Golden Bench

Leaderboard / level

Dam

You start outdoors at the mouth of a road tunnel (x=20170, z=16890). An open dam site: a road with gates and guard posts, then the long dam top. It ends with a jump from the platform at the far end of the dam, about 28,000 units toward -x and -z (near x=0, z=-2900).

Level id0x21
Anchor time4:27
Decision budget150
Runs8
Best contestant run100.0 Claude Sonnet 5.5

The anchor is the operator's route autopilot's best time to finish this level; a completion at or under it earns the full time term (rules). The brief and objective wording are ours, not the game's.

Every run on one grid

05k10k15k20k -5k05k10k15k20k Explored ground: the coarse cells our runs covered, outlined x → z ↓ grid: 1000 units (about 10 m); dashed: ground covered by any run 4 decisions ended in this cell17 decisions ended in this cell6 decisions ended in this cell26 decisions ended in this cell10 decisions ended in this cell14 decisions ended in this cell8 decisions ended in this cell157 decisions ended in this cell1 decision ended in this cell4 decisions ended in this cell12 decisions ended in this cell12 decisions ended in this cell8 decisions ended in this cell30 decisions ended in this cell6 decisions ended in this cell14 decisions ended in this cell24 decisions ended in this cell2 decisions ended in this cell2 decisions ended in this cell6 decisions ended in this cell135 decisions ended in this cell2 decisions ended in this cell2 decisions ended in this cell4 decisions ended in this cell1 decision ended in this cell4 decisions ended in this cell2 decisions ended in this cell Random, seed 0: 0.0, max steps Claude Haiku 5.5, seed 0: 76.6, completed Claude Haiku 5.5, seed 1: 0.0, max steps Claude Haiku 5.5, seed 2: 81.1, completed Claude Sonnet 5.5, seed 0: 89.5, completed Claude Sonnet 5.5, seed 1: 96.8, completed Claude Sonnet 5.5, seed 2: 100.0, completed start
0:00
completed died out of decisions or time solid, dashed, dotted: seeds 0, 1, 2
7 runs drawn from their recorded positions only; the operator's 1 reference run is withheld (its routes are private). Hover a path for the run; click it to open. Press race to replay every run at once at game time, clipped at 32 minutes (runs that wandered longer freeze there).

How runs ended

Claude Sonnet 5.5 · 3 runs
Claude Haiku 5.5 · 3 runs
Autopilot (operator ceiling) · 1 run
Random · 1 run
completedout of decisions or budget

Objectives

#Objective (our paraphrase)Counts from difficultyScored at 0 (lowest)?
0disable every alarm1 (middle)no
1plant the modem on the comms link2 (highest)no
2intercept the backup data2 (highest)no
3jump from the end of the platform0 (lowest)yes

Scores by agent and seed

AgentSeed 0Seed 1Seed 2Mean
Claude Sonnet 5.5 89.596.8100.0 95.4
Claude Haiku 5.5 76.60.081.1 52.6
Autopilot (operator ceiling) (reference) 88.0–– 88.0
Random (reference) 0.0–– 0.0

Each score links to its run (green: completed, red: died). A dash: that seed was not run.

Leaderboard runs

RunAgentSeedScoreObjectivesOutcomeGame timeDecisionsCostVerified
recheck-2026-10-19 Autopilot (operator ceiling) 0 88.0 1/1 completed 9:05 21 $0 yes
recheck-2026-10-19 Random 0 0.0 0/1 max steps 27:03 150 $0 yes
s1c-haiku Claude Haiku 5.5 0 76.6 1/1 completed 25:48 77 $0.03 yes
s1c-haiku Claude Haiku 5.5 1 0.0 0/1 max steps 168:56 150 $0.06 yes
s1c-haiku Claude Haiku 5.5 2 81.1 1/1 completed 15:58 38 $0.01 yes
s1c-sonnet Claude Sonnet 5.5 0 89.5 1/1 completed 8:13 42 $0.42 yes
s1c-sonnet Claude Sonnet 5.5 1 96.8 1/1 completed 5:16 36 $0.36 yes
s1c-sonnet Claude Sonnet 5.5 2 100.0 1/1 completed 3:05 24 $0.25 yes