SentiLight — model checkpoints for the policy-space reformulation study

Trained model checkpoints accompanying the SentiLight study of output formulation versus model scale in policy-constrained affective smart-lighting control. Fifteen checkpoints: the five model families at three seeds each (42/43/44), all trained on the contamination-purged pretraining corpus. scale_grid_runs_clean/{23M,60M}/seed_*/f100/sft/ — generative language models (the A1 condition and the 60M comparator) a2_policy_only_runs_clean/23M/seed_*/ — policy-only supervised fine-tuning variant (A2) tuple_head_runs_clean/{23M,60M}/seed_*/ — 240-way tuple classification head (B2) These weights are not required to verify the analysis. The code, the frozen evaluation set, per-row predictions and run summaries for all 36 clean and 36 contaminated grid cells, and the preregistrations are in the software record (DOI 10.5281/zenodo.21930495), and every reported statistic and figure is computed from those. This record exists so the models themselves can be inspected and rerun. SHA256SUMS lists a digest for every file; verify with sha256sum -c SHA256SUMS. Caveats. Accuracy reported for these models means compliance with a deterministic labelling policy, not perceptual quality. The natural-language evaluation segment referenced in the accompanying software record is contaminated and is a diagnostic, not a benchmark. The tokenizer was derived from the pre-purge corpus and frozen. Labels for natural-language rows were produced by a teacher model, so any comparison against that model is teacher–student.