Comprehensive Analysis: The Calibration Phase (EXP01–07) – Building an Epistemic Gate for LLMs

Developer Eduardo Ayala has released a detailed technical breakdown of his project, 'Beatriz,' an epistemic gate designed to prevent data poisoning during LLM fine-tuning. Over a 17-day experimental period, Ayala conducted 16 experiments to refine a mechanism that distinguishes between true and false information. The core innovation involves a vector-based gate that uses semantic geometry rather than simple lexical matching, combined with a bounded Softplus loss function to prevent model collapse. By comparing various configurations, the project demonstrates that architectural choices, such as the implementation of LoRA, were more effective at controlling perplexity than reference-based loss methods. The findings highlight the importance of rigorous, reproducible methodologies in AI research. The project, which was developed on legacy hardware, is now available as open-source, providing a transparent look at how to build more robust and reliable fine-tuning pipelines for large language models.