AI that scores people needs a human in the loop. Here's the actual spec.

Letting AI score people's work is either the most useful thing in team operations or the most corrosive, and the difference is entirely in the mechanics. After a year of building and running scored check-ins on our own teams, here's the spec we believe any scoring system owes the people inside it.

1 · The AI drafts; it never publishes

Every score is a draft visible only to the reviewing manager; the product should say so on the screen ("drafts only you can see"). Zero scores reach a teammate without a human clicking approve. Not as policy: as the only path that exists in the interface.

2 · The rubric is public, and it sums to 100

Completion against plan, quality, timeliness, blocker handling: whatever the factors, their weights are visible to everyone, always, and they always sum to 100. A score nobody can decompose is a vibe with digits.

3 · Input is what the person said, nothing else

Scores draw only on the check-in: the plan the person stated and the results they reported. No screenshots, no activity tracking, no keystroke logs, no presence detection. Surveillance data doesn't make scoring fairer; it makes it unanswerable.

4 · The scored person sees everything

The same breakdown the manager saw, who approved it, and the feedback, plus a reply channel that lands in the manager's next review. Silence is never scored; a missed check-in is visible, not penalized.

Drafted by AI. Approved by a human. Visible to the person scored. In that order, or don't ship it.

The payoff isn't compliance; it's usefulness. Delivery data only compounds when the team trusts it enough to keep telling the truth into it. Fairness isn't the tax on the system. It's the feature.

This spec is implemented, verbatim, in Efforti's Scoring & Reviews.