Arena Live · 25 Sep 2026

Four AIs. Three real‑time games.

Nagi‑ENORMOUS 83/90 points · won 23 of 30 rounds vs Jev, OpenJev and Laya.

ENORMOUS points
83/90
Round wins
23/30
Games × rounds
3×10
Late, invalid or error replies
0

Leaderboard

3 games × 10 rounds

3 / 2 / 1 / 0 points per round by place. Max 90.

  1. Loading results…

P50 = median decision latency, range across games. Nagi ran on a local H100; Jev is a remote API (network included). Lockstep play: latency is shown, not scored.

Game 1 · 240 Hz physics

Rotorwash

Fly a helicopter through gusts with a swinging sling load. Stay up, go far.

35.8 s vs 13.8 smedian airborne, ENORMOUS vs Jev
9/10rounds ENORMOUS outlasted Jev (p = 0.02)
25–25points tie with Jev. OpenJev and Laya crash in ~3 s

No model completed a sling-load delivery, so points follow distance flown.

Mechanism
Integrator
semi-implicit Euler, 240 Hz, float64, deterministic
Decision
every 250 ms; controls zero-order held for 60 substeps
Action
3 simultaneous queries: collective Δc ∈ {−2…+2}×0.05 · cyclic u ∈ {−0.6, −0.3, 0, 0.3, 0.6} · winch {up, hold, down} → 75 joint actions
Thrust
T = K_T·(Ω/Ωₙ)²·c·(1 − L_vrs)·(1 + 0.25·e^(−h/R))
Vortex ring
loss up to 40% when descent > 0.6·v_h and edgewise speed < 0.3·v_h
Rotor/engine
governor + 0.5 s engine lag, P_avail 330 kW; over-pitch droops RPM; Ω/Ωₙ < 0.85 for 1 s = rotor stall
Wind
seeded Ornstein–Uhlenbeck gusts (τ 2.5 s) + a shear/downdraft zone per leg
Sling
300 kg point load on a 5–25 m winch rope; |φ| > 50° = crash
Course
seeded cave, gap narrows 70 → 24 m; 4 ghost helis, same seed, no contact
Score · cap
metres flown + 500 × deliveries · 90 s · late reply = hold
Full rounds · tap 🔊 for sound

Game 2 · four-player Tron

Lightcycle Royale

Boost, cut off, survive. The arena shrinks.

8/10rounds won by ENORMOUS
10/10rounds with more points than Jev (p = 0.002)
Mechanism
Board
48×48, obstacles exactly invariant under 90° rotation R(x,y) = (47−y, x); spawn p at Rp
Decision
every 125 ms, 6 options: {straight, left, right} × {normal, boost}
Boost
2 cells this tick; 3 charges max, +1 every 2 s
Resolution
simultaneous, 2 substeps; swap or same-cell = all involved die
Shrink
one ring per second from 15 s; fully closed at 35 s (hard cap)
Observation
egocentric 15×15 window, rival head vectors, charges, time to next ring
L1 sensors
per option: crash flag, free run ahead, flood-fill area (cap 400)
Rank · default
elimination order, same tick = tie · late reply = straight
Full rounds · tap 🔊 for sound

Game 3 · four-board battle

Stack Attack

Tetris, four boards, garbage lines sent to rivals.

10/10rounds won by ENORMOUS
10/10rounds ahead of every rival (p = 0.002 each)
Mechanism
Boards
4 × (10×20 + 2 hidden rows), Guideline SRS, one shared 7-bag sequence
Decision
one piece per 250 ms, two stages: rotation (≤ 4 distinct) → column (≤ 10); hard drop, no hold, no spins
Garbage
clear 2/3/4 → 1/2/4 rows to every live rival; hole column = f(seed, receiver's batch count), never the sender
Pressure
from 30 s, every 10 s all boards get k+1 rows (1, 2, 3 …), same hole
L1 sensors
per column option: lines cleared, max height, holes after the drop
Rank · cap
survivors by lines sent, then elimination time · 90 s · late reply = spawn-column drop
Full rounds · tap 🔊 for sound

Inside the Nagi family

Bigger line, better play

SMOL 0.5B → BIG 4B → HUGE 12B → ENORMOUS 27B, same three games.

  1. Loading results…

Lab notes

Round by round, for the nerds

Derived from the scored records. Pick a lineup.

Placement per round

Cell = finishing place (1 = best). Same seed per column for every model. Hover or tap a cell for the outcome.

Paired tests vs ENORMOUS

Exact two-sided sign test on per-round points, paired by seed, ties dropped. n = 10, so the smallest possible p is 0.002. Holm–Bonferroni over the 9 comparisons of each lineup.

Decision latency

All replies pooled over the three games, half-octave bins, log scale. Lockstep: latency is recorded, never scored.

Policy fingerprint

Share of each applied action, and H = normalized Shannon entropy (0 = the same action every time, 1 = uniform). A constant policy in a closed-loop game usually means the model is not reading the state.

Fairness

Same inputs. Same clock.

  • Same observation

    Same text, same closed option list for every model.

  • Simultaneous moves

    No model sees another's move before choosing.

  • Lockstep

    The game waits for every model. Network or hardware speed can't decide.

  • Rotating seats

    Starting positions rotate every round.

  • Pinned vendors

    Jev, OpenJev and Laya run their published, pinned versions. One attempt per decision.

  • Zero bad replies

    0 late, invalid or error replies across all decisions.

  • Real replays

    Every video replays the logged decisions of the scored run.

  • Seeds pre-registered

    Seeds fixed before play. Records archive pinned by commit–reveal, sha256 of records_20260925.tar.gz:

    dd122d17c0d7c45cf97e0e5a49cbc14d041ca43a0a0aa0c41a3898ec8a550000

Read before quoting

What this does and doesn't show

  • An exhibition: 10 rounds per game, one day (2026-09-25). Not a universal ranking.
  • Every model gets the same shared sensor features.
  • Nagi lines differ in backbone family and training recipe, not only size.
  • Measures skill under known rules. On reading brand-new rules, 27B does not yet beat 12B: our next target.
  • Engines and raw records stay private to keep the benchmark sealed; the hash above pins them.

Full tables ↗ · data.json