Data S2: harness-ablation replays and engine scores for "Checking language-model reasoning against ground truth in chess played through a robot arm"

Supplementary Data S2 (related to Figure S4). The 3,024 replays of the harness ablation (31 October 2026): the 63 forced-mate positions of the full-size games replayed three times, with the exact prompt of the game, to Claude Haiku 5.5, GPT-6 Luna, GPT-6 Sol and MiMo-V2.6-Flash-MOPD under uniform, vendor-default and as-played harnesses (and, for MiMo, through Claude Code and the server's chat endpoint). Prompts, replies, agent outputs, token counts and timings of every replay, the engine score of every replayed move, the proxy-ledger links of the MiMo requests and the analysis behind the reported numbers. README.md describes the conditions and files; MANIFEST.tsv lists every file with its size and SHA-256.