Data quality and lineage: can we reproduce the number?

1. The problem this lesson solves

Lesson 4.1 asked which decision a digital tool should change. This lesson asks the next industrial question: if we publish a carbon total, can an independent reviewer reconstruct how it was produced? In practice the answer often fails quietly. A meter was replaced mid-month. The historian stores fifteen-minute readings in local time. The ERP invoice is in MWh while the calculation expects kWh. A weekend gap was filled by a model. The electricity factor changed version. An AI extraction service wrote two invoice lines. The published total still looks precise to one tonne.

Numerical precision is not evidence quality. Reproducibility is necessary, but a flawed formula can be reproduced perfectly — and a correct-looking total can be irreproducible if source evidence and versions are missing. Inventory practice from the GHG Protocol and IPCC Volume 1 treats quality, QA/QC and uncertainty as part of the claim, not as optional footnotes. Digital MRV literature such as Körner et al. (2025) and World Bank dMRV materials make the same point for software-heavy systems: credibility is an end-to-end evidence design.

2. What “reproduce the number” means

Teams collapse several claims into one word. Repeatable means the same team and setup obtain the same output. Reproducible means an independent reviewer can recreate the output from documented artefacts. Traceable means every material output connects to identifiable inputs, transformations and responsibility. Valid means the method measures the intended construct for the stated purpose. Verified means competent checking against defined criteria. Accurate means sufficiently close to the relevant true or accepted value. Defensible brings method, evidence, controls, uncertainty and governance together.

A reconstruction challenge therefore asks seven questions: What is the claim (boundary, period, gases, method, unit)? What was observed? What changed in transformation? Who acted? Which versions of data, code, factors and models applied? Can the run be rebuilt? Can the claim be defended under QA/QC and exception policy?

3. Quality is fitness for purpose

Data quality is not a single score. Accuracy, completeness, consistency, timeliness, validity, uniqueness, integrity, representativeness, granularity, transparency, accessibility and security each answer a different test. A monthly supplier estimate may be fit for operational screening and unfit for an externally assured claim. A value adequate for trend monitoring can be inadequate for product-level allocation. Requirements must be derived from the decision, claim, materiality, method and acceptable uncertainty — the same logic as GHG Protocol uncertainty guidance and ISO 14064-1.

Large inventories make the stakes visible. EDGAR’s global atlas and city-level work such as Cai et al. (2019) show how statistics, factors and coverage choices dominate uncertainty. Independent checks can go further: Manning et al. (2021) use atmospheric measurements to test UK HFC inventory estimates — verification as a second evidence pathway, not a re-sum of the same spreadsheet.

In Türkiye, ETS and SKDM-facing reporting raise the same fitness question for plant files: what evidence standard applies when a number leaves the factory for a regulator or a buyer? Official guidance pages from İklim Değişikliği Başkanlığı and the Ministry of Trade are part of that context — not because they replace plant lineage, but because they define which claims must survive scrutiny.

4. Metadata, schemas and identifiers

A number without context is ambiguous. “420” may mean kgCO₂e/MWh, gCO₂/kWh, a country-year factor, or a rounded tile. Minimum metadata includes identity, definition, unit, period, timezone, source, method, owner, status, quality flag, uncertainty, version, access classification and retention. FAIR principles push the same idea for findability and reuse; W3C PROV names the entities, activities and agents that connect them.

Identifiers should distinguish the thing from its changing attributes. You should use stable IDs for facility, meter, supplier, factor, dataset, calculation run, model and report. Names are labels, not keys. Spreadsheet row numbers are not identities. A data contract — field definitions, types, units, nullability, keys, valid ranges, freshness, owner and change policy — prevents silent semantic breakage when systems exchange files that look valid but mean different things. “Electricity emissions” may differ by location-based versus market-based accounting, gross versus net treatment, factor year or renewable-instrument rules.

When teams ingest vendor extracts or training tables for AI, Datasheets for Datasets offer a practical diligence template: motivation, composition, collection, preprocessing, intended uses and limits.

5. Measurement systems and meter data

A measurement is not the physical quantity itself. It is an estimate produced by an instrument, calibration, installation, sampling interval, conversion and environment. You should record meter identifier, serial, location, quantity, unit, range, resolution, accuracy class, calibration status, multiplier, firmware, clock source and maintenance history. NIST’s greenhouse-gas measurements programme and metrology research such as Röttger et al. (2021) underline why “the meter said so” is incomplete without traceability to references and stated uncertainty.

Errors arrive in familiar families: systematic bias, random noise, gross mistakes (wrong channel, shifted decimal), quantization, drift, sampling that misses process variation, and coverage gaps such as a missing weekend. For fifteen-minute electricity data, a normal thirty-day month expects 2,880 intervals (subject to timezone and daylight-saving treatment). Completeness compares received valid intervals with expected intervals — and reports missing, invalid, substituted and estimated intervals separately rather than merging them into a single “available” count.

Physical balances are independent controls. Main import should reconcile to submeters plus losses and onsite generation interactions within a justified tolerance. An imbalance is a diagnostic signal, not an automatic sentence against one named meter.

6. Missing data, imputation and corrections

Missingness is not one problem. It may be unrelated to observed and missing values, related to observed drivers, related to the unobserved value itself, or structural absence during shutdown. Confusing “unknown” with “zero” invents production that did not happen. You should climb an imputation ladder: recover source evidence first; use a redundant independently governed measurement; apply an approved deterministic substitution; use engineering balance or matched-period methods; only then consider statistical or machine-learning imputation — and if no defensible method exists, you should preserve missingness, disclose the limitation and restrict the claim.

Missingness must not be overwritten in place. You should store raw value, quality state, substituted value, method, model or rule version, uncertainty, reason, approver and time separately so downstream users can distinguish measured, supplier-provided, estimated, allocated, corrected and model-generated values. Turkish plant practice that routes OCR invoice capture into carbon calculations — as in the Brisa–Enerjisa collaboration — is useful only when extracted fields remain labeled, reviewable and versioned.

7. Security and integrity of carbon data

Lineage is incomplete without protection. Unauthorized modification of a multiplier or factor, deletion of raw intervals after aggregation, replayed invoices, spoofed supplier files, poisoned training pipelines, prompt-injected documents, ransomware that removes evidence at deadline, and insider overrides without dual approval all break the claim even when the formula is correct. A matching file hash supports identity relative to a known digest; it does not prove authenticity, authorship or correctness.

Availability and retention fail more quietly: licenses expire, portals change, keys vanish, formats rot, staff leave. Retention design includes raw evidence, readable exports, exit rights, dependency documentation, backups and recovery tests. Boundary and life-cycle choices also reshape “the number”: Jain and Rawal (2022) show how a net-zero energy building claim can still leave material carbon emissions outside the operational window — a reminder that integrity includes what the claim omits.

Lesson 2.3 asks whether digital support is trustworthy enough for rights, oversight and procurement. Unit 3 treats the same evidence chain as a cyber-physical asset under attack and recovery.

8. Applied case: LumaWorks and 12,846 tCO₂e

LumaWorks Materials reports 12,846 tCO₂e for a factory month. A reviewer asks a simple question: “Show me how this number was produced.” The sustainability team has a spreadsheet export. The source meter was replaced mid-month. The historian stores fifteen-minute readings in local time. The ERP invoice records MWh. A missing weekend was imputed. The electricity factor changed version. An AI extraction service populated two invoices. The published total is numerically precise; the evidence chain is fragmented.

9. References

• GHG Protocol. Corporate Standard and uncertainty guidance. https://ghgprotocol.org/
• IPCC. (2006). 2006 IPCC Guidelines for National Greenhouse Gas Inventories, Volume 1. https://www.ipcc-nggip.iges.or.jp/public/2006gl/vol1.html
• ISO. (2018). ISO 14064-1:2018 Greenhouse gases — Part 1. https://www.iso.org/standard/66453.html
• W3C. (2013). PROV Overview / PROV-DM. https://www.w3.org/TR/prov-overview/
• Wilkinson, M. D., et al. (2016). The FAIR Guiding Principles. Scientific Data. https://doi.org/10.1038/sdata.2016.18
• Gebru, T., et al. (2021). Datasheets for Datasets. Communications of the ACM. https://doi.org/10.1145/3458723
• Körner, M.-F., et al. (2025). Digital Measuring, Reporting, and Verification (dMRV) for Decarbonization. Business & Information Systems Engineering. https://doi.org/10.1007/s12599-025-00953-3
• World Bank materials on digital MRV / carbon-credit MRV. https://www.worldbank.org/en/topic/climatechange
• Janssens-Maenhout, G., et al. (2019). EDGAR v4.3.2 Global Atlas. Earth System Science Data. https://doi.org/10.5194/essd-11-959-2019
• Cai, B., et al. (2019). China city-level GHG inventory and uncertainty. Applied Energy. https://doi.org/10.1016/j.apenergy.2019.113579
• Manning, A., et al. (2021). UK HFC emissions verification. Atmospheric Chemistry and Physics. https://doi.org/10.5194/acp-21-12739-2021
• NIST. Greenhouse Gas Measurements Program. https://www.nist.gov/greenhouse-gas-measurements
• Röttger, A., et al. (2021). New metrology for radon at the environmental level. Measurement Science and Technology. https://doi.org/10.1088/1361-6501/ac298d
• Jain, M., & Rawal, R. (2022). Emissions from a net-zero building in India: LCA. Buildings and Cities. https://doi.org/10.5334/bc.194