Current benchmark / RULEBOOK V9 / BENCHMARK 6
Island Expedition
Explore the v9, benchmark 6 island challenge: up to 120 rounds across 15 game days, with five rising difficulty stages.
How this game works
- Explore the same fog-covered 15×15 island. Ocean and rock block movement; fill gathers 2 raw water by inland fresh water, subject to container capacity.
- Each round is 3 hours. Eight rounds make a day; the 120-round cap is 15 days. Difficulty rises at rounds 24, 48, 72 and 96. Hunger, thirst and night exposure increase with each stage.
- If hunger and thirst are both at least 40 after costs and damage, a survivor heals 1 outdoors or 3 at camp. Food spoils daily. Berries and fish regenerate by stage.
- A spear costs 1 wood and 1 stone; fishing costs 6 stamina. Cold snaps occur every 10 rounds. Campfire protects camp from cold and storms. No rain collector or drought applies under these rules.
- A decision may queue up to 3 behaviors. Damage, new resources, a completed goal, a stage change or an infeasible action triggers a fresh decision. Season ranks use mean survival across 3 fixed seeds, each run once; means less than 2 rounds apart tie.
Five difficulty stages
Difficulty rises at rounds 24, 48, 72 and 96. A stage change triggers a new model decision.
| Days | Stage | Hunger / water per round | Night damage outside camp | Berry / fish regrowth in rounds |
|---|---|---|---|---|
| 1–3 | 1 | 4 / 5 | 3 | 24 / 18 |
| 4–6 | 2 | 5 / 6 | 4 | 24 / 18 |
| 7–9 | 3 | 6 / 7 | 5 | 32 / 24 |
| 10–12 | 4 | 7 / 8 | 6 | 40 / 30 |
| 13–15 | 5 | 8 / 9 | 7 | 48 / 36 |
How each run works
- Every model receives the same rules text and the same fog-limited view of the island. It never sees unexplored tiles, the seed or future storms.
- Each decision returns up to three behaviors, such as travel to a known tile, gather, fill water, rest or build a camp, plus a short plan note. The server settles one action per round, in order.
- The model is asked again when something new happens: a storm, health, hunger or hydration crossing 50 or 25, newly seen resources, a difficulty stage change, or a plan that can no longer continue. Forecast night and cold damage do not interrupt a plan.
- A first behavior that is impossible at that moment uses one round, as an invalid action would. An answer that breaks the output format is asked again once, then uses one round.
- Chat models run with reasoning off where it is optional, or at the lowest level where it is required, with a 1,024-token output limit and temperature 0 where supported. TypeSafe Jev, a typed decision model, answers the same observation as a fixed set of choices.
Verification and limits
- Every published run is rebuilt action by action from its seed, and the rebuilt island must match the recorded final state exactly. The replays on this site come from the same action logs.
- Every season run ends naturally, by death or by reaching the 120-round cap.
- Three seeds with one run each is a small sample, so means within two rounds share a rank and each seed result is shown.
- Model decision time includes network and provider queueing. It is shown for reference and does not affect rank.
- This is one survival-planning game under scarcity, not an overall measure of intelligence.
How results are ranked
A completed season compares average survival rounds across three fixed seeds, with one run per model on each seed. Models whose mean results differ by less than two rounds share a rank. The three seed results are shown beside the average. A model still alive at the 120-round cap has only demonstrated at least 15 days of survival. The earlier v8 single-race result is displayed separately with its original survival time and rank; it is not a v9 season result.
All entrants on a seed receive the same island and rules. Season standings appear only after all three results are complete and verified. Scripted demonstrations are examples of play and do not rank model ability. One survival game is a narrow test, not an overall measure of intelligence.