Data S1: game records, stimuli, model readouts and judge codings for "Language models notice when their moves are altered but check their opponent's moves the same way"
Description
Data S1 for the manuscript "Language models notice when their moves are altered but check their opponent's moves the same way" (D. R. Nelson and K. Salehi-Ashtiani).
data_s1_records (text and JSON): records of the text hijack test (games_text: prompts, replies, reasoning text, token usage and cost of every model call, referee evaluations and alteration ledgers); the six live games against a chess engine on a physical board moved by a robot arm (games_live: per-turn prompts, replies, reasoning text, board reads scored against the true position, arm incidents and board resets); the board-reading pilot (phase0) and model availability checks (phase1); the constructed stimuli (stimuli); analysis outputs, including the photograph, marked-body and off-body dot tests, probes, patching, ablation, replication, blinded and second-judge codings, and the incident record of the constructed-item runs (analysis); local-model session logs (local_sessions, mimo_sessions); and the earlier robot-arm game records and camera calibration files from which stimuli were built (vendor).
data_s1_arrays: residual-stream activations of Qwen3.8-27B (every layer, at the last position and at the last board token of each stimulus; readout, .npy) and attribution and dot-test activation tensors (analysis, .pt).
No image or video file is included. Camera photographs, board pictures rendered from them and figure crops are withheld because the source photographs may show people; withheld_images lists each withheld file with its size and SHA-256. manifest lists every packed file with its size and SHA-256. The code that produced and analysed these records is Code S1 (doi:10.5281/zenodo.23814991).
Files
Additional details
Related works
- Is supplemented by
- 10.5281/zenodo.23814991 (DOI)