# Reproducible LLM benchmarks — posttrainllm > Browse reproducible language-model benchmarks with frozen protocols, prerecorded outputs, local replay commands, and specialist comparisons. Canonical: https://posttrainllm.com/benchmarks Benchmark archive # Every score should have a replay. Protocol, raw decisions, local command, and the specialist beside its larger opponent. Successful and failed attempts both belong here when their evidence can be inspected and replayed. ## Available now 3 evidence surfaces · 0 qualified ratings [± ### Cross-game Arena candidate · ratings unqualified Head-to-head Arena Elo and paired-score evidence, with uncertainty and qualification shown before rank. extensible adapters95% intervalsforfeit-awareno FIDE claim Chess sample4 games Provisional gap170 Arena Elo Qualifiedno Enter →](/benchmarks/arena) [♞ ### Character Chess development gate passed Can chess specialization compress frontier move selection into a 30–50M Mac-local model? FEN + UCI20 tacticsfull gamesno tools Frontier65% exact Strongest local9B · 10% Our SLMnot trained Watch →](/benchmarks/chess) [2048 ### Character 2048 failed · advantage not reproduced Can a 30–50M specialist compress game intelligence from a frontier general LLM? character inputstrict + constrainedno toolsfull replays Best general LLMQwen3 4B Instruct Mean score 845 Our SLMnot trained Watch →](/benchmarks/game-2048) **The admission rule** A game is admitted only when a pinned frontier model clearly beats random legal play and the strongest measured local general model. Chess passed that development screen; Character 2048 failed and remains visible as negative evidence.