Hennouni RSD & BlindZone Demo v1.0: Geometric Floor Measurement and Semantic Blind Zone Detection for Embedding Models (RAG & AI Safety)
Hennouni RSD & BlindZone Demo v1.0: Geometric Floor Measurement and Semantic Blind Zone Detection for Embedding Models (RAG & AI Safety)
This record contains the complete release of two complementary diagnostic tools for embedding models, published together on 14 March 2026. RSD Protocol v1.0 (DOI: 10.5281/zenodo.18939230, 10 March 2026) measures the geometric floor of any embedding model: the minimum non-zero residual cosine distance below which two distinct texts become indistinguishable. Validated on 7 models, sigma < 0.10% over 10 independent runs. BlindZone Demo v1.0 is the operational complement: it tests whether an embedding model correctly distinguishes two semantically opposite critical phrases — affirmation vs. negation — in French and English. Result: 4 models out of 7 fail in French, 3 out of 7 fail in English, including in their native training language. This record contains: Theoretical paper (3 pages): semantic blind zones on critical phrases, binary FR/EN results on 7 models. Specification v1.0 (4 pages): full protocol, binary classification, AI Safety implications for medical, legal and financial sectors. User Guide v1.0 (3 pages): installation, execution, result interpretation. rsd_blindzone_demo_fr.py: French phrase pair test — 7 models. rsd_compare_EN_v1_0.py: English phrase pair test — 7 models. Combined workflow: RSD identifies the geometric floor → BlindZone Demo identifies dangerous phrase pairs → targeted fine-tuning with hard negatives. Estimated gain: 60–80% reduction in correction time vs. blind fine-tuning. Author: Mourad Hennouni (ORCID: 0009-0009-7735-2526) — AI Safety Engineering & Systems Analysis
semantic embeddings · embedding model diagnostic · residual cosine distance · geometric floor · blind zone · negation detection · cosine similarity · RAG safety · AI Safety · pre-deployment validation · contrastive learning · Lipschitz constant · French NLP · English NLP · critical phrase testing · embedding evaluation · dense retrieval · sentence transformers · HuggingFace · CI/CD pipeline
semantic embeddings · embedding model diagnostic · residual cosine distance · geometric floor · blind zone · negation detection · cosine similarity · RAG safety · AI Safety · pre-deployment validation · contrastive learning · Lipschitz constant · French NLP · English NLP · critical phrase testing · embedding evaluation · dense retrieval · sentence transformers · HuggingFace · CI/CD pipeline
selected citations These citations are derived from selected sources.This is an alternative to the "Influence" indicator, which also reflects the overall/total impact of an article in the research community at large, based on the underlying citation network (diachronically).0 popularity This indicator reflects the "current" impact/attention (the "hype") of an article in the research community at large, based on the underlying citation network.Average influence This indicator reflects the overall/total impact of an article in the research community at large, based on the underlying citation network (diachronically).Average impulse This indicator reflects the initial momentum of an article directly after its publication, based on the underlying citation network.Average
Citations provided by BIP!Popularity provided by BIP!