| 📁 experiments/corpus/ | The 38 synthetic voice memos used throughout Module 2 (scripts and answer key in memos.json, names.md, transcripts from Grok and a second ASR API) |
| 📁 experiments/2_1_stateless/ | Section 2.1: how billed input tokens grow in a stateless multi-turn chat, and the hidden cost of the system prompt |
| 📁 experiments/2_1_workspace/ | Section 2.1: a one-shot script from a chat box vs an agent working in a local workspace (3 independent runs each) |
| 📁 experiments/2_3_names/ | Section 2.3: transcription plus a correction step for names and terms, without vs with names.md |
| 📁 experiments/2_4_skill/ | Section 2.4: first-try pass rate with vs without a cleanup skill (keep the speaker's voice, fix names, format rules) on 28 memos |
| 📁 experiments/2_5_perspectives/ | Section 2.5: evaluation-note skill alone vs with perspectives.md, on 10 proposals that follow industry consensus |
| 📁 experiments/2_6_overgeneralize/ | Section 2.6: a rule distilled from a single edit vs from a batch of past records, applied to work and life memos |
| 📁 experiments/2_8_skills/ | Sections 2.7 & 2.8: putting the whole skill pool in the window vs an INDEX.md with progressive disclosure, plus description rewrites (30 questions) |
| 📁 experiments/2_9_isolation/ | Section 2.9: cleaning up 28 memos in one shared window vs one isolated window per memo |