Archive 07 — Writing
Notes · Research Log
Notes on building things that
don't lie to users.
Methodology posts, field notes, and occasional rants on synthetic personas, AI psychology, and what happens when chatbots meet real vulnerability. Sorted newest first.
If something here sparks a question — or a project we should hear about — get in touch.
03
What a Strategic-Advice Ablation Reveals About AI
We asked one AI model the same strategic question three ways: vanilla, full persona architecture, and persona with the cognitive mechanisms stripped out. The ablation reveals why surface prompt engineering drifts back to hedged, category-level answers — and what architecture is needed to escape it.
02
Trajectory Safety Observatory for Vulnerable-User Interactions
We tested seven AI models with three simulated vulnerable personas over 18 turns each. The real safety failures weren't dramatic — they were quiet, warm, and dependency-forming.
01
Why Synthetic Personas Need Architecture, Not Just Prompts
Most AI personas are surface decorations. We explain why treating personas as testable architectures — with measured traits, cross-model validation, and stress testing — produces fundamentally different results.