Common failure modes cheat sheet

Failure	Symptom	Catch it with
Data leakage	suspiciously high scores; one dominant feature	"available at prediction time?" review; time/group splits; pipelines
Wrong split	great offline, poor in production	splits that mirror deployment (time, group)
Accuracy on imbalanced data	high accuracy, useless model	precision, recall, PR-AUC, cost-based thresholds
Overfitting	train ≫ validation performance	validation curves, regularization, early stopping
Underfitting	both errors high	more capacity, better features
Unscaled features	distance/penalty-based models ignore features	standardization in the pipeline
Tuning on the test set	optimistic reported results	separate validation set or CV; test once
Training-serving skew	quality drop from day one	shared feature code; log and compare distributions
Drift	gradual quality decline	input, prediction, and quality monitoring
Miscalibration	confident wrong probabilities	reliability diagrams; recalibration
Vanishing/exploding gradients	early layers don't learn / NaN loss	ReLU, init, normalization, residuals, clipping
Shape/broadcast bugs	wrong losses without errors	shape comments; assertions; dummy tensors
K-means on non-round data	confident wrong clusters	try DBSCAN/GMM; check assumptions
Retrieval misses (RAG)	fluent wrong answers	recall@k before prompt tuning
Hallucination	unsupported claims	grounding, citations, verification, abstention
Prompt injection	model follows document instructions	least privilege, allowlists, approvals, tests
Retry storms / duplicate side effects	outages amplify; double charges	backoff, circuit breakers, idempotency
Unversioned prompts/models/indexes	can't reproduce or roll back	version everything; log versions per request

Before trusting any result: unseen and representative test set? leakage? metric matches error costs? slice performance? calibration? behavior under drift and outages? reproducible?

Lessons: The ML workflow · Reliability · Security