← Jason Penrod Available
Note AI governance One page

Govern the loop

A human clicking approve at the end of a pipeline is a signature, not oversight. If you want the review to be real, almost all of the work happens before anything runs. This is the version I keep.

Before you build it

Four decisions and a habit

None of these are features. None of them demo well. They are closer to how you would run a room than how you would build a product, which is exactly why they get skipped.

  1. Write down what the system may decide alone

    Put the boundary somewhere people can argue with it. An unwritten boundary is not a boundary, it is whatever the model did last week.

    Test: can somebody who was not in the room read the line and tell you which side a given case falls on?

  2. Decide what evidence ships next to an answer

    A reviewer looking at output alone is guessing. A reviewer looking at the reasoning, the inputs, and the confidence is reviewing. The difference is what you chose to display.

    Test: can the reviewer say why it decided that, without opening another system?

  3. Define the escalation, then make it cheaper than approving

    Most review tools do the opposite. Approve is one click and flagging means writing a paragraph to somebody who will be annoyed by it. You can predict the outcome from the size of the buttons.

    Test: count the clicks and the keystrokes on each path. If approve is cheaper, you have built a rubber stamp.

  4. Do not measure the reviewer on throughput

    Whatever you measure is the behaviour you will get. Measure a human on how fast they clear a queue and the fastest path through their day is yes. That is not a character flaw, that is the design working as built.

    Test: look at the time stamps. If the median review is a handful of seconds, nobody is reviewing anything.

  5. The habit: sample what nobody escalated

    Read a random slice of the approvals nobody questioned, on a schedule, forever. Every other check looks at the cases somebody already flagged. This is the only one that can find the blind spot, because a blind spot is by definition in the pile nobody looked at twice.

    Test: is it on a calendar with a name against it? If not, it will not happen.

Written by Jason Penrod, 2026. I figure out what to build first, then build it. More of these are at jpcreativesolutions.com/notes, and the work itself is at jpcreativesolutions.com.