OBLIVION TextLM: A Query-Relative Control Plane for Efficient Generative Language-Model Inference

Long-context language models can ingest far more information than a particular query actually requires, yet retaining every represented distinction is neither equivalent to using it well nor automatically computationally efficient. We present OBLIVION TextLM, a model-agnostic query-relative control plane that operationalizes unresolved semantic obligations around a separately licensed pretrained language-model executor. The system constructs a live obligation state, routes evidence against that state, invokes the executor, permits new obligations to be born from intermediate evidence, generates evidence-bound completion certificates, verifies discharge conservatively, and retires an obligation only when verification and an explicit net-gain audit both succeed. Three matched modes—full-context baseline, retrieval-augmented generation, and OBLIVION—use the same executor to separate foundation-model quality from retrieval and control effects. The v2.0.0 reference implementation adds hybrid routing, citations, memory, streaming OpenAI-compatible clients, verifier separation, tracing, usage accounting, benchmarking, safety utilities, and advisory KV-cache ownership hooks. We validate the artifact by compilation, 29 automated tests, release-metadata checks, and deterministic three-mode smoke runs. We deliberately distinguish text-level semantic retirement from unmeasured physical KV-cache or GPU savings. The contribution is an executable, falsifiable systems framework for testing when verified query-relative state control improves quality–compute trade-offs without claiming a separately trained foundation model of its own.