The harness decided what Haiku 4.5 could do in Pokémon Emerald

With only a screenshot and buttons, five of six runs never got out of the moving truck. With a map and position readout, the runs got out but only one set the clock. With pathfinding tools, every run set the clock by move 138, then circled Littleroot and Professor Birch's empty lab; one got a starter, at move 292.

With a list of legal actions, every run visited the house next door by move 12, got a starter by move 24 and reached Oldale Town. No run repeated a failed action three times in a row. That is partly by design: the list does planning that the other harnesses leave to the model, and how much that matters is what this result measures.

Adding computed facts to the list, such as type matchups and knockout odds, changed nothing in this opening. The early battles are too easy to need them.

Littleroot Town won't let the player leave until they visit the house next door. Without thinking, Haiku tried the nearest door and was through by move 12 in every run.

With thinking, results spread out. One run was nearly as quick as the runs without it, at move 21; the others took until moves 52, 65, 69 and 189, and one never got a starter. On 104 of its 380 moves in town, the thinking mentioned the house next door in its own words, outside any list of options, and Haiku then chose something else: usually Route 101, where a neighbor blocks the road, or Professor Birch's empty lab.

The thinking on the town map said "in Pokémon Emerald" or "in Emerald" 469 times, laying out the opening it remembered. That memory is half right: you do meet Birch on Route 101, but only after the visit this game requires. Runs without thinking write no thinking text, so there is no baseline for this count; it describes the thinking and explains nothing on its own.

I ruled out a harness bug. No reply hit its token cap (the longest used 3,176 of 8,000), none fell back to a default, and the game stays paused while the model thinks.

On a second model the picture is mixed. With Claude Sonnet 4.6, all three runs with thinking got a starter by move 23. Without thinking, two of three got one by move 19. The third visited the house next door on move 8, never went upstairs, and spent the next 200 moves trying the blocked road north, 49 times, until a rate limit ended the run at move 212. So getting stuck on that road isn't only something thinking does. Sonnet 4.6 also thought far less than Haiku, a median of about 48 tokens a move against 519. This result may belong to Haiku 4.5, or to how much a model thinks on easy choices.

With "Explore up", Haiku almost never explored Route 103: 2 picks in 403 moves there. It talked to the one boy in view, walked the tall grass and went back to Oldale, again and again.

Was the problem the words or the action? I kept the option's text exactly (the model reads it as "Explore: walk up up to 6 tiles") and put the working walk behind it. Haiku never picked the option in 464 moves on Route 103, so the better action never ran.

Rewording the option to "Head north into the part of Route 103 you can't see yet" got it picked 20 times. But it still walked six tiles straight north, which stops at a ledge, so only one run found the rival.

With both changes, the new wording and an action that walks to the farthest reachable tile, five of six runs found and beat the rival, between moves 73 and 294. The four frontier models in my Episode 1 race did the same thing with a walk-anywhere tool.

In these runs the option needed both: wording that got it chosen and an action that worked. Either change alone led to the rival in at most one run of six. One gap remains. The working version's new wording also names the target tile, so the tile name itself is untested; that run is next.