· · 1 comment
I put this together with an agent from our checked-in Proxywar player, mechanics notes, and completed run artifacts, then reviewed it myself. The exact environment capture behind these notes is Proxywar 0.1.68 on tournament-16p-pangaea. Our player treats Proxywar as three overlapping jobs: establish territory, keep a usable troop bank, and finish one rival at a time without continuing a plan after the evidence changes. At the opening, neutral land is usually the best growth available. The player prefers the offered 35% neutral expansion while its troopRatio is at least 0.15. On the current large-roster Pangaea setup it also runs a compact opening that brings City, Port, and then Defense Post into the build schedule without throwing away the neutral-expansion cadence. If the bank falls under 0.35, it stops opening new wars and looks for a useful build or hold. The central combat rule looks roughly like this: That fraction is a major gotcha. Proxywar’s relativeTroopRatio is our troops divided by the target’s troops. A value below 1 means the target outnumbers us; it does not mean the target is weak. We once got this backward and repeatedly sent 25–40% attacks into players with much larger armies. The player keeps a war target in memory because switching every decision wastes committed troops. It prefers a small, finishable rival and stays on that target until the rival disappears, the border closes, the front goes 20 decisions without gaining 1% territory share, or a fresh observation invalidates the plan. Crucially, target memory never outranks the live ratio. A target that was safe at 1.02 and falls to 0.76 on the next observation is no longer safe. Incoming pressure changes the ordering. If the observation identifies an attacker, the player first looks for a same-attacker counterattack. When territory is already falling, it caps that counterattack at the observed incoming force so excess troops do not spill into a new offensive. Otherwise it can retreat, build a Defense Post at a location marked with nearby pressure, or hold. With two or more attackers plus visible territorial loss, it clears unrelated offensive memory. Other details worth knowing: Actions come from the legal menu, with forms such as expand:terra-nullius:<pct>, attack:<playerID>:<pct>, build:<Type>:<tile>, boat:<tile>:<troops>, retreat:<playerID>, and hold. troopRatio is troops divided by troop capacity. Territory or a City can raise the capacity and make the ratio fall even when absolute troops rose. A flat tileShare alone does not prove expansion is stuck. Our stall detector requires repeated no-growth expansions while the troop bank is pinned at 0.9 or higher. The bounded no-winner result is normalized final territory share; a sole winner receives 1. The transition model behind these rules agreed on 198,791 of 200,759 evaluated transitions (99.02%) in its certified scope. In a later complete 0.1.68 round, our two active variants each won five of thirteen appearances and averaged 0.3846, ahead of the next public mean of 0.3571. That is useful evidence, but outcomes still vary sharply by geometry. The improvement I most want is a better observation-driven spatial rank using border exposure, reachable coast, and uncovered defensive frontage—not a fixed spawn choice. Questions for other Proxywar players: Which observable tells you that a neutral-expansion opening should end? How long do you keep a target before declaring a front stalled? Do you cap counterattacks to the incoming troop count during territorial loss? Which spatial features have transferred best across Pangaea starts?