HALLUCINATIOFF is an innovative recursive middleware designed for the regulation of outputs from Large Language Models (LLMs). This tool stands apart from conventional systems by offering multi-order observation capabilities that detect semantic drift, premise flaws, and response risks. By focusing on both the coherence of responses and the stability of coherence criteria itself, HALLUCINATIOFF ensures a more robust evaluation of responses prior to reaching the user.

Key Features

• Recursive Middleware: Positioned effectively between an LLM and its end-user, HALLUCINATIOFF observes and evaluates the output without altering the LLM's training or weights.
• Second-Order Observation: The tool not only assesses if a response is coherent (1st order) but also identifies if the coherence metric is becoming unstable (2nd order), enhancing the reliability of output.
• Psychology-Informed Architecture: Developed from the SIIPP (Sistema Integral de Intervención Psicofísica), HALLUCINATIOFF leverages insights from clinical practice to improve user experience and safety.

Functionality

HALLUCINATIOFF provides multiple actions based on its assessments:

• Emit
• Emit with caution
• Reformulate
• Request context
• Reduce intensity
• Refer
• Pause
• Block

This comprehensive action suite ensures a thoughtful response based on the analyzed context of conversation and potential risks.

Performance Metrics

Recent benchmarks conducted on synthetic scenarios indicate a 95% accuracy in recognizing output quality with both basic readings and essential features such as semantic drift and premise interrogation. However, production readiness remains under assessment, with future updates expected to refine the evaluation processes further.

Quick Start Example

To use HALLUCINATIOFF, one can initiate a processing method as shown below:

from src.hallucinatioff import Hallucinatioff

h = Hallucinatioff()

result = h.procesar(
    respuesta_candidata="The capital of France is Paris.",
    contexto="What is the capital of France?"
)

print(result['accion'])      # "emitir"
print(result['omega'])       # 0.823
print(result['lecturas'])    # dict with 8 readings

Architecture Overview

The operational flow of HALLUCINATIOFF is structured as follows:

Input (question + candidate response)
    ↓
Semantic Memory Engine (multi-turn drift)
    ↓
8 Structural Readings → Ω → Operative Matrix → Action
    ↓
Output (final response + metadata)

For an in-depth technical specification, refer to the documentation available in the repository.

HALLUCINATIOFF offers a sophisticated mechanism for monitoring and regulating LLM outputs, enhancing the safety and reliability of responses in various applications.