AI Systems Evaluation / LLM Auditing

What to call it: AI red-teaming, hallucination detection, epistemic verification of AI-generated output.

Real evidence

• Caught a fabricated confidence-fusion result (invented 0.95/0.48 numbers) presented with false confidence — full retraction demanded and given.
• Caught a real, live citation-integrity gap: the same source cited with different numbers in two different internal documents.
• Caught a filename overclaiming its own content (claimed "15 stress tests," actually 13).
• A whole methodology — eight codified, executable rules — for catching exactly this class of error systematically, not just by luck.

Honest calibration

This is a real, emerging, currently under-supplied professional category (AI workflow auditing / prompt engineering / LLM QA), and the evidence here is concrete and repeatable, not anecdotal — worth its own pitch entirely, separate from the technical architecture work, since it's a different skill being demonstrated: directing and auditing AI, not writing code by hand.