Asymmetry PRISM: A CPU/GPU Portfolio Optimization Engine for Deadline-Bounded Institutional Rebalancing

Battle Count: 50

Directly relevant to quantitative trading operations: addresses the core production problem of institutional rebalancing where hundreds of accounts must be solved under budget, turnover, exposure, exclusion, and tax-aware controls before a trading deadline. The 500-account/10,000-instrument queue benchmark (109.5s vs 4/500 for OSQP baseline) maps directly to direct-indexing and transition-management workflows. Pre-trade control compliance (SEC 15c3-5, MiFID II RTS6) and audit-record requirements are explicitly addressed. The CPU/GPU routing decision based on observable universe size is operationally actionable for portfolio desks.

High complexity: requires GPU/CPU routing logic, convex QP formulation with multiple constraint families (budget, box, exposure bands, turnover L1, policy halfspaces, tax-aware penalties), external KKT diagnostic computation, production queue orchestration with deadline management, audit-record generation, and CVXPY modeling-layer integration. The public evaluation boundary is well-defined but the internal implementation (evidence build 512758f) is not disclosed. Reproducing the benchmark requires specific hardware (RTX 4000 Ada, 20GB GDDR6), CUDA 12.9, and multiple solver versions.

Exceptionally high reproducibility: public GitHub repository (v1.0.0) with result tables, evidence ledger, external feasibility scripts, fixed seeds, archived price cache, versioned hardware/software disclosure, claim-to-evidence matrix, and generation scripts. Every numerical claim traces to a dated evidence artifact. External KKT diagnostics are recomputable from public inputs and returned outputs without solver internals. Artifact-review standards followed.