Research Notes
Research notes
Engineering decisions, safety design notes, and technical write-ups from the team building Ophraxx AI.
Designing the S1–S9 threat model: how we categorized safety risks
A research note on the decisions behind Ophraxx AI's nine-category threat model — why each category exists, how the safe-topic allowlist prevents over-refusal, and the reasoning behind the S8 and S9 categories that cover jailbreaks and persona abuse.
Routing complexity without asking: how the prompt classifier works
A research note on the design of Ophraxx AI's prompt classifier — the scoring system that automatically routes every query to the right model tier without requiring users to choose, and the tradeoffs behind the thresholds we selected.
Session memory design: why we chose 15 minutes and 20 messages
A research note on the design of Ophraxx AI's conversation session system — the tradeoffs behind the 15-minute session timeout and 20-message history cap, and why sessions are stored in memory rather than persisted to a database.