Analytical workflows for product testing with users in agriculture

AI-ready reasoning from rankings to decisions

Kauê de Sousa, Brayden Youngberg, Joost van Heerwaarden, Hugo Dorado-Betancourt, Jacob van Etten

The main analytical approaches used at Bioversity International and Wageningen University for tricot and on-farm ranking data, written so that researchers and AI agents can follow how we get from observations to scientific conclusions.

This repository is the analysis layer of ai-ready-documentation. The other layers answer what we know (tricot-knowledge-base), what we observed (tricot-data) and how standard ClimMob reports are produced (ClimMob-analysis). This one explains how we reason with the data and why.

About 1000FARMS

1000FARMS is the Africa-wide on-farm trial network behind most of the data these workflows were built on. It runs decentralized on-farm testing with the tricot approach and ClimMob, with CGIAR centres and national agricultural research and extension systems (NARES), and it continues the development of ClimMob). The team also refers to tricot trial data from before the project as 1000FARMS data, since it comes from the same network.

These workflows record how researchers at Bioversity International and Wageningen University analyse that data, so partners and AI assistants can reuse the reasoning, not just the code. The trial data are published in tricot-data and the evidence base in tricot-knowledge-base.

How to use this repository

Start from your scientific question, not from a method name.

Your question	Use case	Decision record
Which varieties do participants prefer, and how sure are we?	which-varieties-do-participants-prefer	ranking
How many participants do we need? Is the response rate enough?	how-many-participants-do-we-need	ranking
Which traits drive overall preference?	which-traits-drive-preference	trait associations
Do different groups of users prefer different varieties?	do-user-groups-prefer-different-varieties	socioeconomic heterogeneity
How do climate and environment shape performance?	how-does-climate-shape-performance	climate and environment
Do on-farm and on-station results agree? Is preference heritable?	do-on-farm-and-on-station-results-agree	genetic gain
Which measured traits (lab or station) explain preference?	which-measured-traits-explain-preference	trait associations, genetic gain
Which product concept do consumers prefer?	which-product-concept-do-consumers-prefer	consumer and concept testing

Structure

• decisions/ holds the methodological decision records. Each states the default approach, the alternatives, and what the results cannot support.
• use-cases/ holds generic, question-first workflows with synthetic data and tests. Each use case is its own RStudio project and follows the layout of template-repo-data-analysis (data/, script/, output/), plus tests/. New use cases start from use-cases/_template/.
• SKILL.md lets AI assistants use these workflows without cloning the repository. It routes each question to a use case and decision record. llms.txt lists the core files.
• archive/ holds the original project scripts from about ten years of work. They are kept for provenance and credit, are not maintained and are not a source of code to reuse. Use the use cases instead. AI assistants leave the archive alone unless the user asks for it.

Running a use case

Open the use case's .Rproj file in RStudio, or set the working directory to its folder. Then

source("script/01-analysis.R")   # the workflow
source("tests/run-tests.R")      # the tests

From a terminal, run Rscript script/01-analysis.R and Rscript tests/run-tests.R from the use-case folder. GitHub Actions runs both for every use case at each push.

Status labels

Every use case carries one label.

• historical means the code was used in a past analysis and has not been reviewed since.
• reviewed-not-executed means the code and reasoning were reviewed but not run.
• tested-on-synthetic-data means the workflow runs and passes its tests on synthetic data.
• validated-against-original-results means the workflow reproduces the results of the original study.

Guidance, not prescription

These workflows document how our team has approached particular questions. They are methodological guidance, not mandatory procedures, and they do not prove that a method is optimal for every question. Users and AI agents should check the assumptions, consider the alternatives, and adapt the analysis to their own question and data. AI agents should also read AGENTS.md.

Authors

• Kauê de Sousa, Bioversity International and University of Inland Norway
• Brayden Youngberg, Bioversity International
• Joost van Heerwaarden, Wageningen University
• Hugo Dorado-Betancourt, Wageningen University
• Jacob van Etten, Bioversity International

Citation and license

Cite this repository using CITATION.cff. Code and documentation are released under the MIT License.