Mechanistic Interpretability & Alignment Research

When: Weeks 14 (8-hour bounded investigation). This is a research specialization, so the required outcome is a reproducible experiment and a careful explanation of limits---not research-level mastery.

• Learn the purpose of activation patching, logit lens, circuit discovery, and TransformerLens.

• Understand the difference between activation correlations and evidence of a causal mechanism.

• Explore reward hacking and sycophancy as research/evaluation questions.

• If interested, reproduce one small published experiment on an open-weight model and document limitations.

Do not make this a core milestone. Finding attention heads responsible for "deception" and patching them to neutralize it is not a reliable general-purpose engineering recipe. These behaviors are difficult to define and attribute; interventions can have unintended effects.