Three real-time games get harder until every player fails. On each seed, every player faces the identical schedule; the one who stays alive longer wins that head-to-head. The leaderboard is the share of head-to-heads a player won.
Board 1 uses 10 seeds per game. That is enough to screen, not to settle close comparisons: no “X beats Y” claim is made from it. The Nagi entry at the top was picked as the best of four Nagi entries, which inflates its numbers (winner’s curse). That is why a separate, preregistered confirmation against Jev was run on fresh seeds; it passed, and it is reported in its own panel with its own n.
New · Nagi Burst V2 · frozen eval seeds …
Nagi Burst V2 (…) played the three Arena games once, on eval seeds that were sealed before it existed. Nagi Burst V1 and Jev are their recorded rounds on exactly the same seeds and rules. Each game is scored by the metric frozen for this release, not by Board 1’s survival time.
| Loading… |
Mean over the sealed seeds, 95% t interval in brackets. Higher is better in every game.
On each seed the higher score wins the head-to-head, an equal score is a tie (½). W is the share won. The test is the release’s frozen gate-A form: the t lower 97.5% bound of W above ½ and an exact one-sided sign test.
…
…
01 / Leaderboard · Board 1 · run 2
Each cell is a paired per-seed win rate: against every other ranked player, on the same seed, whoever failed later scores 1, a tie scores ½. “All games” is the mean of the three games, shown with a 95% bootstrap interval over seed indices.
| # | Player | All games · 95% CI | Snake Rush | Rotorwash-Ramp | Booster Gauntlet |
|---|---|---|---|---|---|
| Loading results… | |||||
Interval bar: 0–100%, dashed line at 50% (an even split). Green edge = Nagi entry. Pink edge = not ranked.
The rule was committed before any Board 1 result existed. Selecting the best of four inflates the winner’s own numbers, and two entries share weights (T-dual Burst and T-dual K-call). The confirmation on fresh seed indices is the next section.
02 / Preregistered confirmation · fresh seed indices 10–33
Seeds 0–9 chose T-dual Burst as the best of four Nagi entries, so its Board 1 numbers are optimistic. To check the lead without that bias, one comparison was frozen before any new data: Nagi-ENORMOUS Burst against Jev on fresh seed indices. The leaderboard above is unchanged and stays on seeds 0–9.
…
of head-to-heads in which Burst survived longer than Jev
95% CI … · n = … seed indices × 3 games · sign-flip p = …
Only Booster Gauntlet is individually significant. On Rotorwash-Ramp and Snake Rush the per-game rates are not distinguishable from an even split, so no per-game claim is made for them.
Jev ran as the same vendor version as in Board 1 (jev-1.13.0). Its answers on a fixed set of requests differ from call to call, so a fingerprint of those answers cannot pin the vendor; the preregistered check compares the version string, which matched in every run.
03 / How to read the results
Difficulty is a fixed function of the game clock, never of what a player does. So on a given seed, every player faces exactly the same ramp, and comparing two players on that seed is fair even though the ramp itself is uneven.
For each pair of players and each seed: fail later = 1, same tick or both alive at 60 s = ½. Average over opponents, then over the three games.
It does not depend on how steep the ramp is, so seconds in an easy level and seconds in a hard level are never mixed.
Survival curve: % of rounds still alive at each second (Kaplan–Meier). RMST: the area under that curve up to 60 s, in seconds. Level: the 10-second band (1–6) where the player failed.
Rounds alive at 60 s are censored, not counted as failures. In Board 1, none reached 60 s.
Three fixed bots set the scale in every game: naive fails around Level 2, heuristic around Level 3, a search planner around Level 5. The bots play with the same one-tick reply delay the models get.
Ladder position maps a survival time to 0 (naive), 0.5 (heuristic) and 1 (planner).
Worked example. On Rotorwash-Ramp seed 0, T-dual Burst crashed at 24.8 s and Jev at 20.1 s: T-dual Burst scores 1 for that pair, Jev 0. Over 10 seeds, T-dual Burst outlasted Jev on 9, so its Rotorwash-Ramp rate against Jev is 90%. Its game score is the mean of such rates against all five opponents.
| Game | Naive | Heuristic | Planner | Calibration seeds |
|---|
Median time of failure, and the level it falls in, for each bot on the calibration seeds, played through the real-time runner with a 1-tick (250 ms) reply delay. Each rung fails at the same level in all three games (within ±0.5 level), the planner is never alive at 60 s on these seeds, and adjacent rungs are separated (paired intervals exclude 0).
04 / Survival curves
Pick two players to compare; the other four stay as thin grey context lines. Shaded bands are the six levels. Dotted lines mark the bot anchors: N naive, H heuristic, P planner.
n = 10 rounds per player and game, so each step is one round (10 percentage points).
05 / The three games
Every game runs at 4 decisions per second for at most 60 s, over six 10-second levels. The world never pauses for a model. The clips replay one scored round each, for the top Nagi entry against the external competitors.
Loading games…
06 / Speed is part of the score
A reply is applied at the first 250 ms tick after it arrives, so a slower model acts on older information. A fresh action is one decided on the latest observation. Nagi and SemIf-4B run on the same pinned H100 path with three game streams per GPU; Jev is a remote API and its times include the network.
p50 = median decision latency, end to end (prompt, forward pass, readout; network for Jev). Fresh = share of live ticks controlled by a reply decided on the latest observation. Stale = replies older than 1 s, rejected. Resets = controls fell back to safe defaults after 1 s without an applied reply. 3-vs-1 = median-latency inflation with three concurrent streams versus one. T-dual Burst and T-dual K-call share weights: the difference is the readout, one forward pass for all controls versus one call per control.
07 / Rules and fairness
Difficulty depends only on the game clock. Seeds 0–9 are committed per game and shared by every player.
A decision every 250 ms, one request in flight per game stream. Replies apply at the next tick; the world never waits.
A reply older than 1,000 ms is rejected. Controls declare hold-or-one-shot behaviour; after 1,000 ms without an applied reply they reset to safe defaults. Before the first reply: defaults.
If the harness processes any tick more than 100 ms late, the round is void: replayed once, then dropped for both sides of every pair. More than 1 void in 10 rounds invalidates the whole cell. Board 1 run 2 had 0 voids.
Self-hosted players on H100, merged bf16 weights, fused kernels, greedy readout, 3 streams per GPU with request micro-batching gated by a parity check against single-request answers.
All 180 scored rounds (6 players × 3 games × 10 seeds) were replayed from the records and matched exactly. Clips are drawn only from verified records.
No reruns conditioned on outcomes. Run 1 of Board 1 was stopped for a harness defect and is never pooled with run 2.
Nagi’s T-dual and ENORMOUS-CL entries were trained on similar control tasks (intended, disclosed below). Training on the arena engines, seeds or states is forbidden; the audit found none.
“Near-transfer” training on similar machine-control tasks is informational and never changes a rate. External players’ training data is not auditable by us. Booster Gauntlet’s new skills (hoverslam without hover, multi-phase hops, relights) make it the least-seen game for every entry.
08 / Read before quoting
09 / The audit trail
Source receipt sha256 … · job … · generated by site/arena/build/make_v3.py.
10 / Archive
Kept as published, with their original data. They used different games, scoring and protocols, so their numbers are not comparable with Arena v3.
