Multiverse Superposition Inference: Deferred Collapse over Externally-Weighted Decoding Paths

Greedy decoding is locally optimal but globally short-sighted; beam search prunes early; self-consistency samples independently and votes only at the end, losing mid-path information. We propose Multiverse Superposition Inference (MSI): a decoding strategy that maintains K parallel paths through generation, with each path carrying an evolving credence weight derived from logprob plus optional external signals (routing-archetype coherence, retrieval-corpus alignment, register-state consistency). At collapse points — sentence boundaries, semantic shifts, archetype transitions — paths aggregate via voting, weighted blending, or semantic merging. Distinct from self-consistency, which commits to independent samples and a final aggregate vote, MSI maintains correlated paths during generation with deferred collapse. The reference implementation is adapter-level (training N LoRA adapters in parallel and blending at inference via the HeartScale Blender); segment-level and token-level MSI with shared-prefix copy-on-write KV cache management are specified in the paper and proposed for v1.1. We discuss compute and latency trade-offs and an empirical program covering long-form generation coherence, reasoning benchmarks (GSM8K, MATH, HotpotQA), and ablations across path count and external signal contribution. Empirical results are deferred to v1.1. The 'superposition' framing is metaphorical; the operations are classical.