redraft: Reactive LLM Inference by Salvaging Stale Outputs as Self-Speculative Drafts

When the context of an LLM call is edited, today’s serving stacks reuse the unedited prompt prefix but regenerate the answer from scratch, even though most of the old answer is often still exactly what the model would say. We present redraft, an inference primitive that propagates a context edit through the generation itself: the stale output is replayed as a self-speculative draft under the edited context, verified in batched windows, and re-anchored after each divergence, so decoding cost scales with how much of the answer the edit actually invalidated. Naive splice reuse is capped by decode instability (a semantically irrelevant edit still rewrites half the tokens); redraft closes that gap with a tolerance-based acceptance rule whose band scales with the local top-𝑘 entropy, calibrated across three models for zero correctness regressions while holding a median 68% of output tokens. Implemented inside llama.cpp as a streaming server primitive, redraft reaches median wall-clock speedups of 2.31× on document summaries, 2.76× on factual answers, and 1.66× on code review over full regeneration on a dense 14B model across a 36-case edit suite. We further give, to our knowledge, the first treatment of self-speculation on recurrent-memory architectures, where rollback is impossible: a reprocess-on-divergence verifier with a parked post-template state makes long, mostly-stable answers profitable (up to 3.6×) even on a hybrid Gated DeltaNet model.