chronos-mesh

ADR-0001: Neuro-Symbolic Boundary — Never Let an LLM Solve Constrained Optimization

Context

Global supply-chain rerouting under simultaneous maritime blockades, fab yield drops, and tariff shifts involves coupled linear equalities, supplier capacity caps, fixed charter activation fees, and portfolio lead-time ceilings. Prompting an LLM to output shipment numbers directly violates conservation laws and produces suboptimal or infeasible plans.

Decision

We separated signal extraction (chronos_mesh.neuro) from combinatorial optimization (chronos_mesh.optimizer). Language components emit typed ShockConstraintPatch coefficients, while an exact Two-Phase Tableau Simplex + Bland’s Anti-Cycling Rule + Best-First Branch-and-Bound MILP Solver computes provably optimal allocations verified by Hypothesis against exhaustive brute-force enumeration (tests/test_milp_properties.py).

Consequences