Spearman beats Tank

every pair of civilizations keeps a running grudge and favor score: all 465 pairs in a 31-civ game, plus the biggest individual reasons with the year each one happened. Grudges come from surprise attacks, declaring war, razing or capturing cities, nuclear strikes, broken deals and caught spies. Favors come from gifts, returned cities, honored alliances and fighting a common enemy. Leaders weight events differently and forgive at different rates, so a warlike civ nurses a vendetta while a peaceful one lets things fade. Cross a threshold and the score starts bending real decisions: attitude, trade prices, who gets targeted in a war, who signs alliances and embargoes. The reasons show inside the diplomacy screen, never as popups, and caps plus diminishing returns stop gift-spam from buying permanent friendship. Severity matters: a capital or a wonder city counts for far more than an ordinary town.

there are three ways in. A peace term ending a war, a demand backed by overwhelming power (roughly 3-to-1 on a composite of cities, population, military and economic strength), or a small civ voluntarily asking a big one for protection. Vassals pay a gold tithe in 10% or 20% steps, contribute science from their Libraries, Universities and Research Labs, automatically join their overlord's wars, and give up independent war and alliance diplomacy. Tributaries pay more gold instead, stay fully free, and can never be absorbed. A long-held vassal can eventually be integrated: gradually, at a gold cost per turn, cheapest cities first and the capital last.

every subject carries a loyalty score from -100 to +100 that drifts a little each turn. What moves it, weighted most first: how the subject's culture, science and military power compare to the overlord's; how harsh the contract is; what the Nemesis ledger says about the pair, including wars the subject was dragged into; and whether their governments match. At -50 the subject stops paying its tithe and science, and starts quietly seeking a secret pledge to a rival. At -100, if the power gap has closed to roughly 2-to-1, it can declare a war of independence. Secret fealty: when the rival it pledged to goes to war with its overlord, the subject defects.

vanilla Civ3 has no war planner at all. Every unit picks its own target, which is why AI armies drip-feed single units into your guns. This stage adds a per-civ war coordinator: assemble stacks before attacking, bombard before assaulting, escort the artillery and the transports, retreat and heal damaged units, focus weak border targets, and never take suicidal odds. The same stage fixes workers, scoring every job by its real yield after government penalties (never irrigate a tile Despotism makes pointless), with a full re-audit after every government change. And it fixes city build choices: what the city actually needs first, current-era units only, garrisons sized to the real threat. No more five archers standing guard in the age of infantry.

vanilla warming is a one-way ratchet. Pollution degrades tiles and it is permanent. The new system makes it two-way: the warming level can fall again when world pollution output stays low, degraded tiles slowly step back (desert to plains to grassland), and workers gain a restore action to speed the recovery along. Push warming high instead, and coastal land starts converting into actual sea: land bridges sever, peninsulas become islands, trade routes and borders redraw themselves. City tiles flood last of all, and a city that goes under is destroyed, which is exactly what makes reversing warming worth doing in time.

two copies of the modded game play the same scripted opening from the same save, and after every turn each machine hashes its full game state. Matching hashes mean both machines simulated the identical game. The hashes are split by subsystem, so when something does diverge, the report names which part of the game broke sync instead of just saying 'desync'. Every feature in the mod is built to the same rule: decisions come only from synced game state, never from local randomness, wall-clock time, or per-machine settings.

profile first, fix what the evidence indicts, measure again. The famous suspect, the trade-network calculation, turned out to be only about 0.5 to 3.6 percent of turn time on the biggest save tested, and the C3X Trade Net X path already replaced it anyway. Over 96 percent of late-game turn time is still unattributed, so the next step is profiling a generated 160x160 huge-map game to find where the time actually goes. Every optimization has to be behavior-preserving: the same game, just less waiting.