Large Language Models excel at parsing unstructured geopolitical, customs, and supplier force-majeure advisories, but fail catastrophically when asked to directly output feasible multi-echelon shipment quantities under coupled capacity, fixed-charge, and SLA constraints.
ChronosMesh enforces a strict Neuro-Symbolic boundary:
src/chronos_mesh/neuro/): Unstructured text $\mathcal{T}$ is compiled into a formal parameter patch $\Delta = (\alpha_s, \mathcal{B}{\text{lanes}}, \delta{\ell}, \tau_s)$.src/chronos_mesh/optimizer/): A Two-Phase Simplex + Best-First Branch-and-Bound MILP solver computes globally optimal primal allocations $(q^, y^, u^*)$.Let $\mathcal{L}$ be the set of transport corridors and $\mathcal{S}$ the set of upstream facilities. Decision variables are continuous lane flows $q_\ell \ge 0$, binary charter/corridor activations $y_\ell \in {0, 1}$, and unmet demand slack $u \ge 0$:
\[\min_{q, y, u} \sum_{\ell \in \mathcal{L}} \Big( w_c c_\ell(\Delta) + w_t d_\ell(\Delta) + w_g e_\ell \Big) q_\ell + w_c \sum_{\ell \in \mathcal{L}} F_\ell y_\ell + P \cdot u\]subject to:
REI): For each node $v$ with effective recovery time $\text{TTR}_v$ and buffered survival window $\text{TTS}_v$:
\(\text{Exposure}(v) = R_v \cdot \max\big(0, \; \text{TTR}_v - \text{TTS}_v\big)\)CVaR95): Expected cost conditional on falling in the worst $5\%$ disruption tail across $N$ Monte Carlo scenarios.