Lecture 12 moves from perplexity to exams, chat preferences, agents, reasoning, and safety. Every benchmark changes the capability definition, scaffold, judge, and contamination risk, so evaluation must first say whether it compares a method, model, or complete system.
Only four drone occurrences have entered Taiwan's official aviation safety statistics — because the threshold is 'substantial damage to a drone over 25 kg.' A hobbyist crash never enters the count. The two published investigation reports are the same model, the same manufacturer, and the same agency, and both probable causes were hardware failures: main rotor servo electrical failure in one, a fractured tail rotor pitch link in the other. In both, the flight control computer and the operator were explicitly cleared.
Checkpointing is not git commits: Claude Code snapshots your files before every user prompt, keeps the 100 most recent per session, and deletes them after 30 days. This piece breaks down the five rewind menu options, the tracking boundaries (bash, subagents, symlinks), and how checkpoints divide labor with git.
The attacks RAG systems face go beyond the technical level — Prompt Injection and Jailbreak are real threats. Both inputs and outputs need independent protection layers.