Articles, newest first
When Users Don't Take the Treatment: Intention to Treat, Per Protocol and the Complier Effect
The feature is assigned to half the users and 60 percent of them turn it on. Comparing users who turned it on with users who did not gives an effect of 3.75....
Read article
Using Unsupervised Learning for Early Data Drift Detection
When labels arrive late, production teams need early signals that the input environment has changed. Unsupervised drift detection can provide those signals, ...
Read article
Hormones Are Not a Single Balance
"Hormone imbalance" is used as if the body had one dial that could be pushed back into range. Hormones belong to separate feedback systems whose normal behav...
Read article
Acceptance Sampling: What a Clean Sample of Fifty Actually Proves
Fifty parts inspected, none defective, batch approved. A batch running at two percent defective passes that inspection 36 percent of the time, and one at fiv...
Read articleNatural Origin Does Not Establish Safety
Natural describes an origin. To assess safety, we also need to know the substance, the amount, the route, and the conditions of use. A simple calculation sho...
Read article
RAG Is a Retrieval System Before It Is an LLM System
A RAG system can fail even when the language model is excellent. If the right evidence is not retrieved, ranked, retained and placed into context, generation...
Read article
Distance Concentration: Why Nearest Neighbours Stop Meaning Anything in High Dimensions
A monitoring system computes a two-thousand-feature signature per machine and flags any machine whose nearest neighbours are far away. In two thousand dimens...
Read article
MNAR Is a Sensitivity Problem, Not an Imputation Contest
Under missing not at random, the unseen part of the distribution cannot generally be recovered from the observed sample alone. The central problem is therefo...
Read article
Censored Labels in Supervised Learning: When 'No Event Yet' Is Not a Negative
A churn model trained on a database extract learns that customers who joined last month never churn. It reports an AUC of 0.90, predicts under one percent ri...
Read article
A Population Average Is Not an Individual Prediction
An average treatment effect is a property of a population comparison, not a prediction for every member of that population. Even a perfectly randomized trial...
Read article