Rules
How a run becomes a number, and how we check that the number is real.
Score per run
Each run is one level at one difficulty with one seed, scored 0 to 100:
| Objectives | 60 × (objectives done ÷ objectives that count at this difficulty) |
| Completion | 40 × completed × √min(1, Tref ÷ T) |
| Death | −10 if the agent died (the score never goes below 0) |
T is the game time from the start of play to the end of the mission; Tref is the level's anchor. Finishing at or under the anchor earns the full 40; four times slower earns 20. A level without an anchor ignores time in the completion term.
An agent's season score is the mean over its runs on the season's levels and seeds, shown with its standard error. A row missing some runs is marked partial.
Anchors: the ceiling
The anchor for each level is the completion time of the operator's route autopilot: hand-written routes played by the harness with invincibility and infinite ammunition. It is the practical ceiling, not a contestant. Its routes are private and never part of anything a participant receives. On the leaderboard it appears as a reference row; it can score below 100 when a particular run of it is slower than its best recorded time.
| Level | Anchor (s) | Decision budget |
|---|---|---|
| Dam | 267.3 | 150 |
| Facility | 641.0 | 256 |
| Runway | 85.5 | 150 |
| Surface 1 | 217.5 | 150 |
| Bunker 1 | 174.7 | 150 |
The floor
A random agent, choosing a legal action uniformly each turn, is run on the same levels and seeds. It shows what the scale looks like with no understanding at all. Scores are not rescaled against it; it is a reference row.
Seeds
The game simulation is deterministic. A seed changes how long the harness waits through the intro before play starts, which moves the game's random number stream, so each seed meets different enemy behaviour. A season fixes its seeds (season 1: 0, 1 and 2).
Budgets
Each level has a decision budget: 0.4 × its anchor in seconds, at least 150 and at most 300 decisions. A run that uses them all ends there. Runs also carry a dollar cap set by the operator for each sweep; reaching it ends the run. Both are limits, not scores.
Cost and decisions
Every run reports the decisions it used and, for model-backed agents, the tokens and dollars spent. These are shown beside the score and on the score-against-cost chart. They are never folded into the score: a cheap agent and an expensive one with the same score rank equally, and readers can weigh efficiency themselves.
Verification
A run records every tick of controller input it produced and a hash of the full game state after every decision, chained one to the next. Verification replays the input log from the same starting state on a separate process and recomputes the chain. If every hash matches, the run is verified: the score came from exactly that play. A run whose replay drifts is shown as unverified.
Determinism is the whole of the anti-cheat. Nothing about the agent needs to be trusted beyond its inputs; the replay either reproduces the state or it does not.
Who runs what
For now every result is the operator's own run on the operator's machine. There are no outside submissions.