Reproducibility deposit — A small language model detects behavioural faithfulness gaps that frontier judges and human raters miss

Reproducibility package for the manuscript "A small language model detects behavioural faithfulness gaps that frontier judges and human raters miss": evaluation code, evaluation sets (including the 90,314-item auditor held-out set), tokeniser, corpus manifest, measured results, rater protocols, and the hierarchical-specificity ablation. All numbers are measured; scenario content is fully synthetic. See README_DEPOSIT.md inside the archive for the folder-by-folder manifest. Version 2 (2026-07-19): record title updated to match the current manuscript title; demonstration memory items in the retention scripts replaced with a fictional persona for privacy (an early demonstration log quoting those items was removed); four-condition experiment results unchanged; one malformed row in the held-out set repaired; per-item matched-geometry evidence added; requirements.txt and CHANGELOG.md added. See CHANGELOG.md inside the archive.