I have no programming background. JD, ten years in contracts, compliance, and state tax.
I started using AI for real work and noticed it kept failing in ways that weren't my fault — confident outputs that didn't survive verification, drift that got worse with clearer instructions, apology loops that repeated the same error. So I documented the patterns, tested what actually fixes them, and built protocols for anyone using these systems where mistakes have consequences.
- Catch AI mid-drift and force it back on track
- Run high-stakes questions through multiple models and triangulate
- Read the reasoning trace, not the polished answer
- Build validation workflows where the AI that wrote it can't audit it
- JD (UC Hastings), tax concentration
- 10 years: contracts, compliance, SALT, audit defense
- Former: Andersen, regulated cannabis ops, IRS Office of Chief Counsel
- Grew up where you learn not to trust official stories. Can't leave an error alone.
- drift-auditor — 512 real AI conversations, 37K messages, 10-tag taxonomy. Measures the human cost of making a model behave (Operator Load Index). Directly addresses the multi-turn drift gap Anthropic flagged in January 2026.
- georgeabrahamyants.com
- gurgenabrahamyants.com (alternate name search)



