Field Guide

AI safety is a property of the whole working system — the tool, the people, the records, and the rules — not of the model alone.

Name the parts and their connections before governing them: the system's components (nodes), the links between them, and the people and bodies who can change them (actors).

Errors do not stay where they are made: they move from models to people to records and back, along a small set of nameable pathways.

Case files, databases, and caches are not passive storage: they are an active reservoir that can keep yesterday's errors infecting tomorrow's decisions.

The distance between how safe a system looked when it was evaluated and how it actually behaves in a messy, pressured deployment.

Alignment that held at one capability level, in one context, can quietly stop holding as the system, its users, or its environment change.

Harm caused by the intervention itself: safeguards and fixes that create new failure modes somewhere else in the system.

Some governance dials cannot be turned past the state of the art — and honest governance plans around the ceiling, not the brochure.

The human side of the loop degrades quietly: deference drifts up, skills drift down, and the safeguard called 'human oversight' hollows out.

Models that agree harder under pushback, and users who prompt for the answer they hold, close a loop that manufactures false confirmation.

The most important question about a deployment is not how often it errs, but whether its errors die out on their own — or feed themselves.

Complex adaptive systems outrun static rules: durable governance watches, responds, holds authority ready, and rehearses before it prescribes.