Deep Reasoning: From Statistical Plausibility to Constrained Semantic Reasoning

This preprint formalizes Deep Reasoning as an operational framework for constrained semantic reasoning. Its objective is not to deny the usefulness of statistical language models, but to define an additional level of verification in which an answer is tested through constraints, admissible inversion, and pure negative.

V1.0 preserves the previous methodological clarifications and strengthens the scientific framing of the document: Deep Reasoning is not presented as an attack on LLMs, but as a complementary layer for stabilizing, bounding, and certifying selected linguistic outputs.

Deep Reasoning is independent of any specific LLM architecture: it applies to any system producing linguistic outputs. Its self-positioning module addresses a structural inversion specific to language models: unlike a human child who learns to read before learning through reading, a language model receives the tool and all the subjects simultaneously, without sequence. Self-positioning is the mechanism that compensates for this absence.