CommitStage: raw experimental data for "Generated but Not Committed: Locating a Commit-Stage Failure behind a Salient Geometric Misbinding in an Open-Weights LLM"

Raw experimental output supporting the paper "Generated but Not Committed: Locating a Commit-Stage Failure behind a Salient Geometric Misbinding in an Open-Weights LLM" (Manabu Higashida, D3 Center, The University of Osaka). This record is a Tier-1 minimal-reproducibility archive: it contains only the raw output of the study's frozen, pre-registered experiments and backs the numbers reported in the paper's freeze-registry ledger. It comprises four experimental arms (953 files, ~8.3 MB): - closure-erasure — the closure-erasure battery (paper Section 4.1, predictions P-C1/P-C2): the salient geometric misbinding is unmoved when the closure cue is erased. - hypothesis-tracks — the budget sweep, the teacher-forced Delta-balance scoring, and Track R (Sections 4.2 and 6). - jlens-pilot — the Jacobian-lens readouts, pass-2 trajectories, and step-1/step-2 calibration (Section 5). - galileo-screening — the belief-domain generality battery of 420 runs (Section 8), screening the effect across models and domains. Each arm mirrors the results/ directory of the corresponding experiment. Every file follows a fixed convention: *_config_frozen.json pins the pre-registration (model id, quantization, sampling parameters, system prompt, prompts, seeds, and thresholds); *_runs / *_readouts / *_scores / *_trajectories hold the raw rollouts; *_verdict_*.json holds the frozen per-arm decision; and *.log files are kept as run provenance. An included MANIFEST.sha256 lists the SHA-256 of every file for integrity verification, and a README documents the full data dictionary. Subject models are open-weights: Qwen3-32B (bf16 and q4), gpt-oss-20B, and Llama-3.3-70B. In the galileo-screening arm, per-item scoring was performed by an LLM judge (claude-sonnet-5); each scored file records the judge's structured verdict together with its rationale and evidence quotes. The Jacobian-lens checkpoints (model weights) are deposited separately and are not duplicated here: DOI 10.5281/zenodo.21449254 (concept) / 10.5281/zenodo.21449255 (version). The archive is released under CC BY 4.0 for reproducibility and verification. The JSON files contain verbatim model outputs generated under the respective providers' terms of service; reuse of any output to train, fine-tune, or distill a competing AI model may violate those terms. This is not a training corpus. This work was supported by JSPS KAKENHI Grant Number JP26K06399.