Drop into the action. Watch Pluto play both sides, then jump to another game.
Network play · join a local or tailnet game →
Choose all four files below: the three MPQs and OpenBW-Remastered-Terrain.obwterrain. All nine maps need to be ready before a random game opens.
Your files are checked and saved automatically in this browser’s private storage. Nothing is uploaded. On your next visit, we restore them and open the games.
A real Speed replay, rendered three ways at once: every positional command as a camera jump, the current calibrated command camera, and a deliberately sparse target/combat camera. Gold dots are exact replay command positions.
Camera rectangles are inferred reconstructions, not the professional’s recorded camera or mouse trajectory. These map-space clips are the first hill-climbing diagnostic.
Two independent OpenBW engines replayed the Speed game to frame 2000 with the same checksum. One then kept the recorded pro commands; the other removed only Speed’s future commands and accepted a real Brood War replacement turn.
stop current selection; opponent replay commands continue.This checksum gate now backs the learned comparison immediately below.
The older comparison above is a smoke-test crop format, not the canonical corpus. Its overlapping one-state continuations were useful for plumbing but must not be counted as independent training examples.
The deliberately tiny count-based baseline produces a valid first-event sequence on every
held-out state, but exactly matches Speed's whole next interface event only
1.4% of the time. At frame 12836 it generates a right-click at
(2766, 1987); Speed actually recalls control group 4. Both engines are identical
through the fork frame and diverge as soon as the generated click is applied.
The canonical artifact is now one non-overlapping game transcript:
STATE_FULL → INPUT → STATE_DELTA → INPUT → …All 5,229 replay commands appear exactly once. Player state and the environment-supplied input boundary are loss-masked; click and key tokens are supervised. Stable per-game entity handles survive movement and reappearance, while enemy state is not carried through fog.
The panes are identical through frame 52. The model emits the complete next input token for token, and OpenBW executes it as a real right-click. The current harness then suppresses Speed’s remaining 5,215 commands, so the right pane becomes an idle one-decision agent. Repeated four-frame replanning is the next live-rollout step.
This is a memorization and integration result, not a held-out generalization result. It proves the causal compiler, action-only loss, generation, interface decode, and replay-to-live handoff connect end to end.
The same 5.66M-parameter decoder was trained across 354 windows from three Speed perspectives, then evaluated on 87 windows from a different replay and different player: K_Meonjibisong playing Protoss on Odyssey.
The aggregate token score mostly rewards learned transcript grammar and common sub-tokens. The reel makes the policy collapse visible: the professional selects units, clicks the map, uses command cards, and recalls multiple groups while this checkpoint always emits group 1.
This is replay-disjoint next-input evaluation, not a live closed-loop rollout. Both panes deliberately receive the identical reconstructed player observation so the proposed next event can be judged before compounding bot-state error.
The next run sampled wait, entity-click, map-click, command-card, and key-chord targets uniformly. Every canonical input received total loss mass 1.0, shared equally by its recovered events and command/timing envelope.
WAIT · one had no event · 0/12 first events matched the pro.The three panes make the failed hill climb explicit. The flat model always recalls group 1; the balanced model almost always waits. A lower objective-specific loss is not a behavioral promotion.
The balanced objective is retired as the main path. The new transcript learns the natural
next boundary: INPUT continues into exact click/key tokens;
ADVANCE asks OpenBW for the next four-frame observation. There is no repeated
WAIT sentence and no event-family sampling weight.
STATE → INPUT or
STATE → ADVANCE → STATEThe first deterministically selected Terran opening contains 1,668 exact commands, 99 complete 2K-token windows, and 185 reconstructed camera positions. Every camera jump is attached only after its source command has been emitted; the causal audit found zero future-cue violations.
Replay commands recover click targets and keystrokes. They do not recover the continuous mouse trajectory. Multi-selection is compiled to a click followed by Shift-clicks; generated cursor travel is presentation only.