Breaking the Chains of Probability: Neutrosophic Logic as a New Framework for Epistemic Uncertainty in Large Language Models

TL;DR

This paper argues that the Softmax layer at the end of an LLM isn't just an implementation choice — it's an epistemological constraint. By forcing the probabilities of mutually exclusive outcomes to sum to 1, we collapse four distinct phenomena into the same format: ignorance, paradox, vagueness, and genuine moral conflict. The authors propose evaluating LLMs with neutrosophic logic (Truth, Indeterminacy, and Falsity as three independent dimensions in [0,1], with no sum constraint) and report that in 66% of unconstrained evaluations T + I + F > 1 — a regime they call hyper-truth that is structurally unreachable under probabilistic prompting (Proposition 1).

What I find interesting

The experiment is well controlled: same JSON output format under Strategy 1 and Strategy 2, the only thing that varies is the sum constraint communicated to the model. That cleanly isolates the effect of normalization from the effect of output format.
The largest representational gap appears exactly where it matters most for alignment: ethical contradiction (95% hyper-truth rate, ΔT = +0.267) and epistemic ignorance (ΔI = +0.383 — i.e., 38 percentage points of indeterminacy that normalization reroutes into confident assertions).
Mason (2026) independently replicated the effect across 5 additional vendors at 84%, so it doesn't look like an OpenAI-specific artifact.
The authors are explicit that this is a representational affordance, not a latent variable measured inside the model — I think that's the right framing and it avoids overclaiming.
Limitations worth flagging

Effective N = 100, with 5 stochastic replicates per cell — these are not independently sampled stimuli, so the Wilson CI and χ² should be read at the cell × repetition level.
Only 5 phenomena, one per class. Downstream validation is still open: does declared hyper-truth predict performance on trolley-problem reasoning, legal dilemma resolution, or paradox-tolerance benchmarks?
The plithogenic extension isn't pursued here, which is where Mason has been pushing and where Proposition 2 (non-injectivity of the scalar projection π) naturally points.
Question for the community

Does it make sense to think about a neutrosophic head as an optional post-Softmax evaluation layer — something a model could activate when it detects a domain sensitive to epistemic conflict (ethics, law, forecasting) — rather than replacing Softmax globally? Or does the audit/calibration cost of having two coexisting output formats make that impractical in production?