Field notes

Ideas for intelligence people can challenge and trust.

The original notes remain here. They are also connected to projects and downloadable resources in Perspectives.

Governed intelligence

Design the risk envelope before the agent

Autonomy should be a bounded operating condition, not a personality trait.

Start by naming what the system may observe, recommend, change, and never touch. Add confidence thresholds, escalation triggers, spend limits, and rollback paths before choosing orchestration patterns. The result is an agent whose authority can be inspected—and narrowed—without relying on good intentions.

Read as an article →

Governed intelligence

Evidence before eloquence

A persuasive answer is not the same as a decision-ready answer.

High-stakes AI should make its sources, assumptions, gaps, and transformations visible. Grounding is not merely retrieval; it is a chain of accountability from evidence to interpretation to recommendation. If a user cannot challenge the chain, the interface is hiding the most important part of the system.

Read as an article →

Governed intelligence

Human authority is architecture

Human-in-the-loop only matters when the human has real power.

A review button is theater if the reviewer lacks context, time, or the ability to stop the action. Design explicit decision rights, usable evidence, override paths, and ownership for exceptions. The human role should be specific enough to test, train, and audit.

Read as an article →

Governed intelligence

Evaluation is the operating system

The learning loop belongs in the product design, not at the end of it.

Define success measures and failure modes alongside the workflow. Capture disagreement, low-confidence cases, overrides, and downstream outcomes. Those signals should shape prompts, retrieval, policy, and product behavior through controlled change—not silent model drift.

Read as an article →