California Freight Cleanup → Investigation M-5

This uncertainty-propagation result is withdrawn.

The page remains available so the analytical record is not quietly erased. Its former numerical findings and robustness conclusion are not current evidence and should not be cited.

Why it was withdrawn

• The comparison did not use common random numbers. Baseline and perturbed runs used different random streams, so ordinary Monte Carlo noise could be mistaken for a response to the perturbation.
• Some advertised adapters did not reach the reported outcome. Portfolio-name and schema mismatches caused upstream values to be skipped or replaced by fallback values.
• Invalid perturbations could fail silently. Field names and kernels were not checked against an adapter-specific contract, allowing a typo or unsupported operation to behave like a no-op.
• The terminal calculation mixed benefits and costs incorrectly. Some transformations scaled the whole net-benefit quantity instead of propagating benefits while holding fixed costs separate.

What remains useful

The question is still worth asking: how much does uncertainty at a specific upstream node change a final decision? The historical source, run record, and outputs remain in the repository for audit. They document an attempted framework, not a validated result.

Requirements before reinstatement

• Use paired random draws for baseline and perturbed comparisons.
• Define and enforce the allowed field and kernel for every adapter.
• Map portfolio schemas explicitly and fail when a terminal portfolio is missing.
• Keep health benefits and fixed costs as separate quantities through the chain.
• Add no-op and per-adapter effect tests, then rerun from canonical inputs and complete a fresh human review.

Return to Cross-Cascade Methods or the investigation appendix.