05
Engineering principles
The boundaries I use when designing governed AI and distributed workflows.
01
Deterministic core, probabilistic edge
Use models to explain, classify, summarize, or route. Keep authorization, scoring, and state transitions explicit where risk demands it.
02
Auditability by construction
Treat evidence, citations, and decisions as first-class records rather than trying to reconstruct them from logs later.
03
Durable state over ephemeral memory
Design agents and workflows to survive retries, cold starts, and asynchronous boundaries without losing investigation identity.
04
Minimize model exposure
Send only the information required for a model task, preserving explicit boundaries around sensitive source data.
05
Human authority remains explicit
Automation may propose, route, summarize, or assist; privileged decisions retain clear ownership and review boundaries.