The ledger. Hit, unclear, missed and overruled — per person, and with the counts shown rather than a rate the sample cannot support.
The problem
Most launches are argued about beforehand and forgotten afterward. The argument is detailed, confident and expensive; the follow-up is nobody’s job.
Which means a decade of launches can leave you with a decade of experience and very little judgment, because the loop that turns one into the other — predict, ship, check — was never closed.
What it does
Predict
The claim, the metric, the target and the date, recorded before the thing ships.
Baseline
Where the number started, so the verdict is not argued from memory.
Settle
On the date, against the data. Unclear counts as an answer.
Record
How good these predictions have actually been, per person and overall.
What it showed
Settled predictions
8
Of eleven recorded. Three still open, one with no baseline.
Settled cleanly as a hit
13%
Against 88% unclear. The tool’s own note: that is about how the targets were written.
The uncomfortable result is not the hit rate. It is that most predictions did not settle at all: a target that cannot be checked against a number on a date was never really a prediction, and eleven records produced one clean verdict.
That is the finding a ledger is for. Nobody would have said in the meeting that the success criterion was uncheckable — it only becomes visible when something tries to check it.
What it cannot do
The tool says this about itself, on the page: the spec calls for a server that evaluates predictions on their date with nobody present and emails the accountable person. This runs entirely in the browser, so the arithmetic, the ledger and the record are real and the schedule is not.
It also refuses to show rates below five settled predictions and shows counts instead. A confident percentage over four items is the thing that makes a ledger worth ignoring.
The long version of the argument behind it is Everything is repricing.