Preliminary researchCivilian reconstruction · Ukraine

Can AI close Ukraine's reconstruction capacity gap?

Communities with stronger administrations may produce better project applications—not necessarily better projects. This study tests whether that presentation advantage affects access to reconstruction funding, and whether AI can reduce it by putting every project into the same factual format.

The research question

Does administrative and digital capacity help communities advance reconstruction projects—and does AI amplify polished packaging or mitigate it through standardization?

How the study works

One problem, tested in three steps

Measure the real-world gap

Link public DREAM projects to community digital-capacity data and test whether capacity predicts project volume, completeness, progression, and funding evidence.

Isolate presentation bias

Give AI the same verified project facts in weak and professional formats. Any difference is a controlled packaging effect.

Test a practical remedy

Convert projects into one standardized factual template and test whether presentation-driven differences shrink.

Where the research stands

The question is viable. The evidence is not yet final.

Public DREAM currently lists 14,613 project records. The site uses a deliberately small 100-project pilot to verify fields and linkage—not to estimate a national effect.

14,613public DREAM records available for the full collection

98/100pilot projects exactly linked to the community index

0AI evaluator runs completed; no packaging result yet

Current conclusion

Proceed with limitations.

The public data are rich enough to build the study, and the mechanism can be tested experimentally. But the current pilot is non-representative: it does not establish that digital capacity causes funding advantages or that AI reduces them.

Why it matters: reconstruction funding should reflect need, social benefit, readiness, and feasibility—not a community's ability to produce polished English-language documentation.

Can AI Close Ukraine's Reconstruction Capacity Gap?

Preliminary research brief · Snapshot 2026-08-12 · pilot-0.1.0

1. Research question

Do digitally and administratively stronger Ukrainian communities have an advantage in preparing and advancing reconstruction projects, and can AI standardization reduce that advantage without losing decision-relevant information?

2. Why Ukraine is a valuable setting

Ukraine's reconstruction requires many communities with unequal administrative resources to present civilian public-investment needs through digital systems. The setting makes information capacity economically consequential while offering structured public project data.

3. Conceptual mechanism

Capacity may affect underlying readiness and the completeness, structure, and language through which reviewers observe a project. AI could amplify polished packaging or mitigate it by normalizing facts.

4. Data collected so far

100 public DREAM records; 98 matched to official community digital-capacity scores; 63 communities.

5. Coverage and limitations

The sample is the most recently updated records and is not representative. The index period is not exposed; funding semantics and stage-history coverage remain incomplete; need and fiscal controls are limited.

6. Variable definitions

Completeness is weighted field presence. PQS combines six interpretable components. Funding success is evidence level at or above a configured threshold, defaulting to level 2.

7. Descriptive findings

Mean completeness is 88.5/100 and mean PQS is 83.4/100. 45 records meet the default staged funding threshold.

8. Adjusted findings

The preliminary adjusted PQS model estimates 2.26 points per 10 index points (95% CI 0.06 to 4.46). Funding models are imprecise. These associations are not causal.

9. AI experiment design and results

Thirty factual cores and deterministic pilot variants are prepared. Manual equivalence review and evaluator runs are not complete, so no packaging effect is reported.

10. Mitigation results

Mitigation has not been estimated. T5 AI-normalized variants and paired evaluator outcomes remain required.

11. What cannot yet be concluded

This pilot cannot establish that digitalization causes funding, that DREAM discriminates, that AI improves allocation, or that any project deserved more funding.

12. Proposed next steps

Expand the representative snapshot, add time-aligned need and fiscal controls, complete manual validation, run the randomized multi-model pilot, estimate treatment variance, and pre-register the powered experiment.