FY2024 · Indonesia · Frozen analysis
Environmental Disclosure–
Performance Intelligence
An NLP and regulator-grounded analytics framework for examining how corporate environmental disclosure aligns with observed environmental performance.
Built from sustainability reports, human-validated NLP, Indonesian PROPER data, entity resolution, statistical validation, and reproducible analytics.
DPD0–1 standings
DISCLOSUREPERFORMANCEThe research problem
Companies describe their environmental story. Regulators observe facilities from another angle.
The difficult question is not whether one source is “good” or “bad.” It is whether the relative standing implied by corporate disclosure aligns with regulator-grounded performance.
compared with
How does this research work?
Three steps, without the jargon
The analytical system is sophisticated; the central comparison is simple.
01
What companies say
Companies publish sustainability reports describing environmental targets, programs, actions, and reported achievements.
Sustainability report NLP analysis Disclosure score
02
What regulators observe
Indonesia’s PROPER program evaluates environmental performance at regulated facilities.
Industrial facility PROPER assessment Performance standing
03
Compare the standings
DPD subtracts regulator-grounded performance standing from environmental disclosure standing.
Disclosure standing Difference DPD
Does a positive difference prove greenwashing?No.
It identifies a mismatch signal for investigation—not deception, corporate intent, or a binary company label.
Educational simulator
See what divergence means
Move the two standings to understand the arithmetic. This demo never changes or recalculates real company scores.
0.72Disclosure
−0.61Performance
=+0.11Positive Divergence
Research architecture
From source documents to an honestly tested construct
Select a stage to see what happened, why it matters, and what was retained.
What happened here
Acquire and freeze FY2024 corporate reports
Why it matters / validation
Period and extraction QA
Frozen output
Report inventory
Research at a glance
A compact view of the evidence base
36Company-years
final DPD analysis
120Mapped facilities
verified PROPER links
40Reference companies
frozen EDS calibration
24,947Core sentences
disclosure corpus
271Regression tests
collected in frozen project
The final construct
One difference, two independently built standings
Environmental Disclosure StandingRelative position in the frozen 40-company disclosure reference
−Performance StandingMean official PROPER midrank across mapped facilities
=Disclosure–Performance DivergenceDescriptive misalignment on a common 0–1 standing scale
Did it actually work?
The null result is part of the product
After building DPD, the study tested whether higher positive divergence corresponded with more independent adverse environmental evidence.
External criterion validation
Two validation phases remained close to zero.
Improved evidence collection, reduced search-visibility bias, and human review did not produce a meaningful relationship.
Phase 6Aρ = 0.0383Initial independent criterion
Phase 6Bρ = -0.034395% CI [-0.3458, 0.2823]
External criterion validity was not supported.
DPD is a reproducible disclosure–performance divergence measure, not a validated greenwashing classification system.
Scientific integrity
A model should not claim more than the evidence supports.
The project retained the null result instead of tuning it away.
What it measures
- Environmental disclosure extent
- Disclosure substantiation
- Targets and governance
- Regulator-observed facility performance
- Disclosure–performance divergence
What it does not prove
- Corporate intent or deception
- False advertising
- Complete corporate environmental performance
- Confirmed greenwashing
What I built
Analytics and engineering judgment, in one end-to-end project.
Research design, NLP, human annotation workflows, statistical validation, entity resolution, regulatory-data integration, reproducibility, evidence traceability, and an accessible analytical frontend.
Read the methodologyPythonPandasNLPSentence Transformersscikit-learnHuman AnnotationBootstrapPermutation TestsEntity ResolutionPROPER IntegrationData QualityReproducibilityNext.jsTypeScriptInteractive Visualization
Continue into the evidence