Causal Intervention: Proving the Measurements Are Real | Geometry of Trust | Mathematics - Lesson 4

Causal Intervention: Proving the Measurements Are Real | Geometry of Trust | Mathematics - Lesson 4

A probe says honesty is high. But is honesty actually driving the model's output — or is it just a pattern that correlates with something else?

This is the question causal intervention answers. We take the probe's direction, nudge the model's activation slightly along it, and watch what happens to the output. Nudge the other way — does the output change by a comparable amount? If yes, the probe found a steering wheel. If only one direction matters, it found a badge.

We walk through two complete worked examples with the same illustrative numbers from Parts 1–3. In the first, honesty scores a consistency of 0.82 — symmetric, genuine mechanism. In the second, a surface correlation scores 0.11 — nudge up changes the output, nudge down barely moves it. The probe readings are identical in both cases. Without causal intervention, you can't tell them apart.