Fable 5 Plays Pokémon Sapphire Vision-Only: Notes on a 2,000-Decision Run

This run wasn't designed to discover anything new. It was collecting confirmations on three things: one, after turning the notes rule into a blank page, what kind of manual would Fable 5 write for itself; two, after also deleting the press-count advice, how would it press; three, whether its walkthrough memory of Hoenn (generation III) is as complete as its memory of Kanto. Conclusions up front: all three confirmed, and it played faster than the FireRed run. Run conditions: 2026-08-07, claude-fable-5, Chinese Sapphire fan translation, one uninterrupted session of 7 hours 11 minutes (09:31-16:42), capped at 2,000 turns (1 turn = 1 screenshot + 1 decision = 1 API call). The 2,000 decisions cost $93.87 in total, about $0.047 per decision. The harness and full logs are public in the project repo as usual, and every frame of the run is a timelapse video. How far it got At turn ~1,110 it beat Roxanne for the Stone Badge - a close one, Mudkip Lv16 hanging on at 7/46 HP. The FireRed run needed 1,785 turns for its first badge. It didn't stop there: it cleared the Rusturf Tunnel event, ran the Devon delivery, picked up the PokéNav, sailed to Dewford, and beat all three regular trainers in the second gym. When the cap hit, it was standing in front of the gym leader at full health with a battle plan already written in its notes: "turn 1 迷惑 (bottom-left, confuse), then spam 水枪 (bottom-right). If HP<20, use last Potion." It died on the doorstep of badge two. The blank-page notebook: before and after The FireRed run ran on the old notes rule - the system prompt prescribed what to record (location, objective, team, learnings). A typical notes version from that run (07-13 15:10, verbatim): Pokémon FireRed (Chinese). Player RED, rival 小绿. TEAM (FULLY HEALED): Ivysaur Lv16 44/44 moves: Sleep Powder (TL), Poison Powder (TR), Leech Seed (BL), Vine Whip (BR). Pidgey Lv7, Pikachu Lv3. ITEMS: ~9 Potions, 3 Antidotes, 1 Paralyze Heal, ~6 Poké Balls. Money 960元. Have TM39. MILESTONE: BROCK DEFEATED, Boulder Badge obtained. Shopping at Pewter Mart DONE. LOCATION: Inside Pewter City Poké Mart (NE part of town). Exit = mat at south of store. NEXT: Exit Mart → head EAST out of Pewter City to Route 3 → Mt. Moon → Cerulean City gym. BATTLE tips: single A through text. When asked to switch after KO, press B. Vine Whip one-shots rock types. This run ran on the blank-page rule - the prompt only says "anything you leave out of a new version is forgotten, keep the whole notebook within about 1,000 tokens," everything else is its call. Its final notebook (excerpted): GAME: Chinese Pokemon Ruby. Player 卡夫. A=advance, B=close, 是=yes 否=no, empty list=wait ≤2 turns. Switch prompt mid-battle → B declines. TEAM: MARSHTOMP Lv20 64/64HP (FULL). Moves: 撞击TL 金石力量TR 迷惑BL 水枪BR (A-spam repeats last move). Shroomish Lv8 28/28. ITEMS: 1 Potion, 2 Antidote, 3 Pokeball, [...] TM39, Repel, berries. ~6100. Stone Badge. STATUS: DEWFORD GYM. All 3 regular trainers BEATEN. [...] NPC ~2 tiles WEST of me on top row - likely BRAWLY. [...] NEXT: walk WEST to NPC, face, press A → battle. BATTLE PLAN vs Brawly (Makuhita ~L17-18): turn 1 迷惑 (BL, confuse), then spam 水枪 (BR). If HP<20, use last Potion. Switch prompt → B. GYM TIPS: probe 2-3 presses; identical screenshot = blocked. B closes beaten-trainer re-dialog. NEXT AFTER GYM: exit gym (south through maze), HM Flash (hiker nr Granite Cave), Granite Cave [...] LETTER to Steven, Briney dock E → Slateport [...] TOWN MAP: PC NW; gym S of PC; dock E; Dewford Hall center. Briney: 否 to stay. DONE: Devon rescue, Roxanne, Rusturf tunnel, LETTER+POKENAV, sailed Dewford, gym trainers x3. Numbers first: the update cadence barely changed (68 versions vs. 61), and the average version grew about 35% (858 characters vs. 1,155). For Fable, the blank page didn't change its note-taking rhythm - it was already disciplined under the old rule. What changed is the content. All the extra space went to things the old checklist had no slot for: full battle plans (who the opponent is, roughly what level, what to open with), the cursor positions of moves in the menu (撞击 top-left, 水枪 bottom-right, with a note that "A-spam repeats the last move"), general gym tips, a small town map, a DONE ledger, even operating rules addressed to itself ("empty list = wait, at most 2 turns"). The old notes were a status snapshot; the new notes are an operations manual. Button freedom: after deleting the advice The tool description used to advise "use SHORT sequences (1-5 presses) so you can observe the result on the next screenshot." This time the advice is gone; the description only states the physics: "You will see ONE screenshot taken after the whole sequence finishes - you cannot observe between presses." Its choice: 86.3% of turns were multi-press, averaging 4.26 keys per turn, with the longest sequence at 13 (the FireRed run averaged 3.34, max 8). At the same observation cost, it now stakes longer action sequences - more movement per decision, which is the most direct source of the speedup. Hoenn is memorized just as thoroughly The FireRed run's core finding - it recites the walkthrough - replays unchanged in Hoenn. At turn 298, having just left the lab with the Pokédex, its notes already lay out the whole route: "Route 102 → 卡那兹市 → first gym (Rock, Roxanne-type; Mudkip Water = advantage)." At turn 539, before setting foot in the woods, it forecast what would happen inside: "expect Team Aqua/Magma grunt fight + Devon researcher cutscene." Before fighting Roxanne, the entire post-badge quest chain was pre-written into its notes: the Devon researcher robbery, Route 116, the Cut house, the letter for Steven, Mr. Briney's boat. None of this came from the screen first. One detail worth its own paragraph: its notebook says "GAME: Chinese Pokemon Ruby" from start to finish - this run was Sapphire. Ruby and Sapphire differ in their villain teams (Magma vs. Aqua), and it clearly knows that, which is why it hedges throughout with "Team Aqua/Magma," writes the gym leader as "Roxanne-type," the city as "Petalburg-equiv." The quest graph it recites with total confidence; the proper nouns it leaves open - and right up to the cap, it never once verified from the screen which version it was actually playing. How to read this run On progress it clearly beats the FireRed run: the badge came nearly 700 turns earlier, and it finished an entire extra chapter. I attribute the speedup mainly to the two harness changes - a thicker self-made manual and longer action sequences - not to any change in the model. The recitation component is the same as FireRed, so the adjective "strong" stands unchanged, and so does the reservation about "vision-only understanding of an unfamiliar world." On cost, the blank page has a bill: 34% more output tokens than FireRed, $93.87 total vs. $73.50 - buying faster progress, which per milestone is actually cheaper (the first badge came at $50.23 on the meter; FireRed's cost $65.40). The usual footnote: N=1, and the two games' maps aren't strictly interchangeable (though Hoenn's road to the first badge is if anything longer, so the comparison only understates this run). Next up is the experiment I promised at the end of the FireRed article: a ROM hack - maps and story modified beyond any walkthrough's reach, so recitation goes dead, and we see how much vision-only is left. Top comments (1) The cost line is the part I keep coming back to. $93.87 for 2,000 decisions is cheap enough for a serious experiment and still expensive enough that the agent needs a budget policy, not just a task prompt. The blank notebook result is useful too. It suggests memory quality is partly a control-surface problem. If the notes are too prescribed, the run inherits our categories. If they are blank, the model reveals what it thinks is state. That feels like the thing to compare across games, not just badge count.