Abstract
This publication describes a system and method for verifying the intermediate quantities used in multi-body language-model fusion by means of an independent recomputation channel — a "shadow channel" — that runs parallel to the primary fusion pipeline. In multi-body fusion, a primary model (e.g., a transformer language model) is combined with one or more auxiliary bodies (e.g., an n-gram count model and a nearest-neighbour/datastore retrieval model) by interpolating their per-position probabilities. Conventional fusion consumes each body's reported probability directly and is therefore blind to two failure classes: (a) crash-faults, in which an auxiliary quantity becomes non-finite (e.g., a softmax underflow producing NaN), which can silently defeat downstream weight selection because a non-finite candidate loses every ordered comparison; and (b) silent value-faults, in which an auxiliary quantity is finite but wrong (e.g., an off-by-one alignment error in the n-gram body), which no finiteness or plausibility check on the reported value can detect. The disclosed shadow channel independently recomputes each auxiliary body from its own raw inputs — the datastore keys/values and the raw n-gram counts and alignment offset, not the primary pipeline's intermediate tensors — compares the independent result against the primary's reported value to produce a per-position disagreement signal (flagging any position whose relative disagreement exceeds a calibrated tolerance or whose value is non-finite), and folds an auxiliary body into the primary output only at positions where the two agree, deferring to the primary channel alone at flagged positions rather than propagating a suspect value. The independence of the recomputation is the essential feature: it is what permits detection of value-faults originating upstream of, or within, the primary pipeline's own computation of the auxiliary quantity, which a check applied to the reported value cannot reach. The method emits an audit record (per-body flagged fraction, median and maximum disagreement, non-finite indicator) that renders a corrupted run visible rather than silently absorbed. The disclosed novelty is limited to this specific combination; the constituent bodies (kNN-LM, interpolated Kneser-Ney) and general fault-tolerance techniques are acknowledged prior art.
Creative Commons License
This work is licensed under a Creative Commons Attribution-No Derivative Works 4.0 License.
Recommended Citation
Wise, David Lee and wise, avan lee, "Shadow-Channel Verification for Multi-Body Model Fusion: An Independent-Recompute Witness for Detecting Silent Value-Faults in Interpolated Language-Model Inference", Technical Disclosure Commons, (July 07, 2026)
https://www.tdcommons.org/dpubs_series/10827