A Research Arc from the Limits of Pointwise Optimization

This document presents the Structural Invariant Framework (SIF) as a single research arc rather than a set of separate modules. The arc begins with the limits of pointwise optimization in low-data neural machine translation and follows the sequence of constraints that emerged from one formally derived structural limit and four controlled empirical hypothesis rejections. The central claim is not that a model can be made more fluent or more accurate in general, but that critical structural invariants in metrological translation require admissibility control outside the generator. The overview connects the published components of the framework: the loss-limit argument, contradiction-structured corpus experiments, representation-level diagnostics, output-failure taxonomy, span-local bridge repair, SIF 1.0 Registry Supremacy, and the SIF 2.0 pilot protocol for audit-ready structural control. The framework is positioned at the intersection of metrology, neural machine translation, and TRIZ-derived contradiction analysis. TRIZ is used operationally as a source of controlled stress probes, not as a generative reasoning layer. Metrology supplies the admissibility requirement: critical terms and structural invariants must be defined, versioned, checked, and audited independently of the neural model. SIF 1.0 established registry-based terminology control for EN→UA metrological machine translation. On the frozen SIF 1.0 evaluation set, anchor identity precision increased from 0.19 to 0.71, with detection recall of 1.000 and zero false activations on null probes. SIF 2.0 extends the principle toward formula, constant, unit, audit-trail, and bounded-repair control under development-stage pilot conditions. The document is intended as a map for readers entering the framework from metrology, NLP, TRIZ, or trustworthy AI. It does not replace the individual technical papers and reproducibility packages; it explains how they form one cumulative argument: optimization can improve proposals, but admissible identity must be defined by an external, versioned, machine-checkable authority.