Everyone measures carbon. No one measures how reliable these numbers are. Carbonbench is the accuracy layer for corporate carbon disclosures: it compiles public reports via machine reading, cross-checks them against official records, scores them with an open formula, and cryptographically seals every record.
Our mission is one sentence: that everyone who looks at a carbon disclosure — auditor, bank, importer, journalist, citizen — can find a measured answer to the question "how much can I trust this number?"
Carbonbench is not an audit or assurance service; it is a statistical plausibility analysis produced with previously published formulas applied to public disclosures.
Widely divergent emission factors are used across different reports for the same national electricity grid. A significant portion of records never discloses the value of the emission factor used — these disclosures cannot be independently recalculated. For the majority of listed companies, no public report can be found at all in our scan.
Even so, indices, credit terms, supplier selections, and public judgments are built on top of these numbers. Each one rests on an unmeasured assumption of trust.
We turn this assumption into a measurement. Next to every disclosure, we place three things: Accuracy Estimate (how defensible the disclosure is given its evidence), Transparency Score (whether the calculation can be independently reconstructed), and Target Quality (whether the commitment is consistent with the measured pace). All three are derived from a single published formula; the same input, worked out by any third party with pen and paper, leads to the same score.
No evidence, no score. For an organization that does not publish a report or provide a measurable inventory, we do not produce a modeled estimate; we honestly write "report not within our scope" on its profile. This label is not a judgment about the company — it only states that no processed report was found in the public sources we scanned. The record is opened once evidence arrives.
The formula is public. Score components, weights, and constants are in published configuration files; every change is versioned and its history is visible on profiles. There is no black box: every score has a "why" panel, and every value cites its source document and page number. Registry links are open; links to company reports open from the record page after sign-in with a company email.
Our margin of error is public. As a first in the industry, we measure and publish our own reading error: the disagreement rate on records resolved in the independent second-reading round is 6.6%. Honesty caveat: this is a machine-to-machine reconciliation and is the lower bound of our margin of error; the upper bound will be published once the human-labeled 200-report gold set is completed.
The company never pays for our reading error. If the report is correct and our extraction is wrong, the record is corrected and re-sealed as a new version; the correction is not used to penalize the company's score, and the case is logged in our error matrix. An inconsistency within the report itself, however, is separately flagged — because that is a problem with the disclosure, not with us.
No record is ever deleted. Our ledger operates on an append-only basis: every record is sealed with SHA-256, a correction means a new version, and the old version remains accessible under the "superseded" label. The score's history and the reason for every change ("formula version changed," "universe grew") are disclosed on the profile. The seal does not prove that a record is correct; it proves that it has undergone a specific verification process and has not changed since that day.
Carbonbench is not a large company and does not act like one. It was conceived and built by a small team of people, and an AI team extends its reach (see Our Team). AI adds breadth: collecting, reading and rule-checking reports at scale, and pressure-testing the methodology across more disciplines than a small team could cover alone. Our people add depth, taking each open question further than a first pass would. The process is human-in-the-loop by design: every medium-confidence automated reading is confirmed by an independent second reading, disagreements are resolved in human approval queues, and the final say on the score formula, the correction policy and publication decisions always rests with people.
This structure is a deliberate choice: a small operation + open rules + full traceability is more auditable than a large but closed structure. We do not want you to trust us — we publish the formula, the source, the margin of error, and the history so you can check for yourself.
If you'd like to contribute to the methodology, file a dispute, or collaborate with us, our door is open: write to us via the Support page.
All requests — disputes, data corrections, profile claims, API, press — are handled through the Help Center; our first-response commitment for disputes is 10 business days. Email: support@carbonbench.org. Ready-to-use press materials are in the Press Room.