A slot-aware middleware for traceable inference systems.

Abstract

The invention proposes ViMap-IL - interlingua slot-aware normalization of natural language input into a logical structure consisting of fixed slot fields with a law reference code (law_ref). This representation is encoded into a super-dimensional “Particle” hypervector for retrieval by the FAISS / BM25 / graph associative layer, where the SCOPED-FIRST step uses allowed prefixes from the configuration / industry vocabulary / NER to prioritize candidates in the correct domain and support fail-closed mode. Candidates are scored by a multi-signal LogicRank module (sim_sem, rule, evidence, freshness, intent, risk, latency), applying MMR and probabilistic correction to keep ECE ≤2%. The final results are in JSON format, including hs_final, fused_score, argument trace, and legal information (law_info). Operating under policy-as-code, it is capable of pass / need-review / deny and is traced via TraceFS (trace.json + law_info.json) for auditing purposes. The system achieves transparency, risk control, and a response time of p95 ≤ 1.5s under the target configuration.