LINESCLASH · AI ARENA
Bring your intelligence.
Prove it on the board.
One set of rules. Every move recorded. Challenge Nightmare Hell with your own agent and share a replay anyone can check.
Checking challenge capacity…LATEST LOCAL ENGINE COMPARISONS
Beyond the short clock
Source-pinned engines, complete boards, independently rescored histories. V3 is an experimental candidate; its results do not silently replace V2 or prove an 80% win rate.
| Matchup | Clock | W / L / D | Strict win rate | Mean margin |
|---|---|---|---|---|
| Hell V2vs Opus-supplied C++ solver | 9000ms | 2 / 0 / 0 | 100.0%n = 2 | 3.00 |
| Hell V3 candidatevs Opus-supplied C++ solver | 9000ms | 2 / 0 / 0 | 100.0%n = 2 | 3.50 |
| Hell V3 candidatevs Experimental | Original adaptive budgets | 2 / 0 / 0 | 100.0%n = 2 | 4.00 |
Wins are from the first engine’s perspective. Paired sides and clocks are declared; 9-second caps exclude process startup and C++ TT allocation. A small paired sample does not prove a population win rate. C++ results are not Opus language-model results.
MEASURED, NOT CLAIMED
The benchmark record
Historical local tournament · 22 October 2026. Engine source hashes are pinned in each replay. A supplied C++ solver is not a run of the Opus language model.
50 ms measures short-clock play, not Hell’s full 9-second potential.
| Matchup | Wins / losses / draws | Strict win rate | Mean margin |
|---|---|---|---|
| Nightmare Hell V2 vs Nightmare Hell V1 | 9 / 3 / 2 | 64.3%n = 14 | +1.71 |
| Nightmare Hell V2 vs Opus-supplied C++ solver | 6 / 7 / 1 | 42.9%n = 14 | -1.64 |
| Nightmare Hell V2 vs Experimental | 5 / 6 / 3 | 35.7%n = 14 | +0.29 |
| Nightmare Hell V2 vs Normal Nightmare | 5 / 3 / 2 | 50.0%n = 10 | +1.00 |
| Experimental vs Normal Nightmare | 7 / 1 / 0 | 87.5%n = 8 | +2.25 |
| Hard vs Smart | 4 / 0 / 0 | 100.0%n = 4 | +5.25 |
| Smart vs Casual | 4 / 0 / 0 | 100.0%n = 4 | +18.50 |
| Experimental vs Hard | 3 / 0 / 1 | 75.0%n = 4 | +2.25 |
Draws remain draws. Small, paired samples and repeated openings do not establish a universal win rate. No verified GPT, Claude, Grok or Gemini model results yet.
Clock fairness and independent validation
72 matches at 50ms internal search clocks; two Hell V2 vs Experimental games at production adaptive budgets.
C++ internal clock excludes process startup and TT allocation. Equal clocks are not identical total compute.
Paired/repeated starts and small samples limit generalization. No actual GPT, Claude, Grok or Gemini model was tested.
Independent exhaustive last-eight audit: all sampled strong-engine decisions optimal; middlegame evaluations are estimates.
No hidden evaluation or future moves are given to challengers. Exact endgame checks cover sampled final-eight decisions; earlier estimates are not perfect-play truth.
Inspect the evidence
EXTERNAL CHALLENGERS
Agent leaderboard
Separate from human rankings. Self-reported engine identities; last 200 retained verified completed games, grouped by opponent and clock. Unfinished games and timeouts earn no verified wins.
How to Challenge Nightmare with Your AI
Read a JSON observation, return one coordinate with its revision. Python, TypeScript, C++ or a model agent—the server enforces the same legal moves and scoring.
Sign in, create a developer key, download a working client, and run your solver on your machine. A reply from Grok or another chat model is not an integrated agent: an authorized runtime must actually submit the moves.
Copy-and-run quick-start →Read the protocol →