apart-research

outlier is an open-source, local-first policy engine and governance framework designed to measure the true cost and risk of AI code generation. Enforce Zero-Trust and protect Human Mastery.

Protecting tech-disadvantaged individuals from scams using the power of AI

Hybrid prompt-injection detection engine (regex · Sentinel v2 · LLM judge) — 95.1% PINT balanced accuracy. Detection core of TUP AIGSMP. Built at Apart Research Global South Hackathon 2026.

Ad-funded LLMs as the default route to commercial secret loyalties — Secret Loyalties Hackathon paper + black-box audit harness.

A dual-control guardrail for Vietnamese public-service, education, and healthcare chatbots (for Apart Research)

Multi-signal pandemic surveillance: attention-based signals (Google Trends, Wikipedia) decay 5-23x across COVID-19 waves but not across flu seasons. Wastewater (NWSS) is the only signal robust across the emerging-disease lifecycle. Apart Research AIxBio Hackathon submission, April 2026.

Attribution-honesty toolkit for detecting secret loyalties in language models using matched controls, refusal-resistant scoring, and weight-delta forensics.

Detection audit of three Qwen2.5-7B "secret loyalty" organisms (Apart Research, Jul 2026). One is byte-identical to base; the other two carry a rank-16, attention-only delta. Ten arms tested for a covert principal — none survived a second control. What we found instead: a replicated safety regression in B model organism.