Skip to content

Human decisions

North Star treats human judgment as part of the system, not as an exception after automation fails.

The current posture stops for decisions that can materially change:

  • public visibility or release semantics
  • access for another person or organization
  • credentials, authentication, or security policy
  • destructive or difficult-to-recover data changes
  • meaningful uncapped spend
  • product direction when the available evidence does not support one path

The AI can gather evidence, compare options, and prepare a recommended action. It should not hide the decision inside technical language.

A clear request includes:

  1. the choice in plain language
  2. why the system stopped
  3. the expected consequence of each option
  4. the evidence already collected
  5. the default safe state while waiting

The dashboard uses Needs your decision rather than the implementation term gate. The result should be understandable even when the operator is not a software engineer.

Approving one action does not grant permanent authority for unrelated work. The permission belongs to the named project, objective, and protected change.

Next: Limits and recovery