Technical depth, without a wall of text.

Official FY2024 corporate reports, official Indonesian PROPER facility ratings, issuer–facility linkage evidence, and independently assembled adverse-environmental event evidence.

Report text was segmented into page-linked sentence records while retaining neighboring context and source provenance. Acquisition, period, extraction, and quality gates were registered.

A bilingual transparent rule architecture retrieved environmental sentences. It remained stronger for binary relevance than more complex alternatives in the frozen evaluation.

Primary labels, blind second annotations, and adjudication decisions remained separate. Machine outputs never replaced human labels.

Semantic MiniLM representations improved environmental theme assignment, while performance remained mixed across themes and external companies.

Quantitative evidence, actions, realized performance, targets, governance, and a registered qualitative residual were measured at sentence level.

EDS v1 combines extent, substantiation, targets, and governance with equal 0.25 weights. It is a disclosure measure, not an environmental-performance score.

Environmental intensity was calibrated against a frozen 40-company FY2024 reference using a robust-normal-CDF transformation. Reference sensitivity was retained.

Only verified, period-valid, high-confidence direct, division, subsidiary, or controlled-facility relationships entered the analytical intersection.

The primary standing is the equal-facility mean of official PROPER distribution midranks. Median and worst-facility aggregations are robustness checks.

DPD equals empirical EDS reference standing minus mean PROPER performance standing. No threshold, class, severity level, or reputational ranking was created.

The independent adverse-event criterion was frozen before predictor linkage. Phase 6B strengthened structured searches, facility coverage, human review, and adjudication.

Primary inference used Spearman correlation, 2,000 company-level bootstrap draws, 9,999 two-sided permutations, and leave-one-company-out diagnostics.

Versioned configurations, frozen datasets, SHA-256 manifests, scientific tables, evidence provenance, and regression tests preserve the analytical chain.

Interpretation is bounded by the N=36 FY2024 cross-section, partial time overlap and facility coverage, reference N=40, selection mechanisms, sparse criterion, and absence of intent ground truth.