What an auditor checks in a financed emissions file.
The calculation is the easy part. The hard part is proving it was performed the same way twice, with a named approver and a file that stands on its own.
Updated October 2026 · 7 min readFive questions, every time.
The wording changes with the firm. The underlying checks do not.
| Dimension | The question | What a defensible answer looks like |
|---|---|---|
| Completeness | What was in scope, what was left out, and why? | A written scoping rule, a coverage percentage and a reason for each exclusion. |
| Method | How was this calculated, and who approved it? | One named document, locked at a version, with a single approver and a date. |
| Data integrity | Was the source file changed after it was read? | A hash at ingestion and a sealed copy once the file is validated. |
| Reproducibility | Same inputs, same result? | A deterministic run over pinned inputs, checked against the recorded output hash. |
| Controls | Who can change this after approval? | Nobody silently. A change is a new version with its own approval. |
A folder or a file.
Both contain a number. Only one contains the reasoning that produced it.
The method is described, but the version that was actually used is not identified.
The knowledge walks out of the building with the person who built the sheet.
The file is assembled after the question is asked, and the order of events is lost.
Inputs, factors, method and approvals sit together, in the order they happened.
The decision belongs to a person, and the record says which person.
The reviewer can rebuild the number with standard tools and compare the result.
What sits inside the record.
| Component | What it proves |
|---|---|
| Computation snapshot | Every input as it stood at the moment of calculation. |
| Factor provenance | A source, version and date for each factor used. |
| Score derivation | The score for each exposure, what data existed and why that band was chosen. |
| Proxy record | Which stand-in was applied where data was missing, and what it cost in score. |
| Method snapshot | The rule in force, locked, with the approver and the timestamp. |
| Verification hashes | A hash over each part, so a reviewer can tell whether the file was altered. |
Reports describe what happened. The record is what lets someone else check it. That is why the file is sealed after approval: a later correction is a new version, not a quiet edit.
Three ways a disclosure fails.
Each failure is structural rather than numerical. The figure is often right; the evidence around it is what breaks.
The number that cannot be re-run
A factor was updated mid-cycle and nothing recorded which version was used. The reviewer runs the calculation again and gets a different total. The fix is a factor set that is pinned at approval, with a mid-period change treated as a restatement rather than an edit.
The method nobody can explain
Three analysts over eighteen months, and the original author has left. The documentation is a document with unclear versions. The fix is one locked method per period, owned by the institution rather than by the person who wrote it.
The file that changed after export
Source data was edited after it left the system, and the change was found months later. The fix is a hash at ingestion and a sealed version after validation, so a later edit is detectable rather than invisible.
When everyone approves, nobody does.
A signature sheet with five names answers a different question from the one being asked. The reviewer wants to know which person is answerable for the number.
After approval the workspace is sealed. Changing it means opening a new version with its own approver and its own date, which is visible in the record.
Replay in four steps.
The test is not whether the vendor can explain the number. It is whether someone outside the bank can rebuild it from the file.
A high score is not the point.
A portfolio measured at score 5 is not a failed disclosure. A portfolio whose score 5 cannot be explained is.
For each exposure the record should say what data existed, which band was assigned, which stand-in was applied and what that choice cost in score. With that, a high score is a stated position. Without it, the score is an assertion.
The improvement path is then a disclosure in its own right: what changed between periods, why it changed, and what the earlier period looked like before it did.
Questions we get asked.
What is the first question an auditor asks?
Can you reproduce this number? Everything else follows from whether the answer is a file or a search through old email.
Does a low data quality score fail an audit?
No. Early disclosures are expected to sit at score 4 or 5. What fails is a score with no written reasoning behind it.
How many people should approve a disclosure?
One, for the method and one for the computation, each named. A group approval diffuses responsibility until nobody owns the figure.
Does the reviewer need our software?
No. The record carries its inputs and replay instructions, so the work can be checked with standard tools after the engagement ends.
Ask the auditor's first question early.
Bring one period. We will show you where the trail currently breaks and what the record would need to answer it.