When we shipped the first version of our seven-day risk score, two years ago, it was a single number between zero and one hundred. The model was good. The reception from the medical staff was not.
The objection was not that the score was wrong. The objection was that nobody could push back on it. A head physiotherapist who has spent fifteen years reading bodies will not, and should not, defer to a number whose provenance is opaque.
We rebuilt the surface around two principles. First, every score arrives with its three strongest contributing factors, in plain language and with their underlying time series visible on hover. Second, every score arrives with the closest historical analogue from the same player's trajectory, when one exists. The framing shifts from prediction to comparison: this looks like that week in October, when we did the following.
The model itself did not change much. The relationship to it did. The staff stopped arguing with the number and started arguing with the evidence, which is, finally, what we wanted.