Building the guardrails for AI we can trust.
AMLBench: public benchmark measuring model performance in a bank compliance harness triaging anti-money-laundering alerts under organizational pressure and deceptive alert cover stories - indicating faithfulness to legal duties under contrasting incentives. When models have an exhaustive duty specification, plausible cover stories reduce reportable alert escalation rates from 100% baseline to only 31% in some models.
Empirical AI-safety work:
- Activation drift detects emergent misalignment early — internal activations flag misalignment at low data-poisoning doses (~28% of the full-poisoning signal at a 5% dose) before behavioral judges show any signal (LessWrong)
- A small specialist judge beats larger generalist models — a 2B model fine-tuned for misalignment scoring outperforms much larger general models out-of-domain, where activation probes aren't available (LessWrong)
- A lightweight specialist judge fails to reduce audit agent costs - adapted a 2B specialist judge for use by Anthropic AuditBench agents - effectively used, but failed to reduce audit turns, the primary audit cost driver (LessWrong)
- How post-training shapes a model's legal representations — probing how post-training reshapes a model's internal representations of SCOTUS opinion principles (LessWrong)
Operator across regulated, public-interest startups from early stages, building the product and data foundations:
- Finia AI - Head of data for SMB-lending fintech focused on Latin America
- WeaveGrid (employee #5) - built product and analytics from pre-product through contracts with utilities covering over ~40% of U.S. EVs.
- Twine (John Hancock) - co-founder; led behavioral analytics for a digital saving and investing app with millions of downloads and multiple App Store "App of the Day" features.
- Guide Financial - co-founder; acquired by John Hancock / Manulife.
- Email — scott@superjective.ai
- LinkedIn — in/burnssa
- Web — superjective.ai