OBLIVION: Query-Relative Neural Computation by Verified Semantic Obligation Retirement

Abstract

Neural sequence models allocate computation to represented context even when a query requires a subset of distinctions. We introduce OBLIVION, a query-relative neural computation architecture whose active state represents unresolved obligations. Obligations may be born when intermediate evidence exposes new requirements and may be retired only after verified discharge and a cost audit that charges routing, certification, transformation, and downstream work. We formalize Semantic Obligation State, Verified Net Discharge, counterfactual discharge, obligation birth, adaptive obligation budgets, and an SHC-aware net-gain condition. Controlled experiments establish query-dependent width, learned gating, conservative uncertainty handling, distractor scaling, natural-input reduction, and multi-hop obligation birth. A pilot uses a frozen 66.36-million-parameter DistilBERT QA model across context policies. On 28 answer-recoverable bridge questions drawn from three disjoint held-out HotpotQA subsets, OBLIVION raises mean answer F1 from 23.31% under exhaustive paragraph scoring to 52.47% while exposing 15.66% of paragraph-token context; the paired-bootstrap F1 gain is 29.19 points with a 95% interval of 8.99–48.95. Audits also reveal incomplete benchmark contexts and cases where routing overhead removes computational gains. OBLIVION therefore provides a falsifiable architecture hypothesis, not evidence of universal Transformer superiority. Negative controls and unreproduced exploratory metrics remain archived for transparent falsification.