Contradiction-Conditioned Response Structure in Low-Resource Neural Machine Translation: Representation-Level Diagnostics, Plurality Analysis, and the Transfer Limit

This record contains the full publication package for Layer 3 of the Structural Invariant Framework (SIF 1.0): a controlled multi-phase probe experiment using Corpus D — a 60-pair diagnostic instrument evaluated against three LoRA-adapted variants of Helsinki-NLP/opus-mt-en-uk (Conservative, ST-CANON, ST-STRONG) across canonical seeds {42, 123, 456}. The central empirical finding is a precisely characterized encoder-to-decoder transfer limit: at the representation layer, the first-order split (ST-CANON > Conservative) holds in 99.2% of cases across all seeds, establishing a stable structural signal. The second-order split (ST-STRONG > ST-CANON) holds at chance level (49.2%), establishing a two-tier rather than three-tier representation hierarchy. At the output layer, no stable regime differentiation is observed across any seed or structural indicator — a null result treated as the primary positive finding, precisely characterizing where the current system's diagnostic reach ends. Contents of this record:— Article: Contradiction-Conditioned Response Structure in Low-Resource Neural Machine Translation (PDF)— Module 5.0 v1.1: Diagnostic Interpretation Layer for Contradiction-Conditioned Response Analysis (full phase structure A–F)— TRIZ_Module5_Pilot: 5-case proof-of-concept demonstrating proposal-plus-admissibility complementarity at the representation-diagnostic level across four contradiction classes— Corpus D Canon v2.0: diagnostic instrument specification and canonical checkpoint protocol— Computation package (CorpusD_Package_FINAL.zip): pre-computed tables and PCA projections for reproduction of all quantitative results (Phase III–IV); no model re-training required. Key empirical results:· First-order split (ST-CANON > Conservative): 250/252 = 99.2%, robust across seeds {42, 123, 456} · Second-order split (ST-STRONG > ST-CANON): 124/252 = 49.2%, indistinguishable from chance (All counts reported at text-level, N=252 units derived from the 30-pair Phase III evaluation subset.) · Plurality (% strict ordering): INTERNAL 18.3%, CROSS 6.7%, RING 13.3%, MULTIAXIAL 10.0% · Output layer bridge threshold: NOT MET (max 1/8 cases, any seed) ·TRIZ_Module5_Pilot: 5 cases, NORMAL trigger state, differentiated admissibility across contradiction classes Scope: All results are strictly bounded to the fixed low-data protocol — Helsinki-NLP/opus-mt-en-uk, LoRA rank 8, EN→UA, Corpus D (60 pairs), seeds {42, 123, 456}. No generalization to other base models, language pairs, or corpus sizes is claimed.