Structural Invariant Framework (SIF 1.0): A Registry-Based Framework for Empirically Verified Terminology Preservation in Neural EN→UA Translation of Metrological Standards

This Zenodo record contains the public release of SIF 1.0 — a Registry-based framework for detecting, classifying, and enforcing domain-critical terminology preservation in neural English→Ukrainian machine translation of metrological standards. SIF 1.0 is a proof-of-concept release: it operationalizes and empirically verifies the terminology-preservation layer under a defined Registry-based evaluation protocol, while neural adaptation, broader real-document deployment, and formula/unit integration are developed in SIF 2.0. SIF 1.0 addresses a narrow but safety-critical problem: neural machine translation systems can generate fluent Ukrainian text while silently corrupting structurally important metrological terms, such as guard band, measurement uncertainty, metrological traceability, acceptance limit, or maximum permissible error. In regulated scientific and metrological contexts, such errors are not merely linguistic; they may affect conformity assessment, uncertainty evaluation, and downstream technical decisions. Framework context SIF 1.0 is the operational release within a broader SIF research series on structural invariant preservation in low-data scientific machine translation; SIF 1.0 itself implements a five-layer (L1–L5) execution architecture. The framework was developed through linked diagnostic stages: structural ordering, behavioral regimes, representation diagnostics, output failure taxonomy, and bounded structural repair. Earlier SIF components, published separately on Zenodo, established the invariant-preservation principle for formulas, constants, units, contradiction-structured data, model-behavior regimes, and representation-level diagnostics. This release finalizes the terminology-preservation and Registry-enforcement layer of that architecture. It does not replace the earlier layers; it operationalizes their central result: candidate AI-generated translations should not be trusted because they are fluent, but qualified through structural verification of critical invariants. Companion diagnostic modules provide broader validity-control context, while the present package demonstrates the bounded Registry-based verification and enforcement mechanism for metrological anchor terminology. SIF 1.0 treats terminology preservation as a structural invariant verification problem, not as a general translation-quality problem. In this release, verification means empirical/protocol-based checking against Registry-defined requirements, not formal verification in the computer-science sense of machine-checked proof. The central component is the Anchor Registry, which defines admissible canonical Ukrainian forms for selected metrological anchor terms. Model output is checked against this Registry, and deterministic Layer 4 enforcement is used to restore or flag violated anchor identities. This record includes - the SIF 1.0 article as a separate PDF;- the public demonstration ZIP package;- Anchor Registry v1.9;- public-demo evaluation data;- external robustness check results;- offline reproducibility scripts;- tokenizer artifacts required to reproduce the public-demo checks;- manifest and citation metadata;- demonstration/video material illustrating the SIF 1.0 structural-check logic. The package is intended to illustrate Registry-based anchor verification, deterministic Layer 4 enforcement, and the public-demo Anchor Identity Precision (AIP) calculation. It allows readers to recompute the public-package AIP and robustness metrics offline from the shipped CSV/JSON files. The record does not claim full production readiness. SIF 1.0 is an empirically verified public demonstration of the terminology-preservation layer under a bounded Registry-based protocol. Neural adaptation, broader real-document coverage, formula/unit integration, and expanded Registry coverage are treated as follow-up stages in the SIF 2.0 development path. Main article-level SIF 1.0 results The article reports the frozen SIF 1.0 evaluation set as 35 sentences: 26 anchor-bearing sentences and 9 null-probe sentences, evaluated against the 22-anchor Registry v1.9. - Baseline MarianMT AIP: 0.1905- Naive glossary baseline AIP: 0.1429- SIF Layer 4 AIP: 0.7143- Frozen null-probe false intervention rate: 0.000 Public demonstration ZIP The ZIP package contains an auditable public subset of the frozen evaluation data. Public-demo AIP is computed over 21 anchor-bearing positions drawn from the 22-anchor Registry v1.9. The 9 null-probe sentences contain no Registry anchors and are used separately to verify the false-intervention rate. This subset is provided for offline reproducibility and reproduces the headline public-package AIP values from the shipped CSV/JSON files. It is not the full article-level frozen evaluation set; it is a reproducible public demonstration subset. External robustness check The external robustness check is evaluated over 50 anchor positions drawn from the 22-anchor SIF 1.0 Registry: - Raw MarianMT baseline AIP: 0.1800- Naive glossary baseline AIP: 0.2800- Layer 4 without EII: 0.3600- SIF Layer 4 full AIP: 0.9800 These external robustness numbers are reproducible offline using the scripts and data included in the ZIP package.