The claim arrived at 11:47 on a Tuesday, and there was nothing wrong with it. A coronary angioplasty at an empanelled district hospital. Eligibility verified, code consistent with the diagnosis, pre-authorisation valid, amount within the ceiling. Every box a rules engine knows how to tick was ticked. This is where most systems stop reading, and where the interesting part begins.

The first machine had been trained on years of settled cases, each labelled by a human as sound or unsound. It knew the shapes of past trouble – the duplicate, the impossible date, the same procedure billed under two names – found none, and passed the file on. A supervised model is an excellent student of history, and that is exactly its limit. It sees only what it has been shown.

The second had no labels at all. Its task was stranger: to know what normal looks like. It held a statistical portrait of every empanelled hospital – how many cardiac cases........

© Daily Times