Commitment-Based Teacher-Student Fidelity Auditing with Public Randomness and Finite-Population Certification

Deploying compact models trained on reference-model responses calls for a verifiable assessment of behavioral fidelity without querying the reference on every input. We implement an audit combining SHA256 commitments to a finite input population and the Student's outputs, a future public-randomness beacon, and an upper confidence bound obtained by exact hypergeometric inversion. We study binary classification of Portuguese industrial-maintenance reports into electrical and mechanical specialties. A LoRA-adapted Qwen2.5-1.5B-Instruct Student ran locally on a CPU and committed 2,048 responses before sampling. The prespecified primary sample of 768 reports contained 64 disagreements with an API-served Teacher. Its one-sided upper bound was 9.7656%, meeting the declared 10% target at nominal 95% confidence under a fixed disagreement table and uniform sampling without replacement. A subsequent census found 158 disagreements, or 7.7148%. The primary stage used 37.5% of census queries and 37.6653% of reported census tokens; the complement was actually queried for validation. Human-reviewed labels already existed for all inputs, so this comparison concerns acquiring Teacher responses rather than evaluation against an existing answer key. The contribution is an operational, reproducible integration of established tools, separating record integrity, sampling validity, and reference stability. This independent preprint record includes the manuscript PDF and a supplementary reproduction package. Canonical repository: https://github.com/Antoniorjr4/auditable-teacher-student-fidelity. The supplement preserves the original evidence archive and supports offline documentary replay without model inference or API access. The manuscript and author-controlled dataset material are CC BY 4.0; replay code is MIT-licensed; model and third-party materials retain their own notices, as detailed in LICENSES.md. An initial software archive was automatically deposited from the earlier knowledge-distillation repository under DOI 10.5281/zenodo.23851131. That historical archive and the earlier baseline DOI 10.5281/zenodo.23644783 are distinct from this preprint. This preprint has not undergone peer review.