Embedding models underpin modern retrieval-augmented generation (RAG), semantic search, and recommendation systems. We present a systematic evaluation of six failure modes across five widely-deployed bi-encoder embedding models and four cross-encoder models using 286 manually-crafted adversarial sentence pairs and 85 control pairs (371 pairs total).
Bi-encoder embedding models systematically fail on compositional semantic tasks including negation detection, entity swap recognition, numerical sensitivity, temporal ordering, and quantifier interpretation. Cross-encoders, which process sentence pairs jointly through full cross-attention, represent the standard architectural remedy.
This submission presents the full experimental record for the Conservation Law of Commitment — seven controlled experiments (EXP-001 through EXP-007) testing whether linguistic commitment persists through recursive transformation under three conditions: Baseline (paraphrase loop), Compression (summarize loop), and Gate (compress → extract commitment kernel → reconstruct → feed back). The dataset comprises 57 signals, 181 condition-signal runs, and 10 iterations per run using GPT-4o-mini at temperature 0.