KDSPM: Knowledge Distillation via Semantic Pruning and Merging with Contextual Drift Capture

We introduce KDSPM, a framework that shifts the unit of knowledge synthesis from documents to atomic claims. The framework decomposes source documents into minimal verifiable assertions, scores each along novelty, corroboration, and semantic drift axes, and applies graduated pruning with a three-way disposition (keep, quarantine, prune). We introduce two novel contributions: (1) the epistemic drift score, which treats inter-source disagreement as an informational signal rather than error, and (2) the uncertainty-as-novelty principle, which protects genuinely novel ideas by quarantining structurally coherent but unclassifiable claims rather than discarding them.