Drift Detection: When an AI's Values Shift | Geometry of Trust | Mathematics - Lesson 3 A single measurement tells you what an AI values right now. But what happens over thousands of prompts? Are the values stable — or are they drifting? In this talk we build continuous monitoring on top of the ruler and probes from Parts 1 and 2. Same causal Gram matrix, same probes, every prompt. The system builds a statistical baseline using Welford's online algorithm, then watches for deviations. When something shifts beyond a governance-defined threshold, it creates a signed, hash-linked alert that nobody can delete or alter after the fact. We walk through a complete worked example: building a baseline over 50 prompts, monitoring through prompts 51–100, then catching a sharp drop in honesty at prompt 101. The alert fires, the attestation is signed, and the chain creates a tamper-evident audit trail from BASELINE → SNAPSHOT → ALERT.