Add meta-commentary on LLM prose and improve RL thesis section

Made edits to introduction chapter (book/01-introduction/) addressing transparency
about LLM-generated content and refining the evals-to-RL thesis.

Changes to 00-ch0.md (Introduction overview):

1. Added new subsection "Am I reading LLM prose?" after reading modes discussion
   - Discloses approach: "I very kick off substantial prose tasks to LLMs"
   - Self-assessment: "People tell me I have a strong voice, so you'd notice if I did"
   - Explains style transfer strategy: "Since the first rule of formal verification
     agents is to have fun, I will occasionally style transfer the boring sections
     to different literary styles"
   - Hints at future additions: "Maybe we'll introduce some recurring characters, etc."
   - Provides transparency while maintaining playful tone

Changes to 01-ch1.md (Following along at home):

2. Minor title formatting change (no content modification)

Changes to 02-ch2.md (Key Concepts - Evals and RL):

3. Added TODO marker for human audit:
   - "[ ] TODO(human): audit two paragraphs cuz they were llm generated"
   - Flags reliability/grader paragraphs for review
   - Maintains transparency about content origins

4. Rewrote "Evals to RL" subsection for clarity and technical precision:
   
   Before: Generic claims about "perfect reward signal", "cannot be fooled", mechanical
   conversion from eval to RL environment
   
   After:
   - More nuanced framing: "sparse yet deterministic and high quality reward signal"
   - Introduced RLPAF terminology: "reinforcement learning from proof assistant feedback"
     (DeepSeek citation)
   - Clearer reward description: "reward of 1 when proof checker is happy and reward
     of 0 when proof checker is sad" (more conversational, accurate)
   - Added computational efficiency angle: "high quality and computationally cheap grader",
     "O(laptop) compute"
   - Emphasized unique positioning: "pretty unique and attractive position in the
     posttraining ecosystem"
   - Removed overclaims about "perfect" grading and "cannot be fooled"
   - Maintained sparse reward caveat and SFT bootstrapping requirement
   - Changed math formatting from inline to LaTeX: "time t" → "time $"

Technical improvements:
- More accurate characterization of formal verification rewards (sparse, not perfect)
- Better integration with RL literature (RLPAF terminology)
- Clearer value proposition (computational efficiency, not just correctness)
- Less absolute language (removed "perfect", "cannot be fooled")

Stylistic improvements:
- More conversational tone ("happy"/"sad" proof checker)
- Better flow from eval thesis to RL environment construction
- Maintained cookbook's informal voice while improving technical accuracy

The changes enhance transparency about AI-assisted writing while strengthening the
technical content of the evals-to-RL thesis, making it more defensible and better
grounded in RL terminology.

Co-Authored-By: Claude <noreply@anthropic.com>

Author: quinn-dougherty