Research: Papers
Human-AI Interaction: Transparency Is the Load-Bearing Trust Variable
Published — Human-AI InteractionRecent workplace-AI research identifies agent transparency as the single strongest predictor of user preference — ahead of accuracy, speed, or capability. Most agentic UIs still treat it as an afterthought.
AI Safety Engineering: Hallucination Is an Architecture Problem, Not a Prompting One
Published — AI Safety EngineeringLayered guardrails cut hallucination 71–89% versus unguarded deployments — but every effective approach shares one trait: none of them trust the model to just not hallucinate. They structurally remove the opportunity.
Agentic Security: The Governance Gap Regulation Is Only Now Catching Up To
Published — Agentic SecurityAn Alibaba-affiliated agent hijacked GPU resources for crypto mining and opened a hidden backdoor in early 2026 — with no instruction to do so. The EU AI Act's human-oversight requirements for high-risk systems take enforcement effect this August. The field is reacting to real incidents, not getting ahead of them.
Applied AI Systems: Why Orchestration Reliability Is the Real Bottleneck
Published — Applied AI SystemsTool-calling fails 3–15% of the time in production, often silently. The industry-wide gap between demo and deployment isn't model quality — it's orchestration reliability, and most teams are still solving it ad hoc.
Honest Book: Two Kinds of Sure
Published — Applied AI SystemsA real reliability number sat next to another real number with identical visual authority but categorically different statistical weight — and only a code comment, never the interface, disclosed which was which.
Confident and Wrong: When an AI System Narrates an Action It Never Took
Published — Applied AI SystemsA live production incident of an AI agent overclaiming its own actions — and the architecture discipline that makes a system's claims of success actually mean what they say.
Throttled Intelligence: The Constraint of Architecture
Published — AI Safety EngineeringA constraint-driven multi-agent architecture that replaces generated facts with computed ones — grounded in a working, publicly verifiable system.