Reliable clinical prediction requires both accuracy and principled uncertainty quantification, so that a model can flag cases where it lacks a coherent basis for its prediction and defer to a clinician. We introduce the Temporal Epistemic Graph of Thought (TEGoT), a framework that fuses free-text clinical notes with structured temporal events and models uncertainty over latent reasoning structure. Patient trajectories are represented as temporal graphs integrated with note embeddings, and a variational reasoning graph induces a distribution over latent topologies, decomposing predictive uncertainty into aleatoric, epistemic, and structural components, where structural uncertainty measures variability across inferred reasoning graphs. On a retrospective MIMIC-IV ICU mortality task, using structured events available before 12 hours and discharge-summary-derived note embeddings as a proof-of-concept text source, TEGoT reaches AUROC 0.931, AUPRC 0.655, and ECE 0.021 after temperature scaling, comparable to or above tabular, neural, and graph baselines under the same cohort. Structural uncertainty gives stronger error detection (AUROC 0.724) than standard epistemic uncertainty (0.681), and its weak correlation with epistemic uncertainty (𝜌=0.19) indicates a complementary signal rather than a repackaging of dropout variance. Using this signal for selective prediction raises retained AUROC from 0.931 to 0.963 at 50% coverage, with a net benefit of +0.083 over treat-all and a decision-curve-positive range of [0.05, 0.42], supporting reliable abstention. We argue that the social value of a clinical model should be measured not by discrimination alone but by the harm it averts when it defers, and we show that net benefit and risk–coverage trade-offs offer concrete, deployment-relevant ways to quantify that benefit.

Explicit subspace disentanglement, enforcing orthogonality and simplex constraints on learned projections, is an increasingly common recipe for paraphrase models claimed to separate lexical, syntactic, and semantic information. A dual-encoder hierarchical model trained with these constraints on the DiPPer corpus of LLM-generated paraphrases produces near-perfect geometric separation, a pairwise Centered Kernel Alignment (CKA) of roughly 0.0007, yet that separation is not evidence of functional linguistic specialization. Three results establish this. First, near-zero CKA is largely preordained by an explicit orthogonality penalty, which converges to 4×10^-6. Second, each of the three subspaces, including the nominally syntactic one for which we identify no functional probe, classifies paraphrase in isolation at F₁ above 0.986, so the subspaces are not functionally distinct. Third, removing disentanglement at matched capacity costs only 0.013 F₁, leaving most of the reported gain attributable to dual-encoder capacity rather than disentanglement. Capacity-matched ablations and controlled functional probes, not CKA alone, are required to substantiate disentanglement claims; orthogonality penalties otherwise manufacture the very metric used to certify them.

Health insurance policy documents justify coverage decisions through clinical evidence and regulatory criteria, but the underlying reasoning is buried in dense medical and legal text. A pipeline is presented that extracts argument components and relations from these documents and assembles them into directed argument graphs. A dataset of 5,230 sentences and 5,685 relations across 53 documents from Aetna, CMS, and United Healthcare is constructed along with two custom annotation tools. Six classical baselines, three fine-tuned transformer models, and three instruction-tuned LLMs are evaluated, with the LLMs assessed under zero-shot, few-shot, and chain-of-thought prompting. PubMedBERT and Legal-BERT, fine-tuned on the dataset, achieve the strongest results overall (0.731 and 0.653 Macro F1 on component classification and relation extraction, respectively). Notably, the much simpler TF-IDF + SVM baseline is competitive with these fine-tuned transformers (0.709 and 0.649 Macro F1) and outperforms the best prompted LLM by 12 and 21 F1 points respectively. The dataset and code are available at https://anonymous.4open.science/r/LLM-Driven-Argument-Mapping-3ACB/README.md