Reproducible example · farm v4 · simulation
One irrigation rule changes. What actually changes?
Two GardenArena farm policies, identical except for their irrigation intention, complete the same simulated week. At day 7, A retains 600 litres of irrigation water and B 522 litres. The 78-litre difference comes from four accepted actions while preparing day 4. This does not establish an overall winner.
This example separates a policy’s intention, its accepted work and the resulting state. Both programs were prepared by GardenArena; no performance is attributed to Claude, GPT, Grok or another AI provider.
The protocol: one difference, the same farm
Both branches start from the same empty day-0 journal for scenario farm-harvest-365-v4, with identical initial parameters. Each requests seven days with a ceiling of 200 game credits for purchases. This is a spending limit inside the simulation, not the service’s price or a required purchase.
A uses the included “La Prévoyante” policy. B changes only otherwise => dry_only to otherwise => comfort in the irrigation channel. The word comfort names an intention; it does not promise healthier crops.
The engine is lsk-farm-v0.4.0. These reports were computed once for the article. Reading it does not trigger a new simulation or use a player’s execution quota.
Day 3: different intention, identical work
A three-day run detects a different intention but identical commands and final simulated state. The report calls this intent_only. A source change alone is not evidence of a decisive contest.
Preparing day 4: four accepted watering actions
At the first work divergence, A performs no irrigation in that pass. B waters both cereals and both comfrey varieties. These amounts come from the commands recorded by Rust, not estimates from the 3D scene.
| Simulated crop | Water drawn | Game credits spent then |
|---|---|---|
| Wheat | 26 L | 0 |
| Barley | 26 L | 0 |
| Bocking 4 comfrey | 13 L | 0 |
| Bocking 14 comfrey | 13 L | 0 |
The water comes from an existing reserve. Spending zero credits does not mean consuming zero resources. The 3D farmer may illustrate the work, but his animation does not determine these volumes.
Day 7: a result, not an invented ranking
| Metric | A · dry-only | B · comfort intention |
|---|---|---|
| Day reached | 7 | 7 |
| Irrigation reserve | 600 L | 522 L |
| Commands in the journal | 10 | 14 |
| Game credits spent | 0 | 0 |
| Cumulative food-shortage days | 0 | 0 |
| Food-reserve days at day 7 | 7 | 7 |
A retains more water over this period. This table does not demonstrate a harvest benefit for B, agronomic superiority for A or either policy’s resilience over a full year. The export contains full states and intermediate decisions; no such extrapolation is added to the final stocks.
What to check in your own comparison
- The same starting state. Different initial supplies do not establish greater efficiency.
- Actual duration. Three days and seven days cannot be ranked by total spending alone.
- Accepted actions. An intention may produce no work or encounter a refusal.
- Trade-offs. Water, food and purchases do not always reduce to one score.
The first work difference locates a divergence. It does not, by itself, establish the cause of every later difference, especially if several rules have changed.
Reproduce and verify
The proof file includes both sources, the initial request and complete three-day and seven-day reports. To reimport the week into the web comparator, save its seven_days object as a comparison JSON file. An import does not automatically replace a saved farm.
The GardenScript guide publishes policy A verbatim. The MIT-licensed kernel source and garden-husbandry farm-policy-compare command permit local recomputation with the same versions. A hash identifies content, not its author or the truth of a physical observation.
Proof file SHA-256: 4a333f5cbc61ef02e10d66ea4289b5f25d0d25d2e619d220ea1cfb47a94ea504
Binary SHA-256: 06213c3a106223049aecb5ff3efe34d2f9af383c4c0c01834dcabae522d432ac