Structural Invariant Framework (SIF 1.0): a registry-based framework for empirically verified terminology preservation in neural EN→UA translation of metrological standards

Neural machine translation (NMT) can produce fluent output while systematically deforming domain-critical terminology in metrological standards. Under GUM (JCGM 100:2008) Annex D.1.1, an incomplete definition of the measurand is a possible source of measurement uncertainty; terminological corruption in normative translation can therefore propagate metrologically consequential error into downstream interpretation and decisions. We present the Structural Invariant Framework (SIF 1.0), a protocol-verified five-layer architecture that detects, classifies, and enforces canonical anchor forms in EN→UA NMT output for metrological standards. SIF does not target general fluency; it asks whether defined structural invariants are preserved. The framework contributes: (i) a taxonomy of six deformation types and two subtypes, derived from 183 annotated corpus rows based on JCGM 106:2012, GUM, and related vocabulary standards; (ii) the 22-term Anchor Registry, a control object implementing Registry Supremacy by defining admissible terminology for a specified document family; and (iii) deterministic Layer 4 enforcement. On a frozen set of 35 sentences (26 anchor-bearing and 9 null probes), detection recall was 1,000 and the false-positive rate was 0,000. Layer 4 raised Anchor Identity Precision, computed over anchor positions, from 0,1905 for the unmodified Helsinki-NLP/opus-mt-en-uk baseline to 0,7143 (+ 52,4 percentage points). The neural bridge remained a pilot and was not used in these results. Evidence from the frozen corpus, expert-elicited probes, and authentic JCGM documents supports the directional claim that, in the evaluated configuration, Registry-based deterministic enforcement is more reliable than learned generation for preserving controlled terminology.