Why Level 3 Flattens: Statistical Regularity, Enunciative Singularity, and Authorial Voice in AI-Mediated Academic Translation

This preprint proposes a theoretical-explanatory account for a finding already documented empirically within the CF-C Research Program: across two independent studies (Teste Alfa, PT→EN; Teste Delta, EN→PT), discursive-informational flattening (DIF) in AI-mediated academic translation concentrated systematically on Level 3 of Bronckart's Socio-Discursive Interactionism (SDI) model — the enunciative level, where modalization, voice management, and authorial responsibility are textually encoded. The article argues that this vulnerability is structural rather than accidental, resulting from a mismatch between probabilistic choices (governed by statistical regularity in training corpora) and inferential-epistemic choices (anchored in the situated responsibility of a specific enunciator). Drawing on the Taxonomy of Informational Gestures (CF-C), it proposes reading informational gestures as units of discursive meaning and discusses implications for machine translation evaluation and linguistic mediation instrument design. This deposit includes both the English version and a Portuguese translation of the same article ("Por que o Nível 3 achata: regularidade estatística, singularidade enunciativa e voz autoral na tradução acadêmica mediada por IA"). This preprint complements the pilot study preprint (DOI 10.5281/zenodo.22019234), the CF-C Instruments deposit (DOI 10.5281/zenodo.21351930), and the CF-C Program project on OSF (DOI 10.17605/OSF.IO/R7CYW).