Which paper source is authoritative?

The file final_paper.md is actually a complete LaTeX conference manuscript despite its extension. It is the closest source to the reviewed DCASE submission: it has the submitted title, three authors, and the original main/ablation claims. latex/main.tex is a separate Rho-format report with two authors. It repeats the old principal scores, so matching those numbers does not establish that it is the latest correct conference source. latex/main_beyondwham_v1.tex is an older WhAM/augmentation narrative.

The detailed audit and evidence remain in reports_v2/reviewer_revision_audit.md and reports_v2/review_audit_evidence.json. Original files are preserved and were not overwritten.

The revision's single canonical manuscript is latex_v3/main.tex, with generated tables from the saved results and a unified refs.bib. It preserves the conference title and three-author list from final_paper.md, including equal contribution for the first two authors. It uses the official 2026 DCASE template, not the Rho report class.

Important differences introduced deliberately:

• Replaces historical scores produced under inconsistent representation/ID holdouts.
• Distinguishes whole-coda unit pairing from actual cross-channel swapping.
• Lists unit/type/ID and sampling supervision explicitly.
• Uses training-only scalers, validation-only layer choice, repeated random and date-grouped evaluations.
• Corrects date parsing, the year contingency table and Cramér’s V.
• Corrects sample-rate handling, temporal feature floors and identity eligibility.
• Removes unsupported universal-superiority, identity-gain and disentanglement claims.
• Corrects the Andreas author list and harmonizes verified bibliography records.
• Replaces the original unavailable/mismatched figure path with a newly drawn diagram matching the revised implementation.

The revised PDF now contains the completed, verified extension and has passed a five-page rendered-layout check, with references only on page five. The original report's formatting and bibliography are not silently propagated into it.