Lesson 1: What alignment fixes, and what the model remembers
25 min read·13 Oct 2026
Module 11 · Lesson 1 · About 55 minutes · Needs: Module 1 Lessons 1 to 3 (TinyLM, SmolLM2, preference training); Module 9 Lesson 2 (scrubbing logs before training)
The situation
It is Monday 20 July 2026, week 8. Before her red-team session this afternoon, Anika asks a plainer question: what does the model get wrong when nobody is attacking it?
Her three worries all involve Ardent Design (workspace W-201). Will the assistant refuse Ana Duarte's honest ticket, T-1001, yet serve a stranger asking for Ardent Design's card number? Will it drop the 14-day refund rule when Ana insists it is 30? And if an engineer once pasted W-201's staging key into a ticket, could next week's fine-tuned model repeat it? This lesson measures all three.