Real-Time Detection of Authority-Bearing Constructions Under Strict Causal Masking

This datasheet defines a benchmark for real-time detection of authority-bearing constructions under strict causal masking, where models access only left context. It measures how accurately and quickly a system identifies linguistic signals of authority without future tokens. Authority-bearing constructions are treated as Type-0 productions within a regla compilada, binding syntactic and operational constraints to decisions. Three hypotheses guide the study: a compact causal detector with an authority lexicon can achieve reliable precision at low latency; performance depends on construction family and register rather than sentiment; limited buffers can improve stability without breaking causality. Multilingual datasets (English, Spanish, optional French, German, Portuguese) include transcripts, hearings, and policy texts segmented into token streams. Tasks involve streaming span detection and stance classification, evaluated at multiple latency checkpoints and causal budgets (b ∈ {32, 64, 128}). Metrics cover streaming F1, AUCL, and stability index.Baselines (oracle, lexicon-only, sentiment) and strict no-lookahead validation ensure isolation of causal effects. The benchmark shows how form, not intent, governs real-time authority recognition, enabling evaluation of models for compliance and human-in-the-loop systems without right-context access.

DOI

• Primary archive: https://doi.org/10.5281/zenodo.17465070
• Secondary archive: https://doi.org/10.6084/m9.figshare.30465578
• SSRN: Pending assignment (ETA: Q4 2025)