Proving an autonomous decision was right by the outcome it produced, not by whether the run completed.

Grading decisions by the real results they cause, so a passing run that produced a wrong outcome is not counted a success.

The durable, reviewable record that ties a decision to the context it faced and the outcome it produced.

Checking the outcome on every run over time, because a system that worked once does not stay right by default.

Trust in an agent that can be shown to a third party, backed by evidence rather than by a green dashboard.

The space between what a system reports about itself and what actually happened in the world.

A failure every monitoring signal reports as success, because the decision looked correct while the outcome was wrong.

The set of conditions a run must meet to count as done right, defined up front instead of assumed.

The real-world result a decision produced, the thing the business actually feels, not the activity that led to it.

Comparing what an agent decided against what actually happened, to establish whether the outcome was right.

The gap between a system's own prediction of success and success confirmed against the real outcome.

Separating the agent's judgment from the conditions it faced, to answer whether it was wrong or the world moved.

The gradual divergence between what a system used to achieve and what it achieves now, as the world changes.

Evaluation grades output on cases you chose; Decision Assurance grades the real outcome in production. Why you need both.