Internal Verdicts Track Evidence: A Template-Controlled Neutrosophic Reading of Epistemic States in Large Language Models via the Jacobian Lens

Abstract

For twenty-five years, neutrosophic logic [1, 2, 3] has assigned independent degrees of truth (T), indeterminacy (I), and falsity (F) to model epistemic states that classical and fuzzy semantics cannot  express. All previous applications, however, measured these components on outputs: answers, judgments, expert evaluations. This paper reports, to our knowledge, the first neutrosophic reading of components  inside the internal representations of a large language model. Using the recently released Jacobian lens [9] on Qwen3.5-4B, we project intermediate-layer readouts onto lexicons of support, refutation, and  hedging, obtaining layer-wise (T, I, F) profiles under three epistemic conditions (conflicting evidence, first-person false belief, factual control; n = 20 each). A first battery yielded three striking  signatures — a surge of refutation mass under conflict, sustained indeterminacy under false belief, and an apparent "verdict collapse" before the output layer. A second, template-controlled battery showed  that all three, as initially stated, were largely artifacts of prompt grammar: matched templates with agreeing evidence inflate the same lexicon masses, true beliefs elicit the same hedging, and the model  verbalizes its verdict when allowed to generate (17/20 items). What survives the controls is stronger than what died: the neutrosophic verdict balance B = log10(F) − log10(T) tracks the polarity of the  evidence under identical templates (separation of about 3 orders of magnitude, Cohen's d = 2.3–2.5, p < 10-6 in 9 of 9 layers; correct item-level classification 18/20 and 17/20), and residually discriminates  false from true first-person beliefs (p < 2 × 10-4 at every layer). We argue that the independence of the neutrosophic components — the axiom that distinguishes neutrosophy from fuzzy and classical  frameworks — is precisely what makes this measurement and its self-correction possible, and we distill the two-battery design into a reusable protocol for neutrosophic auditing of language model internals.
  DOI 10.5281/zenodo.22249786