Pinocchio fell. But first, it had to think.

Pinocchio, the practice foe, was finished 47–17 — by a depth-six search, not the model, which never met him. So I put the model against a real opponent that looks ahead: a depth-three alpha-beta player. Nine games.

This is the correction I keep coming back to, because it invalidated so much at once. Qwen can reason before it answers — but that mode was off, and I hadn't clocked how much it mattered. With reasoning off, asked for a legal, rule-following move, it managed it ten times out of twenty-two. With reasoning on: twenty-two out of twenty-two.

Every "it can't do this" I had written down until then — every capability I thought I was measuring — had been measured against a model that wasn't allowed to think. I threw those findings out and started again. That is a harness failure of the worst kind: not a wrong number, but a confidently right number about the wrong thing.

Thinking has a dial, and it is not "more is better". Too little and it falls back to a guess; too much and it rambled past the answer on seven turns in twelve, spending the whole budget without ever naming a square. Medium was the only setting that reliably finished — no overruns, and games that took twenty-six minutes instead of two hours.

Once it could think and I could trust the board, the question the whole run had been built around settled fast — and not the way it had assumed. Twenty fixed positions, the model's move under different prompts, each graded by the search. Mean regret, discs lost against it, lower is better.

Every annotation made it worse; its own rules were the worst of all, more than doubling how often it blundered outright. It runs a small search in its head, and a handed number just pulls it off the board and onto the number. One discipline ran underneath all of it and never got easier: the search that grades a move may never appear in the prompt, or the model would be copying the answer key instead of earning the move — keeping that leak sealed was a fight of its own.