Filing-Order-Constrained Semantic Exposure in Japanese Annual Securities Reports

Abstract

Japan introduced a sustainability-information section in annual securities reports for fiscal years ending on or after March 2023. We develop filing-order-constrained semantic exposure, measuring each filing’s position relative to earlier peer disclosures. Within fiscal year, industry, and disclosure layer, we retain up to five earlier documents whose multilingual-E5 cosine similarity exceeds a layer-specific threshold. The network comprises 1,623 firms, 4,746 firm-years, 22,739 document-layer records, and 29,338 directed edges. The design preserves disclosure layers, defines peer sets economically, and enforces source-before-target ordering. Continuous outgoing/source and incoming/follower rankings are stable across threshold, top-K, cutoff, and high-document-frequency weighting choices, whereas common-component removal and alternative representations alter edges and roles. A 1,000-iteration date-permutation null distribution randomly reassigns firm-year filing times within fiscal-year-industry cells and rebuilds the network while holding documents and peer composition fixed. Mean Spearman correlations between actual and permuted score ranks are 0.575 and 0.688. The aggregate role count matches this null, and the Broker count does not survive two-sided Benjamini–Hochberg adjustment across four role-count tests. Same-size density also matches its null mean. Under the declared multilingual-E5 representation, the observed smaller-to-upper asymmetry is stronger than under the null (lower-tail empirical probability 0.002), providing no support for a large-to-small cascade. Primary market, analyst, next-year firm-value, and out-of-time textual estimates have intervals containing zero; none of 13 Nikkei-sourced environmental, social, and governance outcomes survives Benjamini–Hochberg adjustment. These results establish an auditable measure of ordered semantic structure while leaving downstream economic and behavioral associations unestablished in this sample.