Causal vs Retrocausal Attention Boundaries

This study quantifies the minimal right context that changes model decisions about authority bearing constructions under strict causal masking versus non causal access. We formalize flip probability P_flip(b to b+Δ), the instance level threshold τ(x), and the construction level threshold τ_C, and we measure breakpoint sharpness over a right context ladder b in {0, 1, 2, 4, 8, 16, 32}. The dataset contains minimal pair ladders per construction family, deontic stacks, nominalizations, enumerations, defaults, agent deletion, scope setting adverbs, role addressatives, across six languages, en, es, pt BR, fr, de, hi, with balanced length, domain, and register, and human gold labels. We evaluate frozen causal decoder models, non causal encoders and encoder decoders, and causal streaming variants with sliding windows. Masking primitives include hard truncation, stochastic truncation, and delayed reveal streaming with cache isolation and sentinel based leakage tests. Primary endpoints are τ_C by construction and language, P_flip curves, AUC_flip, and a latency accuracy frontier. Results map τ_C to observed compliance deltas on instructed tasks while holding administrative workflow constant. The contributions are a public dataset with right context ladders, an evaluation harness with tested masks, per construction τ atlases and P_flip plots, and a preregistered analysis that links regla compilada constraints, Type 0 production equivalence, to measurable authority judgments.

DOI

• Primary archive: https://doi.org 10.5281/zenodo.17378361/
• SSRN: Pending assignment (ETA: Q4 2025)