Withholding farmers’ initial counts of shade trees planted from seedlings cut the researchers’ estimate of strategic or collusive amendments from about 25% to 11% in an experiment involving 407 cocoa farms in Côte d’Ivoire.
The findings come from a shade-tree planting programme run by an unnamed international cocoa buyer and exporter. The paper, written by Federico Cammelli, Johan Six and Rachael D. Garrett, was published in Science on August 13, 2026.
The study does not show that one-quarter of cocoa audits or sustainability certifications are unreliable. The 25% figure is a statistical estimate, not a count of individual records proved to be fraudulent. The research examined one company programme and a specific reward experiment.
Auditors could see whether a farmer qualified
The farms were taking part in a programme that encouraged the planting of shade trees. These trees can support cocoa production while providing environmental benefits such as greater biodiversity. The public data cover 28 field agents working across 85 villages and 17 cooperatives.
For the experiment, farmers could receive a reward if the auditor’s count of shade trees planted from seedlings fell between 80% and 120% of the farmer’s earlier reported seedling count. Field auditors entered their results into a smartphone system.
The app told every auditor whether the first entry passed the reward test. Auditors could then revise their entries. The researchers randomly divided the farms into two groups: 203 in which the auditor could also see the farmer’s original count and 204 in which that number was concealed.
That setup gave the research team a way to compare routine corrections with changes that moved a farm towards a favourable result.
Changes clustered around the reward criterion
Among auditors who could see the farmer’s count, 49 of 146 initial failures, or 33.6%, were amended. Only five of 57 initial passes, or 8.8%, were changed. The 24.8 percentage-point difference produced the researchers’ estimate of strategic or collusive amendment for that group. The underlying counts are available in the public replication package.
The same pattern did not appear in placebo fields recording shade trees that grew as wildlings or occurred naturally and had no bearing on the reward. The authors inferred false reporting or collusion from the excess amendments to the reward-relevant field and the absence of a comparable pattern in those other fields.
The pattern does not establish the motive behind every altered entry. Auditors did not receive the reward themselves. An ETH Zurich account of the research suggested that maintaining relationships with farmers or sharing a reward could provide incentives, but the experiment did not establish that either explanation applied.
Concealing the target changed the result
Auditors in the second group still saw whether their first entry passed. They did not see the farmer’s original count, making it harder to work out which revised number would satisfy the test.
In this group, 26 of 145 initial failures, or 17.9%, were amended, compared with four of 59 initial passes, or 6.8%. The difference was 11.2 percentage points, down from 24.8 points when the farmer’s count was visible.
On that measure, estimated strategic or collusive amendments fell by more than half. The change did not eliminate the problem, but it improved the result without requiring another farm inspection.
For companies, the finding points to a practical control. A field-data system can conceal a benchmark until an initial observation has been locked. It can also preserve the first entry, record every revision and flag unusual editing patterns for review.
Those controls would not establish that every measurement is correct. They would make it harder to change a result quietly after the desired outcome becomes clear.
The study did not test consumer certification labels
The numerical findings apply directly to this shade-tree programme and experimental design. The researchers did not test Fairtrade, Rainforest Alliance or another consumer certification system, and the farms were not being audited for compliance with the EU Deforestation Regulation.
The study did not prove that every revised record was fraudulent or identify collusion in each individual case. Its evidence comes from the large difference in revision rates, the focus on the reward-relevant field and the reduction that followed when the farmer’s original count was concealed.
The method captures strategic editing after an initial entry. It does not measure every possible source of inaccurate field data, and it did not test whether the shade-tree programme improved biodiversity or forest conditions.
That narrower result is still commercially useful. Companies often invest in traceability software, satellite imagery and supplier databases, but those systems cannot correct a false farm-level record if there is no reliable way to identify the error.
EU deforestation rules raise the value of reliable farm data
The study arrives as cocoa businesses prepare for the EU Deforestation Regulation. The rules are scheduled to apply to large and medium-sized operators from December 30, 2026, and to most micro and small operators from June 30, 2027. Micro and small operators already covered by the EU Timber Regulation come under the new rules on December 30, 2026.
Before covered cocoa products are placed on the EU market or exported, upstream operators must conduct due diligence. This includes collecting the geolocation of all plots where the cocoa was produced, along with the production date or time range. Operators must establish that the products are legal under the relevant laws of the producing country and were not produced on land deforested after December 31, 2020.
The regulation does not make this cocoa experiment an EUDR compliance test. It does, however, make the reliability of plot and production data an operating concern. The European Commission says that an operator unable to collect the required information must not place the relevant product on the EU market or export it.
Satellite monitoring can help verify changes in forest cover. It still needs to be connected to the correct plot. If a location or farm record is wrong at the point of collection, the error can move through the rest of the supply chain with a convincing digital trail.
Audit design can matter as much as audit frequency
More inspections are not automatically better inspections. The cocoa study suggests that companies should also examine what auditors can see, when records can be edited and whether the system retains a usable history of those changes.
Concealing one piece of target information did not solve every data-quality problem. It did produce a large reduction in the study’s estimate of outcome-driven amendments. For businesses making environmental claims or preparing regulatory evidence, that is a reason to test the design of the monitoring process rather than concentrating on data volume alone.
