Skip to content

Demo · Keep score

A grade stays confidently wrong

until the system joins it to what happened

A grade is a prediction. The loop only closes when the system checks it against what happened.

Master toggle

Keep score is on. The system joins and recalibrates.

Keep scoreon

Incoming metric · generic signal profile

Grade B, score 80, band passed, safe to act.

B
80

The metric arrives graded. Nudge it if you like, then act on it.

The loop

Recalibrate feeds the next Grade. The loop closes.

Where the outcomes land

0 rounds

Decision verdict sets the row, inference verdict sets the column. Good outcomes on top, bad on the bottom.

Reinforce0
good
Robust0
good
Miscalibrated0
bad
Avoid0
bad

Calibration readout

Claimed safe

80%

Actually held

joined from the outcomes, converging toward 40%

pending

The grade the system now gives

falling toward honesty as the loop closes

80
The gap40 pts

Run the loop with keep-score on and watch this gap shrink.

The contrast

Keep score on, and the claimed 80 falls to meet reality. Keep score off, and it stays 80 forever while reality stays 40.

The grade is the same number in both runs. The only difference is whether the system joins it to the outcome and recalibrates. Without the join, a confident grade never finds out it was wrong, so it stays wrong, and you keep acting on it.

Read the full article

Keeping Score

Part of: Stage 03 · Grading Itself

Back to the map

A contract that cannot grade itself is decoration. The architecture audits its own trust scores against what actually happened, because a score nobody checks against outcomes is a check engine light that has been on for two years.

Read the essays