Preprint intended for submission to NeurIPS 2026 (Thirty-Ninth Conference on Neural Information Processing Systems). Large Language Models produce stochastic outputs that undermine reproducibility in knowledge extraction tasks. This paper presents the 14-Step FOSS Hallucination Gate, a deterministic post-processing framework that transforms unreliableLLM outputs into consistent, validated causal triplets. Key Results: • 88% precision on DocRED (2,614 triplets, IAA-validated, Fleiss' κ=0.47) • 100% byte-level determinism (SIEB benchmark: 150 extractions) • Multi-model validation: 100% SIEB across Qwen-8B, Gemma-2B, Llama-3B • 2.5× improvement over naive baselines on complex documents (p < 0.001) • 4,355 validated triplets from 1,997 documents in 7.1 hours • Consumer hardware: Mac mini M4 (16GB RAM) The key contribution: reliability emerges not from better models, but from deterministic validation architecture. This package contains the paper, all experimental data, benchmark results, and reproducibility scripts.
David Tom Foss (Tue,) studied this question.