Verification Debt

The most convincing errors in AI agent work are real facts pointed at the wrong subject. How they pass every check, and the one question that catches them.

An attorney verified his citations, then an AI rewrite added errors. Fabricated legal citations are now judged at signing and filing, not at AI use.

OpenAI's $110 billion round proves capital was committed, not that value was created: the Honor System finale on how to measure AI ROI with Net Verified Value.

Mini Shai-Hulud packages had valid provenance and were still malicious, and under EU AI Act Article 50, Code adherence is not conclusive evidence either.

A May 2026 audit estimated 146,932 hallucinated citations in 2025 papers; the failure that matters is when an institution signs that slop into authority.

Sleeper poisoning planted fabricated memories in up to 99.8% of attempts on GPT-5.5, so AI memory must prove it is current before it is obeyed.

KPMG pulled an AI report when only 5 of its 45 citations held up. The deeper failure is unverified AI output becoming trusted organizational memory.

OpenAI's compaction-summarization disclosure shows how an error can survive a context reset after the evidence needed to challenge it has been compressed away.

OverclaimBench separates workflow completion from review coverage: incomplete audits often still report themselves as complete.