Abstract
The rules passed G21 (alone, kite the one that shot you) on 24.5% of its seeds. Read hit by hit, G21 was lost on the direct gun: the kite waited on the mortar's ring, the mortar tied the gun up 12 s, and the step in's window was open a quarter second of each enemy reload. Of three designs, kite_in_alone was kept: with no friend near the target, the step in ends six tiles from it. G21 rises to 35.5% (+8.3 ± 2.6 points paired over 600 seeds), G22 is unchanged, the census reads −1.12 ± 0.92 a seat, inside the noise, and the families hold.
I. Intent
Paul: "its risky, we will recover. i think it indicates faster training." Then: "good, commit when the census holds, then the ring sensor." And: "write a lab report about these redesigns."
II. Decision log
| Decision | Cause |
|---|---|
| G21 read hit by hit before any change | a third of the losses could be the test's and not the rules' |
drones = false added to the scenario form | to read the tanks' own duel; it held 1 seed in 200, and showed kiting needs a recon's eyes |
| G21 keeps its drones | without them no rule can kite: sight and reach are both 10 tiles |
duck_alone dropped | no effect: a gun first seen inside its reach cannot be ducked |
| Holding the mortar dropped | G21 24.5% to 19.0%: its splash hits count |
kite_in_alone, not kite_in moved | kite_in 7 gained G21 a third and cost G22 a third: the boundary |
| 6, from a sweep and 400 held-out seeds | best on both; paired +7.0 ± 3.2 on the held-out alone |
| The enemy in G21 and G22 pins 2.25 | with 6 on every brain G21 read 34.5% and G22 50.5%: a yardstick must not move with the rules |
| G21's floor 0.25 | today's success, 35.5% less two errors of a hundred seeds |
III. Method
G21 and G22 for the rules (make scenario, a commander in seat 0 with the dial in its own table), 200 seeds from 9,500,000; G21 again on 400 from 9,700,000, read seed by seed for the paired difference. The rule before the census: not worse than two errors on 48 seeds of every map, the families holding, the other gauntlets the same. Census on this Mac, JOBS=8, the paste shim under bash. The families with SCENARIO_SET="kite_in_alone=6". The gauntlet with GAUNTLET_WEIGHTS, then as built with the default moved. 165 brain tests.
IV. Results
kite_in_alone | 2.25 | 5 | 6 | 7 | 8 |
|---|---|---|---|---|---|
| G21, 200 seeds | 24.5% | 30.5% | 35.5% | 32.5% | 29.5% |
| G21, 400 held out | 24.5% | 25.0% | 31.5% | 28.2% | |
| G22, 200 seeds | 52.5% | 52.5% | 52.5% | 52.5% | 52.5% |
Paired, 6 against 2.25 over 600 seeds: +8.3 ± 2.6 points; deaths 48.5% to 40.2%. Census −1.12 ± 0.92 a seat (europe −11.76 ± 4.78 on four seeds, the one map outside the noise). Families hold. Every gauntlet floor holds.
V. Discussion
The kite was right in its idea and wrong in two numbers that the game had moved under it: the mortar's reload and the step in's depth. Split by whether a friend is near, the depth is the boundary Paul asked for.
VI. What is not built
G22 taught to the rules; the ring as a sense for the nets; whether the waiting ring should stay at the mortar's minimum.
VII. Paul's decisions
None outstanding.