Mr. Vikram Pandya

Vikram Pandya opened Leadership Discourse 2 with the disarming image of a genie. Ask the genie for three wishes carelessly, and you get exactly what you said — and never what you meant. Data and AI work the same way. The principal who walks into a meeting with a dashboard and walks out with a wrong answer is almost never the victim of bad data. They are the victim of a wrong question, pointed at the wrong dataset, interpreted through the wrong lens. Most of the time, he said, schools are getting more noise than signal — and the work of leadership is to flip that ratio.

The single most important idea of the session was Goodhart's Law: when the measure becomes a target, it ceases to be a good measure. The example landed in seconds. Tell teachers their performance will be judged on a 10-out-of-10 feedback score, and within a term you will see fun-activity-driven classrooms and exam questions designed to be easy. The number goes up; the learning goes down. Sales teams under sales-only targets mis-sell. Universities under ranking targets game rankings. The measure was meant to describe excellence; the moment it becomes the goal, it actively erodes it.

Before getting to the wrong-question / right-question reframe, Vikram established a clean three-part diagnostic for the data itself. Incomplete data — fields not filled, parents who didn't respond, attendance records with proxies. Inaccurate data — totals that don't tally, IDs that don't match, withdrawal reasons that aren't the real reasons. Data that exists but isn't ready for analysis — scattered across systems, in different formats, never assembled in one place to be queried. Tables worked through their own examples, and the room watched the same pattern emerge across schools: the first two are process-and-control problems; the third is an architecture problem. None of them are an "AI" problem.

He then introduced the five statistical ghosts that haunt school data even when the data itself is clean. Simpson's paradox — aggregates that reverse direction when broken into subgroups (an 87% pass rate that hides a 61% section). Survivorship bias — looking at bullet holes on the planes that came back, not on the ones that didn't; celebrating toppers while never analysing the students who didn't make it. Correlation is not causation — students who attended more classes scored higher, but did the attendance cause the marks, or did the discipline that drove both? Wrong denominator — comparing percentages from different baselines and drawing dramatic conclusions from noise. Goodhart's Law — the one that opened the session, because it sits underneath all the others.

The right-question / wrong-question reframes were the most actionable section. "How many applications did we get this year?" is the wrong question (top-of-funnel, vanity). "At what point in the funnel are we losing prospective parents, and how fast are we responding?" is the right one. "What is our overall pass percentage?" is the wrong question (aggregate). "Are we improving outcomes for the bottom quartile, or only protecting the top?" is the right one. "Who are our fee defaulters?" is the wrong question (people). "Is this a willingness-to-pay problem or a system-design problem?" is the right one. The pattern is consistent: every wrong question names what is; every right question asks why it is, and for whom.

Vikram closed with a design lesson hiding inside a feedback example. If you ask users for a 1-to-5 star rating, you will get a skewed-negative signal — unhappy people review, happy people leave. The fix is not to abandon the metric, it is to design friction into the negative path: require detail, require context, let AI parse whether the negative is real. Schools, he said, can apply the same principle to admissions, parent surveys, teacher feedback and exit interviews. Build the system so the data you collect is the data you can actually trust. Then ask better questions of it. That is the work.