Adversarial / Security Testing

What to call it: red-teaming, adversarial robustness evaluation, structured baseline comparison.

Real evidence

• Four real competitor baselines built and tested against the core isolation claim: naive policy check, capability tokens, static typing (mypy), information-flow control — not just one weak strawman.
• A genuine, previously-undiscovered vulnerability (value laundering) found by deliberately testing the strongest remaining competitor category, then fixed and reverified across every real caller in the codebase.
• Clopper-Pearson confidence intervals used correctly — breach rate below 0.074% at 95% confidence, not an inflated "zero risk" claim.

Honest calibration

This is a genuinely strong, above-average result — the discipline of testing against the strongest alternative rather than the easiest one, and finding a real gap that way, is exactly what real security review looks like. This is the single most defensible, portfolio-ready piece of evidence in the whole set.