Corpus D2 — SIF v1.0 Layer 4: Output Failure Taxonomy for Anchor-Induced Fragility in Contradiction-Aware NMT

This record publishes Corpus D2 — the Layer 4 output-failure taxonomy component of the Structural Invariant Framework (SIF v1.0) for low-data contradiction-aware neural machine translation. Corpus D2 is a 22-pair failure-oriented behavioral probe for characterizing output-level structural failures in EN→UK NMT using Helsinki-NLP/opus-mt-en-uk with LoRA fine-tuning. Three failure modes are empirically characterized through controlled contrast pairs and ablation experiments: surface-form corruption (SF-1), anchor token loss (SF-2), and identifier drift (SF-3). The principal finding is anchor-induced fragility as a structural system property: canonical ⟨anch_*⟩ token encoding is sufficient to induce complete tag-set reconstruction failure independent of formula complexity or sentence structure, as demonstrated by the NULL-003/NULL-004 controlled contrast pair. An anchor format replacement ablation confirms that the failure is format-driven (token boundary sensitivity), not semantics-driven, with exact_tag_set_match improvements of +0.250 to +0.375 under plain marker encoding. A seed-aggregated regime probe (3 regimes × 3 seeds, 198 rows per regime) confirms that the representation-level two-tier hierarchy (ST-CANON ≈ ST-STRONG > Conservative at 99.2%) does not transfer to the output layer under any training regime or seed. TRIZ integration across six curated metrological and quantum-metrological cases yields one verified improvement (Q3: +0.5 anchor delta via frame separation), two stable VERIFY outcomes, two Layer B null results, and one negative outcome (Q1), providing proof-of-concept closure of the diagnostic cycle. Within SIF v1.0, Corpus D2 does not aim to improve translation quality. Its role is diagnostic: to identify what breaks, under which conditions, and whether failures are systematic enough to justify bounded intervention. The companion Layer 5 artifact, Module 4.0, implements the subsequent bounded anchor reconstruction bridge and should be cited separately once its DOI is assigned.