Provenance (Carbon Data)
An auditor picks one line out of a Scope 2 disclosure and asks a question that sounds trivial: where did this number come from? The answer that satisfies them is not “DEFRA”. It is the publisher, the document, the edition, the sheet, the cell, the date it was read, and the version of your own dataset that served it.
Provenance is that trail written down. Without it a figure is not wrong — it is simply unverifiable, which in an assurance engagement costs the same.
Provenance is the documented trail from a published carbon figure back to the exact place it was read from — publisher, document edition, cell reference, retrieval date and data version. It answers “where did this number come from?” precisely enough to re-derive it.
A carbon figure is a claim about the world, and a claim is only as good as the route back to its evidence. Provenance is that route, recorded at the moment the number is used rather than reconstructed afterwards — because a reconstruction is a second opinion, not an audit trail.
What provenance means for a carbon number
In carbon accounting, provenance is the metadata that travels with an emission factor so that anyone can find the figure again in its original source. It is distinct from the factor’s value and from its citation. A citation names a document; provenance locates a number inside one.
The distinction matters because emission factors are not published as standalone figures. They arrive as cells in large spreadsheets that are re-issued annually, re-ordered between editions, and frequently revised without a visible changelog. “DEFRA 2026” identifies a publication. It does not identify which of several thousand cells you used, nor whether the number you copied is still in that cell today.
A provenance record is sufficient when a stranger holding only your record and the public source can retrieve the same number without asking you a question. If they need to guess which sheet, which row, or which edition, the record is incomplete — however respectable the publisher’s name looks in the footnote.
This is why provenance is an engineering property rather than an editorial one. It has to be captured automatically at the point of ingestion, because the information it records — which cell, which day, which edition — stops being knowable the moment the source is re-issued.
The five parts of a provenance record
A complete record answers five separate questions. Dropping any one of them breaks a different part of the audit.
| Part | Answers | What breaks without it |
|---|---|---|
| Publisher | Who asserted this figure? | No way to assess authority, licence or independence. |
| Document edition | Which release of their dataset? | Figures from different years get compared as if they were the same series. |
| Cell reference | Where inside that document? | Retrieval becomes a search. Two people find two different rows. |
| Retrieval date | When was it read? | A silent in-place revision is undetectable after the fact. |
| Data version | Which edition of your corpus served it? | A restatement cannot be reproduced; see factor vintage. |
The last row is the one most often missing, and it is the one that makes the other four reproducible rather than merely descriptive. Publisher, edition, cell and date describe the outside world. The data version records what you were holding when you did the sum — which is the only part of the chain you control, and therefore the only part you can guarantee.
One number, traced end to end
Take the UK grid electricity factor, the figure behind every UK Scope 2 line this reporting year. Its value today is 0.13096 kg CO2e per kWh [GreenCalculus grid.gbr.electricity.location_based · DEFRA 2026 'UK electricity'!E25 · v2026.203]. That is the claim. Here is the evidence behind it:
| Provenance field | This factor |
|---|---|
| Canonical name | UK grid electricity — location-based (generation) |
| Publisher / edition | DEFRA_2026 |
| Cell reference | 'UK electricity'!E25 |
| GWP basis | AR5_100 |
| Data version | v2026.203 |
Every value in that table is read live from the corpus rather than typed, which is itself part of the point: a provenance record transcribed by hand goes stale the first time the underlying row is revised, and nothing on the page announces that it has.
In June 2026 the UK grid factor moved within its own source workbook — the value shifted cell between editions. A record naming only “DEFRA 2026” survives that change while silently pointing at the wrong row; a record naming the cell does not. This is the ordinary case, not an edge case: annual spreadsheets get re-ordered, and nothing in the publication announces it.
Provenance as a field, not a footnote
A footnote is provenance for a reader. A reporting pipeline needs it as data, because the check an auditor performs — re-derive this line — is a query, not a reading exercise. That means the five parts have to travel with the value through every hop: ingestion, calculation, export, disclosure.
Carried as fields, a single factor’s record looks like this:
| Field | Carries |
|---|---|
source.id | The publisher and edition, as one stable identifier |
source.cell_ref | Sheet and cell inside that document |
source.retrieved | The date the value was read |
data_version | The corpus edition that served this response |
version_pin | Whether a requested past version was actually matched |
The last field is the one that turns a description into a guarantee. Asking for a past version and getting a figure back is not proof you received the historical value — you may have received today’s. A pin that reports whether it matched is what separates a reproduction from a fresh calculation wearing an old date.
Across the GreenCalculus corpus at v2026.203, every canonical row carries all three external parts — publisher and edition, cell reference, and retrieval date. That is a deliberate ingestion rule rather than an achievement: a row that cannot state where it was read from is not admitted, because backfilling a cell reference after the source has been re-issued is guesswork with a citation attached.
No chart appears on this page, and the omission is deliberate. Provenance is a structural property — a record is complete or it is not — and the only honest comparative series would be one asserting how often other publishers record each part, which is not something we have measured.
Provenance is not chain of custody
The two terms are routinely swapped and they govern different objects. Chain of custody tracks a physical or contractual claim as it moves between parties — that this electricity was renewable, that this batch of fuel was certified sustainable. Provenance tracks a number back to the document it was read from.
| Provenance | Chain of custody | |
|---|---|---|
| Object tracked | A data point | A product, certificate or attribute |
| Question answered | Where was this figure read from? | Who held this claim, and in what order? |
| Fails when | The source cannot be re-read | A link in the transfer chain is unverified |
| Typical evidence | Cell reference, edition, version | Certificates, transfer records, audits |
| Governed by | Assurance and reproducibility practice | Scheme rules (for example GO, REC, RSPO) |
A market-based Scope 2 claim needs both, and confusing them produces a characteristic gap: an organisation with immaculate certificate records and no idea which edition of the residual-mix table it applied.
What the standards actually ask for
No major standard uses the word “provenance” as a defined term. The requirement arrives instead as an evidence obligation: the GHG Protocol Corporate Standard requires that data sources be documented so an inventory can be verified and recalculated; ISO 14064-1 requires documented evidence sufficient for a third party to reach the same result; and verification under ISO 14064-3 tests exactly that. Read together, they describe a provenance record without naming one.
The practical consequence is that provenance is assessed, not declared. An assurance provider does not read your methodology statement and take its word — they sample line items and attempt to re-derive them. A trail that only works when its author is in the room is the failure mode this term exists to name.
Four ways a provenance record fails
Each of these passes a casual review and fails an assurance sample.
- Publisher without edition. “IPCC AR6” or “DEFRA” alone. The publisher is stable; the numbers are not. Two reporting years cite the same string and mean different values.
- Edition without location. The right workbook, no sheet or cell. Retrieval becomes a search across thousands of rows, and two reviewers legitimately land on different ones.
- Location without date. A cell reference that was true when written. Sources are revised in place more often than anyone expects, and without a retrieval date a silent correction is invisible.
- Everything except your own version. Full external provenance, no record of which edition of your dataset produced the figure. The number is traceable to the publisher and not reproducible by you — which is the half an auditor tests when they ask you to restate a prior year.
The fourth is the most common in well-run organisations, because the first three are visible in a footnote and the fourth is not visible anywhere until someone asks for a restatement.
Frequently Asked Questions
No. A citation names a document; provenance locates a number inside one. “DEFRA 2026” is a citation and it is not sufficient, because the publication contains thousands of factors across many sheets and is re-issued annually. Provenance adds the edition, the cell, the date it was read and the version of your own dataset that served it — enough for someone else to retrieve the identical figure without asking you which row you meant.
Because public does not mean stable. Annual factor workbooks get re-ordered between editions, and a figure can move rows without any announcement. A record naming only the publication still looks correct after that happens while pointing at the wrong row. The cell reference is what makes the difference detectable: re-read the cell, compare it with the value you stored, and a silent move shows up immediately.
They track different objects. Chain of custody follows a physical or contractual claim between parties — that this electricity was renewable, that this fuel was certified. Provenance follows a number back to the document it was read from. A market-based Scope 2 claim needs both, and the usual gap is an organisation with complete certificate records that cannot say which edition of the residual-mix table it applied.
Not by that name. The GHG Protocol Corporate Standard requires data sources to be documented well enough for an inventory to be verified and recalculated, and ISO 14064-1 requires documented evidence sufficient for a third party to reach the same result. Verification under ISO 14064-3 then tests it by sampling. The obligation is described rather than labelled, which is partly why the practice is uneven.
Recording complete external provenance and no internal data version. The figure is traceable to the publisher but not reproducible by you, because nothing records which edition of your own dataset produced it. That half stays invisible until someone asks you to restate a prior year, at which point the answer is a fresh calculation rather than a reproduction of the original one.
Need someone who does this? 6 carbon accounting & inventory providers in our directory →