Neuro-Symbolic Task Allocation for Heterogeneous Warehouse Robots with Versioned Knowledge Updates

Dynamic warehouse task allocation must coordinate loading, transport and unloading robots under changing conditions. We propose a versioned neuro-symbolic architecture in which DeepSeek proposes facts and restricted rules, an authenticated semantic shield validates them, and a mivar-inspired knowledge layer supports deterministic allocation. Stage A selects bounded dispatch candidates using CP-SAT; Stage B performs conflict-aware route and unloading-service search. The service graph supports single-carrier, abstract coalition and same-origin batch transport. Frozen evaluation separates allocation (1800 simulation runs), grounding (6102 responses) and cached-DeepSeek effects (60 episodes). Equal-map-family semantic exactness was 99.52% for facts and 93.02% for rules, but only 69.53% for moving-congestion rules. Cached restricted rules met prespecified simulator-level non-inferiority criteria for deadline misses, distance and a distance-based energy proxy. An additional 150 exploratory episodes evaluate candidate size, early unloading-robot selection, semantic shielding and emulated asynchronous integration. A separately frozen test issued 1248 new API requests on 416 synthetic events. With the contemporaneous backend, revised prompting achieved 320/320 correct congestion transactions versus 317/320 with the original prompt, without significant template-level superiority (Holm-adjusted p = 0.50). Typed-field compilation also reached 320/320 but failed validation on 10/32 access-restriction events. Complete trusted structured inputs do not require an LLM. The contribution is an auditable, bounded semantic interface; physical-robot safety, live-network reliability and physical-energy benefits remain unvalidated.