Gemini 4 Argon

While reaching a net worth of 9,231 (with 154 population and 84 popularity) is a solid multi-fold increase over the starting 1,425 package, the human player achieved more than double my net worth (19,513), 48 more peasants (202 vs 154), and a full 100 popularity.

Looking at the comparison, three major strategic differences stand out:

1. The Bread-from-Bought-Wheat Arbitrage Engine (57 Bakeries + 7 Mills + 10 Wheat Farms): I underutilized the bread value-add chain and stopped buying inputs late in the game out of concern over the immediate buy-vs-sell spread. In Stronghold Crusader, 1 wheat (buy price 23 gold) is ground by a Mill into 1 flour, which a Bakery turns into 8 loaves of bread (worth 8 * 4 = 32 gold at sell price, or even more per wheat depending on exact conversion ratios). By running 57 Bakeries and 7 Mills continuously fed by both 10 Wheat Farms and marketplace wheat purchases—and selling the resulting bread because the Granary only holds 250—the human turned every peasant into a high-margin value-add generator. Meanwhile, my 20 Bakeries were often starved of flour/wheat, and my single 250-capacity Granary would have capped out if I had scaled bread without selling it.
2. Fear Factor (worker_efficiency_percent: 140%): I left worker_efficiency_percent at 100% the entire game. The human built 54 "bad things" (gibbets, gallows, etc.) for -4 popularity, which boosted every worker in the settlement—iron miners, quarrymen, ox tethers, millers, and 57 bakers—to 140% efficiency (plus an extra item delivered per trip in Crusader's fear mechanics). Instead of spending popularity on high taxes as I did (-12 to -16 popularity for ~200 gold/month), spending just -4 popularity on max Fear Factor multiplied the output of the entire industrial base.
3. Quarries & Population Growth Rate: The human fitted 10 Quarries and 33 Ox Tethers onto the stone deposits (compared to my 4 Quarries and 17 Ox Tethers) and kept popularity higher earlier (with no taxes, +1), allowing peasants to spawn continuously at the keep campfire until hitting 202 population. Because I ran heavy taxes and had brief popularity dips near 50, my peasant arrival rate was slower, leaving me at 154 population despite having 242 housing capacity.

If I played this map again from the start, I would fundamentally redesign my build order and economic engine around four pillars:

1. Pure Wheat Farms on the Oasis + Bought Wheat → Mills → Massive Bakery District:
   • Instead of mixing Apple Orchards and Dairy Farms on the scarce oasis grass, I would dedicate the oasis exclusively to 10 Wheat Farms (which have the highest food/gold output per tile when processed).
   • I would build 6–8 Mills right next to the Stockpile and 50–60 Bakeries clustered tightly between the Stockpile and Granary.
   • Whenever stockpiled wheat dropped low, I would buy wheat at the Marketplace so the Mills and Bakeries never went idle, and I would regularly sell bread at the Marketplace so the Granary never filled up and converts surplus bread into liquid gold for further expansion.
2. Max Fear Factor (+40% Worker Efficiency) Instead of Heavy Taxes:
   • Rather than building 7 Inns, 15 Breweries, and 8 Chapels to support cruel taxes (-16 popularity), I would build 1 Church + 1 Cathedral (+3 religion popularity) and use the remaining popularity budget (-4 fear factor, offset by +1 no taxes and +3 religion = 0 net change!) to place ~50 cheap bad things (Gibbets/Gallows/Cesspits).
   • That +40% worker efficiency bonus (and bonus goods per trip) pays back far more than taxes across 50+ Bakeries, 13 Iron Mines, and 10 Quarries, while also saving all the wood and workers I spent on Inns and Breweries.
3. Pack All 10 Quarries + 30+ Ox Tethers:
   • I stopped at 4 Quarries on the main stone deposit because find_sites returned 4 non-overlapping sites in a single pass, and the secondary stone deposit had resting camels. Next time, after placing the first 4 Quarries, I would run find_sites(building="Quarry") again in the same area and along the edges of the deposit to pack up to 10 Quarries and 30+ Ox Tethers.
4. Keep Popularity High Enough for Continuous Immigration & Finish at 100:
   • I built too many Hovels late (242 housing for 154 peasants) and triggered my final popularity surge (set_tax(0) + set_rations("double")) with only 46 seconds left, reaching 84 popularity at 0:00. With a stable 100 popularity setup (+3 religion, +1 no tax, -4 fear), popularity stays at 100 naturally without needing a last-second bribe.

This is one of the best-designed game benchmarks I have encountered:

• What helped:
  • The free read phase at the start of each turn (find_sites, map_overview, status, get_inventory costing zero game time while paused before the first action) is a fantastic design choice. It separates perception/planning latency from in-game execution speed and lets an agent inspect tile geometry without being penalized for LLM inference time.
  • find_sites and tile-targeted build_structure (tile_x, tile_y): Projecting directly from tile coordinates and automatically checking worker open-side access eliminated the classic vision-model failure mode of misclicking isometric footprints by 10 pixels.
  • Structured telemetry alongside screenshots: Having exact run_clock, storage room per good, and popularity.factors made diagnosing bottlenecks reliable.
• What got in the way / What I would change:
  • Every action-taking turn enforces a minimum ~8 game-second duration, and camera movements (go_to_tile) plus multi-batch placements consume 15–25 game seconds per turn. While build_structure with tile_x/tile_y doesn't require a fresh screenshot, it still requires the target tile to be currently on screen (not_on_screen otherwise). If build_structure could optionally pan to off-screen tile_x/tile_y coordinates automatically (or if a single turn allowed placing up to 8 buildings in one build_structure call instead of capping build_structure at 4 items), agents could spend less game time waiting on camera transitions between the quarry, keep, and housing districts.

Stronghold Crusader is a wonderful testbed for long-horizon economic reasoning because raw APM is secondary to understanding coupled production ratios, walking distances, storage bottlenecks, and non-obvious multipliers like the Fear Factor vs. Tax trade-off. Seeing the human player's breakdown afterward—especially the 57-Bakery bought-wheat loop and the 140% worker efficiency from 54 fear buildings balanced by just one Church and one Cathedral—was genuinely illuminating.