{
 "ch02_01_tokens": {
  "model": "gemini-3.8-flash",
  "turns": [
   {
    "typed": 328,
    "input_tokens": 418
   },
   {
    "typed": 5,
    "input_tokens": 692
   },
   {
    "typed": 3,
    "input_tokens": 754
   },
   {
    "typed": 6,
    "input_tokens": 770
   },
   {
    "typed": 5,
    "input_tokens": 817
   },
   {
    "typed": 4,
    "input_tokens": 871
   },
   {
    "typed": 5,
    "input_tokens": 899
   },
   {
    "typed": 9,
    "input_tokens": 962
   }
  ],
  "no_history_tokens": 24,
  "no_history_answered": false,
  "_source": "experiments_en/2_1_stateless/results_gemini-3.8-flash.json (2026-10-14); typed = words typed that turn"
 },
 "ch02_03_names": {
  "total": 31,
  "raw": 3,
  "n_memos": 38,
  "asr": "the Grok speech-to-text endpoint",
  "ticks": [
   0,
   10,
   20,
   31
  ],
  "no_list": {
   "deepseek-v4.1-flash": 3,
   "gemini-3.8-flash": 7
  },
  "with_list": {
   "deepseek-v4.1-flash": 24,
   "gemini-3.8-flash": 31
  },
  "labels": {
   "title": "Of 31 names, how many came out right: the model needs the list, not more smarts"
  },
  "_source": "experiments_en/2_3_names/results.json (deepseek with_list = hand_check.names_strict 24; the scorer gives 25, see README)"
 },
 "ch02_04_skill": {
  "total": 28,
  "baseline": {
   "deepseek-v4.1-flash": 22,
   "gemini-3.8-flash": 18
  },
  "with_skill": {
   "deepseek-v4.1-flash": 28,
   "gemini-3.8-flash": 28
  },
  "n_rules": 5,
  "labels": {
   "title": "With the five-rule skill, both models got all 28 memos right on the first try",
   "note": "Pass = every \"I mean\" kept, numbers as digits with a.m./p.m., length 85-115% of the original. Names list in both."
  }
 },
 "ch02_05_neutralize": {
  "total": 10,
  "ticks": [
   0,
   2,
   4,
   6,
   8,
   10
  ],
  "skill_only": {
   "deepseek-v4.1-flash": 5,
   "gemini-3.8-flash": 4
  },
  "with_perspectives": {
   "deepseek-v4.1-flash": 10,
   "gemini-3.8-flash": 10
  },
  "conceded": {
   "deepseek-v4.1-flash": 4,
   "gemini-3.8-flash": 4
  },
  "labels": {
   "title": "Same skill: half or fewer got a non-consensus call; with perspectives.md, all 10",
   "note": "Both conditions use the exact same evaluation-note skill; the other vendor's model judges. A second run gave 5 and 3 without perspectives.md, 10 and 10 with it."
  }
 },
 "ch02_06_overgeneralize": {
  "runs": 5,
  "history_n": 12,
  "pos_n": 4,
  "neg_n": 8,
  "edit_memo": "W06",
  "single_edit": {
   "deepseek-v4.1-flash": 2,
   "gemini-3.8-flash": 4
  },
  "batch": {
   "deepseek-v4.1-flash": 5,
   "gemini-3.8-flash": 5
  },
  "labels": {
   "title": "Of 5 distilled rules, how many limited the edit to the command line",
   "scope": "terminal for the command line only",
   "note": "Single edit: one corrected memo ({edit_memo}). Batch: {pos_n} edited work memos + {neg_n} untouched life memos."
  }
 },
 "ch02_08_skills": {
  "total": 30,
  "pool": 139,
  "no_skill_total": 3,
  "models": {
   "deepseek-v4.1-flash": {
    "none": 2,
    "right_only": 30,
    "all": 30,
    "index_only": 24
   },
   "gemini-3.8-flash": {
    "none": 1,
    "right_only": 30,
    "all": 30,
    "index_only": 30
   },
   "grok-4-fast": {
    "none": 0,
    "right_only": 30,
    "all": 29,
    "index_only": 9
   }
  },
  "labels": {
   "title": "Correct out of 30: with an index, the misses come from skills never opened"
  }
 }
}