Evidence, Not Rulings: Reliable Machine Learning for Molecular Halal Authentication

halal certification of processed food depends on laboratory evidence of species origin: gelatin, hydrolysates and emulsifiers do not show which animal they came from. We study what machine learning can add under one fixed rule: a model may report molecular evidence of a prohibited-source species, but it may not issue a religious ruling. That rule forces a three-way label space prohibited, permitted, and inconclusive and treats the two error directions separately. On a benchmark of 17,941 tryptic peptides from 16 marker proteins across 28 panel species, a “species-specific” label is really a claim that the peptide is absent from permitted species. Checking that claim against 1.59 million UniProtKB sequences for the panel-not only the marker records a pipeline returns removes the prohibited- source label from 12% of implicated peptides and places nine of fifteen recovered literature markers in more species than reported. Two further limits follow: a low per-peptide error rate does not keep clean samples clean, and with swine held out every model still labels a third or more of porcine peptides as permitted