MVU turns many written inputs, such as consultation replies, news articles or discussion contributions, into a map of information and perspectives: what is agreed, uncertain or disputed, and why. A language model drafts the map and a named person is meant to check it.
Problems and Purpose
Public consultations, news coverage and discussions produce more text than decision-makers or participants can read. MVU starts from the concern that summaries can blur who said what, hide disagreement inside averages and present open questions as settled.
MVU tries to keep three things visible: shared ground, open questions and real disagreement. Each is tied to the sources that hold it. The map is meant to help people orient themselves before a decision. It does not recommend a decision or propose a compromise.
The intended use differs by task. In a policy brief, the map is meant to show the people drafting a measure where groups share ground, what is still open and where they split, before they draft. In a media review, it is meant to show readers what diverse sources agree on and where they differ, in information and in perspective, and whether the reasons behind a position are the same or different. It maps the information and the viewpoints, not the sources: two sources can agree on one point and disagree on another, and can reach the same position for different reasons. In public deliberation, it is meant to help the group itself orient: after the event, the participants see the map, confirm it is accurate, and can use it to see where they stand. These uses are intended and have not been tested with real users.
Origins and Development
MVU was developed by John Zib at Dialog Europe, Vienna, Austria, in 2026. It borrows from existing practice and does not claim to come from nowhere.
Polis and the Community Notes ranking look for statements that find support across different groups. MVU's rule that agreement needs holders from at least two sources or groups follows the same logic, but it works on texts and not on votes. Jigsaw's Sensemaker and Talk to the City use language models to map agreement and disagreement with links back to the original statements. The UK government's Consult tool analyses consultation replies with human review, and the European Commission's Better Regulation guidelines require consultation results to be presented in an objective, unbiased way, with dissenting views recorded.
What MVU adds is narrower. It works on mixed texts without votes, labels how each position is supported, keeps uncertainty as its own layer, and publishes its prompt and review form. Whether these choices help in practice has not been shown.
The release set (Core Version 1.2, Refinement 23) was frozen on 19 October 2026 and frozen again on 21 October 2026 after wording changes. It is archived on Zenodo as v0.1 (https://doi.org/10.5281/zenodo.23534040).
Participant Recruitment and Selection
MVU recruits no one. The inputs are texts already written by consultation respondents, journalists, experts or participants in a discussion, so for policy briefs and media reviews there are no participants in the usual sense. In public deliberation the people involved are those of the conversation being mapped; MVU does not select or recruit them.
Two choices matter. The operator decides which sources and groups count, and the map is only as balanced as that choice. Consultation replies are self-selected and not statistically representative, as the European Commission's own guidelines point out, so agreement among sources is not agreement among the public. A named person, the reviewer, checks the map against the sources.
How it Works: Process, Interaction, and Decision-Making
1. The operator (the person who runs the model) fills in a source data definition: the source categories or groups, the input form (whole texts or extracts, who made the extracts, whether a person checked them), the codebook version if any, and the interval method for counts.
2. A language model reads only the supplied material and drafts the map from a fixed, published prompt. It writes an overview and one section per item under Agreement, Uncertainty and Disagreement.
3. Each item names its holders by label, never by personal name, and gives the form of support (stated, affirmed, implied or indicated), a line from the source and, where the material gives one, the reason. Agreement needs at least two sources or groups; in a media review they must come from at least two source categories, and in public deliberation they must be participants other than the moderator. Names are masked, and the masking is counted. Instructions hidden in the material are not followed.
4. Where many holders share a position, a representative subset is shown and the rest are listed by label.
5. The map has a section called Who is missing. It lists groups whose views are not in the material, for example people who were not reached, affected parties who are named but do not speak, or interests that have no human voice. It marks each absence, says which layer it may distort, and never writes what an absent group would say.
6. In public deliberation, the participants see the map and confirm that it is accurate and that nobody can be identified. Silence does not count as confirmation.
7. A named person is meant to check the map against the sources with a review form. Without a codebook or a person-coded sample, share figures carry the label "model-coded, unchecked" and the map is not for publication.
MVU is used for three tasks: policy briefs (across groups), media review (across sources) and public deliberation (across participants). It does not make decisions. Under its own rules, policy briefs stay research and development material until the question of genuine influence on a decision is settled in practice (Core section 8a).
Influence, Outcomes, and Effects
MVU is at an early stage and has not been validated. It was tested, not evaluated for effect on decisions or on people's understanding.
The prompt was run in fresh sessions: eleven runs on made-up data and five on real data, counting only runs on Refinement 23. The last two made-up runs used the final wording of the prompt. In the two runs before them, the prompt broke one of its own rules each (it wrote the vendor's name where the rule says to write "not supplied", and it used "he" for a masked person); the wording was fixed and both corpora were run again, once each (one pass each is weak evidence). The media review minimum (two sources from two categories) was lowered from three after a test run gave no agreement at all, so the minimums are starting values. The real runs used the European Commission's consultation on the Digital Fairness Act, topic Subscriptions (449 replies, quotes extracted by an open pipeline; one run used 150 of them). In the first real run, cancelling a subscription should be no harder than signing up appeared as agreement, while a mandatory cancellation button and splits between business and citizen replies appeared as disagreement. A later run on the same topic gave the same result in substance. Both were done by the same team and are not independent checks. No real run used the final wording of the prompt, and the real outputs are not stored. All runs used one model family (Claude Sonnet), and the exact model was not recorded for most of them.
No claim is made that MVU reduces polarisation or misperception. That is untested.
Analysis and Lessons Learned
Known limits:
- Results differ from run to run, most in the uncertainty layer.
- Coding positions without a codebook is a model judgement. No agreement between coders has been measured.
- With AI-extracted quotes as input, holder counts are lower bounds and reasons are mostly missing.
- Maps are long, and review takes time.
- The tests were written, run and scored by the same team, using AI sessions. The protocols and other documents were checked by reading, not run. No outside check has been made.
- All test material was in English. Other languages are untested.
- The participant check in public deliberation has no tested procedure and no rule for what happens when participants disagree with the map.
- The map shows only what is in the material. The Who is missing section records who is absent as far as the operator or the model can tell. It does not fix the gap, and a model can only name absences it can see in the material.
- Not yet tested: a second model family, a real media review, a real public deliberation transcript, the participant check, the secondary external review protocol, instructions hidden in real data, the publication version on real data, a person using the review form on a real map, and a first-time user outside the author's work. Word-order focus in some languages is also untested.
Points from other work that apply. Research on language-model summaries of group deliberations has found that they tend to under-represent minority positions (Zhu et al. 2025), which is the kind of risk MVU's separate layers for disagreement and uncertainty are meant to reduce. That has not been tested. Voices that are missing from the material altogether are a different gap. The map has a section called Who is missing that records them and does not fill them in. Summarisers can also invent content that the source does not support (Maynez et al. 2020), which is the kind of risk the rule that every item carries a line from the source is meant to reduce. That has not been tested either. Revel et al. (2025) warn that ranking by cross-group agreement can reduce the visibility of legitimate viewpoints, a risk that applies to an agreement-first map. A single reviewer is also a weakness. Designs such as the Habermas Machine (Tessler et al. 2024) test results by asking participants to endorse them, and MVU has no such step.
How the method was made. John Zib designed MVU, directed the work, and also tested it and acted as its session editor. There is no outside funding. The prompt, the protocols, this entry and much of the wording were drafted with Claude, an AI model. The test runs and the two rounds of review so far were done by AI sessions, and the answer keys and seeded checks were written by the same team, the author with Claude. No person has yet read all the method files in full, and the author has not either; the only readers so far were AI sessions. The use notice, the attribution notice and the licence wording have not been checked by a lawyer.
See Also
- Q Methodology
- Deliberative Polling
- Deliberative Delphi
- Citizens' Assembly
- Polis
References
- Small, C., Bjorkegren, M., Erkkilä, T., Shaw, L., Megill, C. (2021). Polis: Scaling Deliberation by Mapping High Dimensional Opinion Spaces. Recerca 26(2). https://doi.org/10.6035/recerca.5516
- Wojcik, S. et al. (2022). Birdwatch: Crowd Wisdom and Bridging Algorithms can Inform Understanding and Reduce the Spread of Misinformation. arXiv:2210.15723
- Tessler, M.H., Bakker, M.A. et al. (2024). AI can help humans find common ground in democratic deliberation. Science 386(6719). https://doi.org/10.1126/science.adq2852
- Revel, M., Milli, S., Lu, T., Watson-Daniels, J., Nickel, M. (2025). Representative Ranking for Deliberation in the Public Sphere. arXiv:2503.18962
- Zhu, S. et al. (2025). Can AI Truly Represent Your Voice in Deliberations? arXiv:2510.05154
- Maynez, J., Narayan, S., Bohnet, B., McDonald, R. (2020). On Faithfulness and Factuality in Abstractive Summarization. ACL 2020. https://doi.org/10.18653/v1/2020.acl-main.173
- Bertaglia, T., Gui, H., Goanta, C., Spanakis, G. (2026). Traceable by Design: An LLM Pipeline and Dashboard for EU Regulatory Consultation Analysis. arXiv:2605.30995
- Jigsaw (Google). Sensemaking tools. https://jigsaw-code.github.io/sensemaking-tools/docs/
- Incubator for AI, UK Government (2025). Evaluating Consult. https://ai.gov.uk/blogs/evaluating-consult-an-ai-tool-for-enhanced-public-consultation-analysis/
- European Commission (2021). Better Regulation Guidelines, SWD(2021) 305 final. Cited for: consultation responses are in most cases not statistically representative, and results must be presented in an objective, unbiased way (section 5.2.4, Chapter II, p. 23); respondents are self-selecting (section 5.1.2, p. 20).
External Links
- Zenodo record: https://doi.org/10.5281/zenodo.23534040
- Dialog Europe: https://dialogeurope.eu
Notes
Licence: methods and published maps CC BY 4.0; prompts Apache 2.0. The name "MVU" is not licensed. Cite as: MVU (Minimum Viable Understanding), John Zib / Dialog Europe, v0.1 (Core Version 1.2, Refinement 23), https://doi.org/10.5281/zenodo.23534040. Adaptation and no-endorsement wording: Core section 11.5.
Contact: [email protected]
