Code and derived data: batch-normalization folding in the MLPerf Tiny ToyCar anomalous sound detector

Study code and derived data for the manuscript "Batch-normalization folding errors shift the reference metrics of embedded anomalous sound detectors: evidence from MLPerf Tiny" (X. Lin). Earlier working title: "Conversion fidelity versus detection performance in a published anomalous sound detector: a batch-normalization folding case study". The MLCommons Tiny ToyCar autoencoder conversion script omits the Dense bias when folding batch normalization. This record contains the code, frozen protocols, test-set manifests and hashes, the four converted models (erroneous/correct folding x FP32/INT8), per-clip anomaly scores for all 2,459 ToyCar test clips and all model branches, saved paired bootstrap indices and replicates, numerical audits, execution-path diagnostics and the data behind the figures. No audio or audio-derived features are included. The recordings are available from the DCASE 2020 Task 2 development dataset (https://doi.org/10.5281/zenodo.3678171). The main statistics can be recomputed with NumPy only via recompute_saved_scores.py; see README.md inside the archive. Version 1.2 keeps toycar-bn-folding-v1.0.zip unchanged and contains the pre-specified extension study, frozen on 2026-10-22 16:17 UTC before any extension test clip was scored: E1, 30 autoencoders with the MLPerf Tiny architecture retrained on the six machine types of the DCASE 2020 Task 2 development data (five seeds each), each converted with erroneous and correct folding to FP32 and INT8; and E2, ten random calibration sets for the original ToyCar checkpoint. toycar-bn-folding-extension-v2-core.zip contains the frozen protocol and input hashes, all code, per-clip scores, per-model metadata, summaries, sanitized run logs and the figure script. toycar-bn-folding-extension-v2-models.zip contains the trained weights and the four converted .tflite models of each of the 30 retrained models; it unpacks into the same directory tree as the core archive. Version 1.1 was published before the model files were added and contains only the first two archives. No audio or audio-derived features are included. Licences: study code MIT; derived data (scores, statistics, models, figures) CC BY-NC-SA 4.0, following the source recordings; upstream MLCommons files MIT (Copyright 2020 Hitachi, Ltd.). See LICENSE, LICENSE-DATA.md and LICENSES/ in the archive. AI assistance: OpenAI Codex (model GPT-6 Astra) and Anthropic Claude (Claude Opus 5.5) assisted with code, computations and documentation; the author reviewed the content and is responsible for it. For versions 1.1 and 1.2, Anthropic Claude (Claude Code, Claude Opus 5.5) drafted the extension protocol, wrote and ran the extension code, and wrote the plotting code.