Numbers that can be traced back to a run.

Are the published numbers traceable?

The final numerical audit found no discrepancies in the checks below. Main table scores can be independently reproduced from saved predictions. This establishes numerical consistency, not correctness of every upstream processing decision or a guarantee of scientific validity.

The auditor recalculates macro-F1 directly from true and predicted class counts without importing the experimental scoring function. Both supported-example primary scores and full-eligible secondary scores match saved metrics within an absolute tolerance of 10⁻¹². Three-run means and sample SDs match the summaries, and the manuscript’s displayed values round correctly to three decimals.

What was checked

• Dataset accounting, identity eligibility, year-by-unit counts, Pearson χ², and Cramér’s V.
• Primary and secondary metrics from 792 task-prediction files across 264 result records.
• 426 summary mean/SD pairs, 96 displayed table entries, and 24 comparison-figure points.
• Additional numerical prose, including branch, pairing, swapping, bandwidth, and held-out-year results.
• Grouped identity support and the selected 1+1+3 diagnostic.
• Train/validation/test disjointness, date and year separation, prediction IDs, and label encodings.
• Recorded training IDs, checkpoint presence, finite histories, uniform sampling where applicable, validation-based layer choice, and saved source/split hashes.

The archive contains 240 primary/control result records and 24 separate branch-probe records. Among the 240, 158 have local neural-training records and 82 use fixed or frozen representations. These totals include historical correction runs and year holdouts; they are not 240 models listed in the final paper.

What this audit did not do

No model, probe, or notebook was rerun. The audit did not regenerate every embedding from raw audio or replay training from checkpoints. Training records and checkpoint presence support provenance but do not independently prove that each checkpoint generated each saved prediction.

Within-unit probes retained run-level scores but not per-example predictions. Their reported means were reconciled against those records, at a weaker verification depth than the main tables. The original audio-to-label mapping and external pretraining membership were not independently established by this arithmetic check.

Website-specific traceability

The website’s main result displays are generated from the final audit ledger rather than manually entered score arrays. Ablation and diagnostic expansions read the saved summary CSVs. Descriptive EDA figures use the clean label table and actual audio samples. The data export records the SHA-256 of the manuscript used as its source.

The download-provenance manifest lists source paths and hashes for files packaged into code, notebook, evidence, report, and paper-source downloads. Notebook source files are preserved byte-for-byte. The site build report records checks on the delivered pages, links, assets, data, and interactions.

Read or reproduce: full numerical audit report, machine-readable website data, code package, and saved experimental evidence.