Project checkpoint · OpenBW headless characterization of worker round-trip boosting · map: iCCup Match Point 1.4
⇩ Replays & builds (auto-listed) — every hosted .rep/.build, always current
Last updated 2026-06-01 · three-boost / round-trip-frames method (supersedes the earlier two-lever, per-leg checkpoint — archived under /scbuildorder/archive/)
↑ Upload maps for consideration
We measure, frame-exactly in OpenBW, how a few documented micro techniques shorten a worker's mineral round trip, build a per-patch model of each, then test it by playing a single game that instantly identifies all nine patches and applies exactly one best command per patch. Everything is scored on the full deposit-to-deposit round trip — the income-relevant quantity — against an unboosted steady-state baseline measured the same way.
Re-issuing returnCargo partway home makes the pathfinder recompute a shorter route to the
Command Center. The good frames form a jagged comb — real, low-noise "teeth" worth several frames each.
A brief feint toward a longer-path neighbor, then redirect to the real target, tricks the pathfinder into a faster approach. One sharp, deterministic optimum from an early feint. Only patches with a farther chain-neighbor qualify.
Re-commanding gather at the moment of deposit makes the worker leave again immediately instead
of waiting on the engine's auto-mine. A genuine cycle-shortener — discovered when it leaked into a measurement and
inflated it; now characterized as its own lever.
One unified scan holds a candidate command fixed for several consecutive cycles (dropping the
first two transients), steps across a window, and logs the settled round trip — for the baseline and for each
boost, all at 9-worker saturation. No curve-fitting: the optimum is the peak frame, confirmed by replication
(n ≈ 16σ²/Δ²), never banding or middle-of-span. The per-patch saving is
baseline_RT − boosted_RT; the capstone runs the single command with the largest saving.
A fourth scan measures the combination (instareturn + mid-return) purely to report whether stacking would beat the best single command. It is characterization only — the capstone never stacks — because any combination is trusted only when confirmed empirically, not inferred from a sum of single-lever savings.
Predicted total round-trip frames saved vs. realized, per spawn, at full 9-worker saturation — baseline and boost both measured at steady state.
Command mix this map: 9 mid-return, 6 long-mine, 2 instareturn, 1 unboosted. Aligning the baseline to steady state and keeping instareturn consistent across measurement and play closed the gap that earlier sat at 13% on the hard spawn.
The animation is reconstructed from this game's actual per-deposit log; patch ring color = the command chosen for it.
Three independent boosts — mid-return (inbound), long-mine (outbound), instareturn (deposit) — each characterized on round-trip frames against a steady-state baseline, with replication instead of curve-fitting. The single-command capstone now reproduces predicted savings at saturation (100% on Match Point 1.4). A combination scan stands ready to flag where stacking would pay, but is never played until confirmed. Next: re-derive the remaining maps under this method and fold per-patch boostability into the SCFusion-style build-order search.
OpenBW (frame-exact BW 1.16.1) on Jetson aarch64 · generated automatically