Monitoring Without Labels

Because the model reduces every input to a small number, so its output is a learned projection of everything it actually depends on. A positive rate moving from 3 to 7 percent overnight means behaviour changed, whatever the underlying cause, and it needs no ground truth, no binning choice and no per-feature multiplicity correction.

When the model is poorly calibrated, because it can then be confidently wrong on novel inputs, producing low entropy exactly where the problem is. The signal fails in precisely the situation it was deployed to catch, which is why calibration quality is a precondition for trusting confidence as a proxy.

Manual override rate. Every time a human rejects or changes the model's output they have evaluated it, and that arrives immediately rather than with the label delay. Escalation rate and appeal rate work the same way. All are business signals correlated with quality that most teams never wire into monitoring.

That \(P(Y \mid X)\) has not changed. Importance weighting, confidence-based estimation and reverse testing all assume it, so none of them can detect concept drift. They estimate the effect of covariate shift and are silent about the drift that matters most, which is the limitation to state whenever their numbers are reported.

None of the usual ones: predictions are stable, entropy is low, inputs are in range. Proxy monitoring needs at least one check that would catch degenerate behaviour, such as prediction variance collapsing or the distinct-output count falling. Stability is indistinguishable from health unless something explicitly tests for it.

Because labels arrive selectively: only approved loans can default, and approval was decided by the model. The eventual labels come from a filtered population, so the estimate describes the model's behaviour on cases it already accepted. Correcting it needs a randomly approved holdout or explicit modelling of the selection.

Because random sampling gives an unbiased estimate with poor efficiency, and low-confidence sampling gives informative labels with a biased estimate. Stratifying with known weights delivers both: efficient labelling effort and an unbiased aggregate, provided the weights are recorded at sampling time rather than reconstructed later.

Because the model reduces every input to a small number, so its output is a learned projection of everything it actually depends on. A positive rate moving from 3 to 7 percent overnight means behaviour changed, whatever the underlying cause, and it needs no ground truth, no binning choice and no per-feature multiplicity correction.

When the model is poorly calibrated, because it can then be confidently wrong on novel inputs, producing low entropy exactly where the problem is. The signal fails in precisely the situation it was deployed to catch, which is why calibration quality is a precondition for trusting confidence as a proxy.

Manual override rate. Every time a human rejects or changes the model's output they have evaluated it, and that arrives immediately rather than with the label delay. Escalation rate and appeal rate work the same way. All are business signals correlated with quality that most teams never wire into monitoring.

That \(P(Y \mid X)\) has not changed. Importance weighting, confidence-based estimation and reverse testing all assume it, so none of them can detect concept drift. They estimate the effect of covariate shift and are silent about the drift that matters most, which is the limitation to state whenever their numbers are reported.

None of the usual ones: predictions are stable, entropy is low, inputs are in range. Proxy monitoring needs at least one check that would catch degenerate behaviour, such as prediction variance collapsing or the distinct-output count falling. Stability is indistinguishable from health unless something explicitly tests for it.

Because labels arrive selectively: only approved loans can default, and approval was decided by the model. The eventual labels come from a filtered population, so the estimate describes the model's behaviour on cases it already accepted. Correcting it needs a randomly approved holdout or explicit modelling of the selection.

Because random sampling gives an unbiased estimate with poor efficiency, and low-confidence sampling gives informative labels with a biased estimate. Stratifying with known weights delivers both: efficient labelling effort and an unbiased aggregate, provided the weights are recorded at sampling time rather than reconstructed later.