Calibration. Grouped by the last confidence recorded before the thing resolved — and honest about how thin eight items make the buckets.
The problem
Writing “Q3” when you mean “probably, if the migration lands first” throws away the only part anyone could have acted on. The date survives into the slide; the condition attached to it does not.
And because nothing records what you believed at the time, there is no way to find out afterwards whether your seventy per cent has ever meant seventy per cent.
What it does
A window
Not a date. The range you actually believe.
A number
How sure you are, recorded, and scored later.
A reason
Every time the confidence moves, why it moved.
Calibration
Grouped by what you said at the time, against what happened.
What it showed
Landed in the window
4 / 8
Of the resolved items on the example roadmap.
Brier score
0.239
Lower is better. Always saying 50% scores 0.25, so this is barely better than refusing to guess.
The calibration table is the part nobody has. Five items were called at 80–100% and 60% of them happened; two were called at 60–80% and half happened. Both buckets read overconfident, which is the single most common and least examined failure in roadmap language.
The Brier score gives that a floor to measure against: 0.25 is what you score by saying fifty per cent to everything, so any number above it means the confidence you attached was worse than useless. 0.239 is above the floor and not by much.
What it cannot do
Eight resolved items make thin buckets, and the tool says so on the page rather than printing a percentage and hoping. A single outcome moves a row.
It scores the confidence you recorded, not the one you felt. A number typed to satisfy a field is a number the calibration table will faithfully hold you to.
The long version of the argument behind it is Everything is repricing.