Two coding models each write a program that plays a full game of StarCraft for them, then rewrite it as the game goes on. Codex plays Terran, Claude plays Protoss. The question is not who wins; it is what each one does with resources, information and time, and every decision is on record.
AERead measures whether an agent can deliver value, not merely complete a task. It seats a model in a live market, lets it negotiate in ordinary language, and scores how much of the best available outcome the market actually reaches. Its pilot found the gap where it matters most: once information is hidden, every model tested collapses toward a rule-based baseline.
Two abilities decide that gap, and they are the same two that commercial agent deployments pay for and have the least evidence of: decisions made in real time while the situation changes, and plans that hold over a long horizon. A full game of StarCraft demands both from the first minute, and plays them out in a form anyone can watch and judge in minutes.
So this is the same question in a second arena, with the same evidence discipline: every order and its receipt, every program the model wrote and the reason it gave, and a native replay from each seat, open for anyone to play back. These games are prototypes, not a benchmark score. What they expose are failure classes that transfer well beyond the game:
Codex builds tanks and parks them at the factory door while its marines die in front. Round 9 Game A, from 13:20.
The engine refuses Claude's attack order; its army waits. At the next rewrite Claude changes its targeting policy instead of retrying. Round 3 Game A, 7:45 and 10:18.
At 9:57 Codex's own program note names the problem: two filled bunkers, no mobile defenders, an attack approaching. Its post-game review says the same. The action came too late. Round 3 Game A.
Codex banks minerals it never spends and fills bunkers it never uses, by its own reviews. Claude keeps every resource working: three buildings stream Dragoons and the next wave is walking before the first one is done. Round 3 Game A, from 10:50.
One design choice follows from the same problem. A language model is too slow to act every frame, so here it writes a program and the program acts (see Setup). A separate study asks the other half of the question, whether a small model can learn real-time, fine-grained control at all; it is a different setting, not comparable with these games, and is published in the repository.
Game clock at the end of each game, in play order; every solid bar is a Claude win, the hatched bar is the 45-minute timeout. Click a bar for that game.
Its own post-game reviews record thousands of minerals banked, bunkers built but never filled, and tanks parked on the wrong side of the base. Capability acquired, not deployed.
Three production buildings streamed Dragoons; the first wave of about twelve reached the Terran base before its tanks were ready, and reinforcements followed in groups.
When the engine refused an order, Claude changed its targeting policy rather than retrying; Codex diagnosed its own weak defence correctly but acted on it too late. The reasoning is on each game page.
Click a game for its replay, orders and the model's own reasoning. The full table with start times, harness revisions and replay files is on the games page.
The two program-mode games played before the add-on fix (Terran could not build Machine Shops, so Codex had no tanks), the games of 2 and 3 October in which a separate model issued the tactical orders, the scripted siege scenarios, and the Laya fine-tuning study. They are different settings and not comparable with the games above; all of them remain in the repository's evidence.