Fact Lifecycles in Next-Concept Modeling: Tracking Candidate-Claim Status Across the Decoding Hierarchy

This preprint presents a descriptive study of candidate-claim status across nine recorded hierarchy levels of one Next-Concept Modeling (NCM) checkpoint on the 11,334-article XSum test split. Sentence-level candidates are assigned automated source-conditioned natural-language-inference (NLI) labels; supported, contradicted, or unresolved; and adjacent-level candidates are matched using TF–IDF cosine similarity with one-to-one Hungarian assignment. The analysis reports article-weighted label fractions, reference-overlap metrics, correlations between ROUGE-L and contradicted-label fractions, and saved heuristic transition counts.

The contradicted-label fraction rises from 0.222410 at level 1 to 0.484776 at level 9, while mean ROUGE-L rises from 0.032465 to 0.279258. These are descriptive outputs of the specified evaluation pipeline, not human-verified factuality judgments or causal evidence. Candidate sentences are not necessarily atomic claims; unresolved labels do not imply falsity; and introduced or removed candidates are heuristic matching outcomes rather than verified factual changes.

The manuscript also documents an unresolved discrepancy between two correlation calculation paths and limitations in source-to-run provenance. No full end-to-end reproducibility claim is made.