flowxai/sentinel-gate

TEXT GENERATIONConcurrent Unit Cost:1Model Size:4BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jul 10, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

FlowX.AI's Sentinel Gate is a 4 billion parameter LoRA fine-tune of Qwen3-4B, designed as an on-device escalation decision gate for regulated decisions. It processes complete case information, including domain facts and policy schemas, to deterministically decide whether to ESCALATE a case to a human or DECIDE it is safe to automate. Optimized for safety-critical compliance automation, it ensures zero false-negative and false-positive rates on its evaluation set for the core ESCALATE/DECIDE action.

Loading preview...

FlowX Sentinel Gate (4B) Overview

FlowX.AI's Sentinel Gate is a 4 billion parameter model, fine-tuned from Qwen3-4B, specifically engineered as an on-device escalation decision gate for regulated processes. Its primary function is to analyze a complete case, including domain facts and an applicable policy_schema, and output a strict JSON decision: either ESCALATE (route to a human) or DECIDE (safe to automate).

Key Capabilities & Features

  • Safety-Critical Decision Making: Achieves a 0.000 false-negative rate (no missed escalations) and a 0.000 false-positive rate (no over-escalation) on its held-out evaluation set, ensuring high reliability for compliance automation.
  • Structured Output: Provides a detailed JSON output including the action (ESCALATE/DECIDE), escalation_category (as a routing hint), confidence_score, confidence_reasoning, human_action_required, and an audit_trail.
  • Deterministic Inference: Requires a frozen inference contract with a specific system prompt, user-turn format, and decode settings (enable_thinking=False, temperature 0, max_new_tokens ~1200) to ensure reproducible and auditable decisions.
  • Multi-format Support: Available in fp16 safetensors (for CUDA/vLLM), MLX-quantized (int4/int8 for Apple Silicon), and GGUF (Q8_0/Q4_K_M for CPU/Ollama).
  • Compliance Automation: Designed to integrate into compliance pipelines, sitting after flowxai/semantic-mapper to determine which cases are safe for automation versus those requiring human intervention.

Intended Use Cases

  • Regulated Decision Automation: Ideal for automating decisions in domains like banking, insurance, logistics, and labor where compliance and safety are paramount.
  • Human-in-the-Loop Systems: Facilitates efficient routing of complex or high-risk cases to human experts while automating routine decisions.
  • Audit & Traceability: The structured output and audit trail provide clear reasoning for each decision, crucial for regulatory compliance and post-hoc analysis.

Finetunes
1 models