Level 3: In-Context Consistency
Document-Wide Terminological Stability
NMT models have no long-term memory — each sentence is translated semi-independently. In multi-claim patent documents, this creates catastrophic term variation where a single component accumulates multiple different translations across claims. Patent examiners interpret term variation as intentional claim differentiation, triggering indefiniteness rejections under 35 U.S.C. § 112(b).
Corpus status:
4 cases
— Active. New cases added as documented from ongoing HITL review.
| ID | Error Type | Case Summary | |
|---|---|---|---|
| L3-001 | Synonym Drift | In-Context Consistency — The "Three-Headed" Device. A medical injector patent consistently referred to a single "control device." Generic NMT drifted between three different translations (commande, contrôle, régulation) across the document, artificially creating a system with three distinct controllers. This variation violates the legal presumption that different words denote different things. The alignment protocol implements "Global Term-Locking" to freeze the first approved term and propagate it consistently throughout the claims. | View PDF |
| L3-002 | Synonym Storm | Catastrophic Terminology Drift — "The Synonym Storm". In a bio-pharma patent, the core invention ("Flow process system") must be semantically invariant. Generic NMT, exhibiting "Goldfish Memory," generated 7 different French translations (e.g., système de traitement..., système de processus...) across 15 claims. This inconsistency fractures the legal scope, implying the existence of multiple distinct inventions. The alignment solution implements "Zero-Tolerance Locking," enforcing a single canonical term (système de traitement d'écoulement) throughout the entire document. | View PDF |
| L3-003 | Oscillation | Polysemy Hallucination & Semantic Drift — "Pace Pulse". In patent EP4142594, the term "pace pulse" refers strictly to an electrical trigger (impulsion). The generic NMT model hallucinated a biological context, translating it as pulsation (heartbeat) in the preamble, but correctly as impulsion in a subsequent step. This oscillation within a single claim violates the Rule of Consistency, creating a fatal legal contradiction where the claim ostensibly defines two different physical phenomena. The alignment protocol implements "Semantic Domain Locking" to strictly enforce the electrical definition document-wide. | View PDF |
| L3-004 | Domain Bias | Domain-Context Hallucination — "Beam Training". In 5G/6G telecommunications, "Beam Training" describes the physical alignment of antenna signals (Formation de faisceau). Generic NMT models, biased by AI training data, hallucinated a machine learning context, translating it as Apprentissage du faisceau (Beam Learning). This fundamentally alters the claimed invention from a physical signal process to a software learning process, risking patent misclassification. The alignment protocol implements "Bigram Context Constraints" to detect the [Beam + Training] compound and enforce signal processing terminology. | View PDF |