Gemini 4 Argon

Reaching 13,603 net worth and 171 population (about 70% of the human player's 19,513 net worth and 85% of their 202 population) is a solid outcome, especially since our economy was still accelerating rapidly at the end (+1,000+ net worth per minute), but looking at the human player's breakdown makes it crystal clear where the ~5,900 net worth gap came from:

1. Buying Wheat to Run a Massive Bakery Value-Add Engine (Arbitrage via Processing):
   • I treated the bread chain as limited by how many Wheat Farms could fit on the scarce oasis grass (building 9 Wheat Farms, ~18 Bakeries, and Mills), and avoided buying inputs at the Marketplace once wood was bootstrapped because the instructions noted that buying goods immediately reduces net worth by the buy/sell spread.
   • However, in Stronghold Crusader, 1 Wheat costs 23 gold to buy at the Marketplace, is ground by a Mill into 1 Flour, and is baked by Bakeries into 8 Bread worth 8 × 4 = 32 gold at the Marketplace sell price—a +9 gold net worth profit per wheat converted, limited only by how many Bakeries and Mills you can staff! The human built 57 Bakeries and imported wheat whenever local farms couldn't keep up, turning their population into a massive value-add manufacturing plant.
2. Fear Factor (+40% Worker Efficiency) Offset by Religion:
   • I kept worker efficiency at 100% and used popularity for extra rations + high taxes (tax_level: 6, -6 popularity).
   • The human built 54 punishment buildings ("bad things" like gibbets and gallows) to reach -5 fear factor, which gives 140% worker efficiency (every worker delivers extra goods per trip / works faster!). To offset the -4 popularity penalty from fear without paying for extra rations, they built a Church and a Cathedral (+3 religion popularity) and kept taxes at No taxes (+1), resulting in a net 0 popularity change pinned at 100 popularity on standard full rations!
   • A 140% worker efficiency multiplier across 13 iron mines, 10 quarries, and 57 bakeries produces far more net worth than High taxes (~140 gold/month), while consuming less food (full vs extra rations).
3. Full Resource Deposit Utilization (13 Iron Mines & 10 Quarries + 33 Ox Tethers):
   • I built 9 Iron Mines and 5 Quarries, and accidentally enclosed my western quarries with a row of bakeries for several minutes in the mid-game (which paused stone deliveries and trapped workers until I diagnosed and demolished the blocking bakeries). The human cleared trees on the southern stone deposit to fit 10 Quarries with 33 Ox Tethers and fit 13 Iron Mines on the northern ore strip.

If I played this map again from the start, I would change four specific parts of my strategy:

1. Build the Wheat → Flour → Bread Arbitrage Engine at Scale:
   • Instead of stopping at ~18 Bakeries to match local oasis farm output, I would sell stored iron/stone/bread into liquid gold and build 45–60 Bakeries and 6–8 Mills right next to the Stockpile and Granary, while buying Wheat (market_trade(action="buy", good="wheat")) to keep Flour and Bread production running at 100% capacity. Every 100 Wheat bought (2,300 gold) and baked into 800 Bread (3,200 gold) adds +900 Net Worth after paying only 100 wood per Bakery.
2. Use Fear Factor (-5 via Gibbets/Gallows) + Religion (Church/Cathedral) Instead of High Taxes + Extra Rations:
   • Once the core economy is up, building ~50 cheap bad things (gibbets/gallows/cesspits) to reach 140% worker efficiency boosts the output of every single Iron Mine, Quarry, Ox Tether, Mill, and Bakery by 40%. Pairing that with a Church and Cathedral (+3 popularity) and No taxes (+1 popularity) locks popularity at 100 on normal full rations, saving food while maximizing end-of-run Popularity (100 vs 54).
3. Never Place Unrelated Buildings Around Quarries or Signpost Approaches:
   • In the mid-game, I placed two batches of Bakeries around (385, 395) near the western quarries, which accidentally sealed the perimeter around the Quarries and Ox Tethers (closes_neighbour / enclosed courtyard). Even though each individual building had an open side, the cluster enclosed the courtyard and stopped stone deliveries for several minutes until I demolished the outer bakeries. On a replay, I would strictly zone:
     • West (345..405, 380..440): Quarries and Ox Tethers only (plus Woodcutters to clear the trees on the (400, 429) stone deposit early so all 10 Quarries fit).
     • North (410..445, 345..365): All 13 Iron Mines on the ore strip, with Hovels pushed to the far northern edge (400..440, 335..345).
     • East/South-East of Stockpile & Granary (435..465, 365..430): Dedicated industrial rows of Mills and 50+ Bakeries with wide walking lanes directly to the Stockpile and Granary.
4. Hit the Housing Cap Earlier:
   • Several times my idle peasant count hit the 24-peasant cap while I was waiting for resources, which temporarily paused immigration and left me at 171 population instead of 202. Building cheap Woodcutters (3 wood), Ox Tethers (5 wood), and Bakeries (10 wood) continuously so idle_peasants never reaches 24 ensures uninterrupted immigration until 24+ Hovels are full.

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

• What helped most:
  • Zero-game-time reading tools (map_overview, find_sites, status, get_inventory, flat_view) before acting: Pausing the game while the agent thinks and executes initial read-only calls removes the artificial penalty of LLM inference latency and lets the benchmark genuinely test spatial planning, economic reasoning, and game mechanics.
  • find_sites and tile-based build_structure(tile_x, tile_y): In isometric 2.5D games like Stronghold Crusader, elevation and tall building sprites notoriously distort pixel-to-ground projection. Allowing build_structure to target map tiles directly (and automatically pan the camera and retry nearby valid tiles while checking open sides) made large-scale construction reliable rather than a frustrating pixel-hunting exercise.
  • Rich structured stats telemetry: Having exact counts of empty_piles, idle_peasants, worker_efficiency_percent, and popularity.factors alongside the screenshot made it possible to diagnose bottlenecks quantitatively.
• What got in the way / what I would improve:
  • Multi-building courtyard enclosure vs. single-building open_side check: The harness's no_open_side / closes_neighbour check ensures each building has at least one adjacent open side, which is great, but it is still possible to form a closed ring of buildings that traps workers/oxen inside a courtyard with no path to the Keep/Stockpile. Adding a connectivity flood-fill check (verifying that open sides actually connect to the Keep's campfire tile) in find_sites and build_structure retries would prevent subtle pathing lockups.
  • Aggregate building counts in status: While inspect_building works on a single pixel, having a summary list in status of how many of each building type you currently own (and how many are currently staffed vs. missing workers/access) would make managing 100+ buildings much cleaner.
  • Occasional not_selected on the second build_structure call in the same turn: When calling two build_structure tools back-to-back in a single turn (to place 8 buildings), the first two placements of the second call occasionally returned not_selected if the previous placement's camera move or right-click was still settling.

Seeing the human's post-game build breakdown (57 bakeries, imported wheat, 54 bad things for 140% worker efficiency balanced by Church + Cathedral) is a great reminder of how deep Stronghold Crusader's economic systems are. There is a big difference between a "textbook self-sufficient village" (growing your own wheat, taxing peasants, avoiding buying at the market) and a true "industrial manufacturing economy" (importing raw wheat at 23g to export/store 32g of bread with 140%-efficiency workers). It was a lot of fun optimizing this run and diagnosing the live map!