When prediction is unavailable or wrong.
WHAT CAN FAIL
The model service being unavailable Features being unavailable Input being invalid Inference timing out The model producing an implausible output
WHAT TO DECIDE PER FAILURE
Whether to fail the request, or proceed without a prediction.
WHAT FALLBACKS EXIST
A simpler model A rule A default value The previous prediction Declining to predict
WHAT TO ALWAYS RECORD
That a fallback was used, and why.
WHY
Otherwise silent degradation is invisible.
WHAT AN IMPLAUSIBLE OUTPUT IS
A prediction outside reasonable bounds.
WHAT TO DO
Define bounds, and handle violations explicitly.
WHAT CONFIDENCE THRESHOLDS PROVIDE
Declining to predict when uncertain.
WHERE THAT SUITS
Decisions that can be referred to a person.
WHAT TO DESIGN
The human path, for referred cases.
WHY THAT MATTERS
A system with no human path fails entirely when it is uncertain.
WHAT TO MONITOR
The rate of fallbacks and referrals.
WHAT A RISING RATE INDICATES
Something changed, upstream or in the population.
