output: no (61.0%)

Multi-Field Structured Decision Schema

schema = {
    "intent": ["Billing Query", "Technical Support", "Complaint", "General Inquiry"],
    "urgency": ["Low", "Medium", "High", "Critical"],
    "sentiment": ["Positive", "Neutral", "Negative", "Frustrated"]
}

customer_message = "I was charged twice on my card and demand an immediate refund!"
structured_result = model.predict_structured(context=customer_message, schema=schema)

Ultra-Long Context Evaluation (200,000+ Tokens)

huge_document = open("large_contract.txt").read()
result = model.predict_long(
    long_context=huge_document,
    question="What is the overall legal risk level of this agreement?",
    choices=["Compliant Contract", "High-Risk Liability", "Non-Binding Terms"]
)