# Evaluating Ukraine’s Reconstruction Projects

**A proposed experiment on presentation effects and AI-assisted evidence preparation.**

Research proposal for discussion · 10 September 2026

## The question and value

When the facts of a civilian reconstruction project stay fixed, does presentation change human reviewers' priority assessments? Can an AI-assisted, human-verified brief reduce that sensitivity at a lower preparation cost than manual extraction into the same template?

Reviewers have limited attention, and communities differ in their ability to prepare documents. If wording changes judgments about identical evidence, a common review format may help. The study informs a practical choice: narrative descriptions, a fixed template, or AI-assisted preparation. It does not assume that AI will win.

## Data already available

Public DREAM records supply descriptions, objectives, costs and document references. The stored 12 August 2026 pilot contains 100 project records, 98 with an exact community-code match and 85 with forecast costs. They were selected by recent update, so they are not representative grant applications. A documented public API was successfully tested in September; a schema adapter is required before migrating the existing collector.

Source-linked dossiers still need manual verification; expert reviewers must be recruited; no human or AI evaluation has been run. The digital index is optional descriptive context because its measurement period and validity as an administrative-capacity measure remain unresolved.

## Proposed experiment

1. Define one eligible civilian project category. Two people verify source-linked facts, uncertainties, units and missing evidence.
2. Create plain and polished versions with the same claims and qualifications and comparable length. Randomly assign versions to reviewers, each seeing only one version per project.
3. Preregister one anchored 0–100 priority-for-further-review score. Estimate within-project presentation differences, accounting for shared projects and reviewers in uncertainty estimates.
4. Independently normalize each version through the same frozen AI workflow, then verify the brief. Measure the presentation difference again and compare with the raw difference. Benchmark against manual extraction into the same template from the same source packet. Count errors, omissions, failures, corrections and total preparation time.

An identical gold fact card removes wording differences mechanically; that is not proof of AI value. AI reviewers provide a separate model-behavior comparison, not evidence about actual funders.

## Contribution and limitations

AI-assisted proposals and AI review sensitivity have already been studied. The proposed contribution combines reconstruction records, human experts, source-level verification and a non-AI preparation baseline. Absolute novelty is not established. A review score is not funding awarded, engineering quality or social impact; these require different evidence.

## Request to the professor

Help choose a sector, agree a review rubric and secure access to qualified reviewers. Begin with roughly 30 manually checked projects as a feasibility pilot. Use pilot variance, workload and recruitment to design a powered, preregistered main study. Deliver an auditable dataset, a tested brief format and evidence about whether AI adds value. A precise null result or a better simple template is useful too.

## Primary sources and audit

- [DREAM open API](https://open-contracting.github.io/dream-api-docs/)
- [European Commission JRC: AI-assisted proposal writing, 2026](https://publications.jrc.ec.europa.eu/repository/handle/JRC146131)
- [Thorne et al.: LLM grant review perturbations, 2026 preprint](https://arxiv.org/abs/2603.08281)
- [Li et al.: Content-preserving rhetoric in AI peer review, 2026 preprint](https://arxiv.org/abs/2608.08975)

See the accompanying data audit and research review dated 10 September 2026. For a print-ready version, use the site's `/brief` page.