A rule that makes the model decline to answer when its inputs cannot support a call — returning no forecast and a reason, rather than a number nobody should act on. It is the reject option from selective prediction, applied to regression on degrading assets.
A prediction system that always answers is easy to build and hard to trust. Real telemetry drops out, goes stale, and arrives incomplete, and a model that keeps producing confident numbers through all of that will eventually produce one that gets acted on. An abstention gate defines, in advance, the input conditions under which no forecast is issued. The output is three-state — proceed, flag, or decline — and the decline is reported as a result in its own right, with the reason.
The gate is defined on the inputs, not on the model's own confidence — which is the point. A model's confidence in its own output is exactly the thing that fails under distribution shift.
If three or more required inputs are absent for the asset at the moment of the forecast, no forecast is issued. The condition is a count, not a judgement, so it is auditable after the fact.
If any required input has not updated within five readings, the same applies. Stale data is more dangerous than missing data, because it looks valid.
Green, amber or red. Red is an abstention: no runway stated, and the reason given. An operator can act on a decline; they cannot act on a silent degradation in accuracy.
In our missing-data robustness study, the gate rejected about 70% of rows and issued no forecast on them.
On the rows it did accept, mean absolute error fell between roughly 1.5× and 10×, depending on how the data degraded. The 10× figure is the worst corner tested — 30% of readings missing, arriving in long blocked gaps. Where the same 30% of readings went missing as isolated single-reading dropouts, the improvement was nearer 1.5×.
The rejection rate is not a caveat on the result. It is the result. A gate that rejects nothing has done nothing, and the value sits entirely in the rows it refuses.
Two limits we state on the public page as well as here. This is a single sample — gap positions were randomly drawn, and the report says plainly that the exact values depend on where they landed, with only the trend robust across scenarios. And it is a claim about the gate, not about our predictor: a gradient-boosted baseline scored marginally better than our model on the accepted rows.
Why this is commercially load-bearing, not just scientifically tidy.
Every false alarm costs an intervention: labour, an outage window, and risk introduced by touching working equipment. A system that predicts aggressively into bad data generates those costs and erodes the operator trust it needs to be useful at all.
It also matters to the parties who price things. An insurer, a lender or a regulator assessing evidence needs to know when a system knew it did not know. A model with no reject option cannot answer that question about its own history.
It is algorithm-agnostic, which is worth noting: the gate sits outside the model and works the same whether the predictor underneath is physics-informed, gradient-boosted, or something else. That is also why we treat it separately from the question of which technique wins.
The reject option has a substantial literature in classification — SelectiveNet and its successors — and a growing one in regression, including work on distributional regression with a reject option and on selective nonparametric regression. Applied prognostics work is much thinner. We could not find published risk-coverage curves for remaining-useful-life regression on named public datasets, which is a gap worth someone filling.
Adjacent and often confused with it: uncertainty quantification. An ensemble spread or a conformal interval tells you how wide the answer is; an abstention gate decides whether to give an answer at all. They are complementary, and conformal methods carry their own difficulty here, since exchangeability is a poor assumption for monotone, non-stationary degradation.
On our own uncertainty language we are deliberately strict: we report an ensemble spread, never a confidence interval, and we label calibration as in progress until coverage is verified.
A read-only assessment runs on data you already have, before any hardware conversation. You see the record; you decide what it is worth.
Choir™ Corp