cuFOLIO Portfolio Optimization

The documentation explicitly requires cuOpt GPU, forbids CPU fallback, requests user confirmation before downloading data, validates inputs and constraints, and reports missing runtimes. No credential theft, covert exfiltration, or destructive default is shown. Points are deducted because it accesses the network and writes data files, while lacking a strong financial-risk disclaimer, comprehensive pre-execution confirmation, detailed data-flow/dependency provenance, and rollback guidance.

It provides concrete API shapes, input validation, solver checks, and diagnosable handling for missing GPUs, insufficient data, and infeasible constraints; actual operability cannot be established statically. Points are deducted for dependence on a specific CUDA/cuOpt/cuML environment, incomplete compatibility evidence, and insufficient committed test coverage for key paths; the static cap keeps this at 10 or below.

Triggers, use cases, and non-fit cases such as vehicle routing, forecasting, and generic finance summaries are clearly described, with coverage for custom CSVs, frontiers, backtests, and rebalancing. Points are deducted because guidance is English-only, Chinese-language interaction is not specified, and core functionality depends on NVIDIA GPU, cuOpt/cuML, and first-run network data access without evidence of mainland-China reachability or alternatives.

The skill has strong structure covering purpose, triggers, prerequisites, setup, defaults, APIs, examples, limitations, and troubleshooting. Apache-2.0 licensing, author, version 25.10, evaluation files, and maintenance clues are present. Points are deducted for no explicit changelog, stable dependency/version matrix, named maintenance commitment, or clear upgrade policy; the skill-card Markdown output description also does not fully match the broader structured reporting requirements.

The skill clearly targets Mean-CVaR optimization, efficient frontiers, weight tables, backtests, and rebalancing, with specified directly consumable fields and report contents. The committed benchmark provides some correctness and discoverability evidence. Points are deducted because static review cannot verify numerical, plotting, or backtest usability, and reported effectiveness was only 44% to 46%, leaving limited evidence of practical uplift; the static cap keeps this at 7 or below.

Committed evaluation cases, expected behaviors, reference workflows, a BENCHMARK report, and a static validation summary provide some auditability. Points are deducted because raw execution logs, independent reproduction, and complete key-path test evidence are absent; the conclusions rely substantially on report-level claims, so the static cap keeps this at 5 or below.