Press Room
All of the findings below are anonymous and aggregate — no company name
is mentioned. The method and sample size are given for every finding; raw data can be downloaded
as CSV. Independently reproducible.
Finding 1
Independent assurance is correlated with disclosure quality
+10.7 pts
Within-stratum (country held constant) weighted median difference ·
n=2,844 · in favor of assured disclosures
| Stratum | n | Difference |
|---|
| Hong Kong | 902 | +13.8 |
| United States | 658 | +9.2 |
| Türkiye | 347 | +7.4 |
| China | 264 | +2.9 |
| International (EU MRV shipping) | 176 | +13.5 |
| France | 173 | +7.7 |
“International” = EU MRV shipping operators, which have no single
country of domicile.
Why don't we publish the pooled figure?
When the entire universe is compared in a single bucket, the difference comes out as
-9.9 points — that is, the sign flips. This is a
Simpson's paradox: the assurance rate varies greatly across countries (in some
markets almost all records are assured but their medians are low), so pooling measures
this composition difference, not the effect of assurance. The right question is
"does assurance make a difference within the same market?" — the answer is in the table
above, and it is positive in every measured stratum.
Correlation is not causation — a company that
obtains assurance may already be more transparent. But the difference is measurable and
repeatable. Stratum threshold: at least 20 records in each bucket.
Finding 2
Disclosing the emission factor value is the exception
| Factor transparency | Records | Share |
|---|
| Discloses the factor value | 1,097 |
10% |
| Gives only the source name (DEFRA, IPCC…) | 786 |
7% |
| None | 8,769 |
82% |
Without disclosing the factor value, an emissions statement cannot be
independently recalculated. This is the fundamental problem with claims of
"comparability."
Finding 3
Disclosure quality by sector
| Sector | Records |
Median | |
|---|
| Cement & Building Materials | 35 | 61.8 |
|
| Finance | 344 | 57.0 |
|
| Food & Beverage | 206 | 53.8 |
|
| Automotive | 73 | 53.4 |
|
| Tech & Telecom | 355 | 52.8 |
|
| Healthcare & Pharma | 195 | 51.5 |
|
| Tourism | 23 | 51.5 |
|
| Energy | 185 | 51.3 |
|
| Textiles & Apparel | 50 | 50.8 |
|
| Transportation | 129 | 50.5 |
|
| Metals & Mining | 99 | 49.6 |
|
| Retail | 64 | 46.9 |
|
| Chemicals & Packaging | 102 | 45.3 |
|
| Insurance | 83 | 42.7 |
|
| Real Estate | 19 | 41.4 |
|
| Other | 3611 | 39.3 |
|
The Accuracy Estimate measures the disclosure's defensibility —
not whether emissions are high or low. Only sectors with n≥15 are shown.
Finding 4
The silent: listed companies that do not publish a report
466 / 733
As of July 14, 2026 (KAP list of Borsa İstanbul companies).
For 63.6% of Borsa İstanbul companies, no
sustainability report could be found in public sources. The scored universe consists of
companies that have chosen to publish one — this is a structural selection bias, and
we also publish the part we do not cover.
"Silent" does NOT mean it definitely does not publish; it means none was
found in public sources. The notification channel is open, and the list is updated if a
report surfaces.
Finding 5
Issues caught by the rule engine (anonymous count)
| Control | Flagged records |
|---|
| No third-party assurance, or narrow assurance scope | 7,825 |
| Incomplete scope coverage | 1,832 |
| Duplicated scope values (Scope 1 ≈ Scope 2) | 468 |
| Unexplained large year-on-year change | 422 |
| SBTi claim not on the official list | 310 |
| Scope 2 dual-reporting gap | 241 |
| Order-of-magnitude outlier | 120 |
| Line items do not add up to the disclosed total | 68 |
A flag does not mean a confirmed error — it is a signal of inconsistency
requiring human review. In the first reviewed batch, some of the discrepancies turned out to
be our own reading errors and were corrected.
Our own margin of error
Discrepancy in independent second reading: 6.6%
51 / 770 adjudicated records.
This is machine-to-machine agreement: the lower bound of our margin of error; the
upper bound will be published with a human-labeled gold set. Asking a data provider
"what is your margin of error?" usually gets no answer — ours is public.
Data
Download it, verify it yourself
sector_medians.csv
score_distribution.csv
data.json
Methodology:
/methodology ·
The protocol and formulas are public (no black box). Contact: support@carbonbench.org
Named data requests: A named list is never shared without running the pre-publication
notice and dispute process. This protects both companies and journalists. The public
package therefore contains counts only — no company names.
This page is not an audit or assurance opinion; it is a
statistical plausibility analysis produced with published formulas applied to public disclosures.
Snapshot (October 10, 2026): 31,948 current sealed records in the full
universe, of which 15,190 carry an Accuracy Estimate
(5,577 company disclosures ·
9,613 official-registry records). The findings on this page are computed on the
narrower disclosure universe — 10,648 company disclosures, of which
5,577 are scored. Each finding states its own n; use that n as the
denominator, not the universe totals.