Blog · arXiv Analysis · Published: July 10, 2026 · Modified: July 10, 2026 · Last reviewed: July 10, 2026

The Curated Corpus Becomes the Public Answer

Hafsteinn Einarsson and coauthors' July 2026 arXiv paper compares curated retrieval with open web search for a public AI information service answering civic questions in Iceland.

For this essay, a source-trust receipt is the record that shows which sources a public AI service was allowed to retrieve, which sources it actually cited, and which cited sources expert reviewers judged unfit.

The Paper

The paper is Curated retrieval versus open web search in public AI information services: a coverage-trust trade-off, arXiv:2607.05217 [cs.CY, cs.CL, cs.IR]. The arXiv abstract and API list the authors as Hafsteinn Einarsson, Hafsteinn Birgir Einarsson, Jón Gunnar Ólafsson, and Jón Gunnar Þorsteinsson. Version 1 was submitted on July 6, 2026, and version 2 was posted on July 7, 2026. The downloaded PDF reports 38 pages and University of Iceland affiliations.

The point is not "RAG good, web bad." The paper shows coverage and source trust pulling against each other: open web search reaches more questions while weakening control over institutional citations.

The Service

The case is Evrópuvefurinn, described by the paper as an independent, government-funded service run by the University of Iceland. It answers questions in Icelandic about the European Union and Iceland's relationship with it. The study was conducted before public relaunch, as Iceland prepared for a referendum scheduled for August 29, 2026, on whether to resume EU accession talks.

The researchers built 287 fixed questions rather than drawing from live user traffic. Each question could be answered through a curated local corpus of Evrópuvefur articles or through open web search. Five domain experts reviewed outputs from early May to the start of July 2026, producing 551 scored evaluations, 449 distinct AI-generated answers, and 128 source flags.

The Trade-off

The curated path behaved like narrow institutional memory. When the corpus did not contain enough material, the system often declined to answer. That visible failure is also a governance asset: it exposes the archive's edge instead of filling the gap with whatever the web provides.

The web-search path behaved like a reach machine. It answered more questions, but in more than a third of reviewed web-search answers, 35 percent or 65 of 187, at least one cited source was flagged. The paper says those web flags were almost always for untrustworthiness or irrelevance. Curated sources were flagged much less often, and the flags there were for being out of date rather than for source credibility.

Ordinary answer quality did not reveal the source problem. The paper reports that fluency and topical fit did not predict source trustworthiness. A polished civic answer can still rest on a source an expert would not want a public institution to cite.

Prompt Steering

The study also tests whether a trusted-domain list in the system prompt steers citation behavior. The ablation used the same 287 study questions and reports that the trusted-domain list increased citations to listed domains from 12 percent to 21 percent. That is a measurable nudge, not a control.

One citation absence is especially useful as a warning sign. Across all 287 web-search answers in the deployed-system evaluation, the system never cited RÚV, which the paper describes as the public broadcaster and the country's most widely used news source. A public AI service can therefore have a trusted-source policy and still route around a central public source. The receipt has to record actual citations, not merely intended preferences.

The Receipt

A source-trust receipt for public AI should include the service mandate, corpus boundary, retrieval mode, search provider or plugin, model route, prompt source policy, date of retrieval, full citation list, flagged source, flag reason, reviewer role, and whether the answer was published, revised, declined, or escalated.

The receipt should also separate three judgments that answer engines collapse: whether the answer addresses the question, whether the cited source supports the claim, and whether the public institution should stand behind that source. Current, relevant, unsuitable, trusted, and stale are different states, and each needs a different remedy.

Governance Reading

The Spiralist reading is that retrieval is an editorial act even when nobody calls it editing. Once a public AI service cites a page in a civic answer, that page borrows institutional weight.

This belongs beside answer engines becoming the front page, AI search reshaping referral power, citation verification as a reward signal, retrieval-augmented generation, and AI governance. The shared lesson is that citations are not decoration. They are the institutional surface where authority, search, and public memory meet.

Limits

The paper is a preprint and studies one service, language, topic, political moment, and system configuration. The questions were generated to span the debate, not collected from real users. Reviewers worked through a shared queue, so the RAG and web modes were reviewed in unequal numbers, and the authors describe the mode comparison as observational rather than controlled.

The paper also cautions that most source flags were individual expert judgments rather than adjudicated rulings. The prompt ablation measured citation-list compliance, not expert-reviewed trustworthiness. Those limits narrow the conclusion: public AI services should measure source trustworthiness in their own configuration instead of importing confidence from a benchmark or prompt.

Source Discipline

This page treats the arXiv abstract page, API record, HTML version, and downloaded PDF as primary sources. It does not reproduce the paper's prompt appendix, reviewer comments, or source list beyond minimal identifying facts. Numerical and bibliographic claims above are limited to facts verified in those records.

The disciplined question is not "does it cite sources?" It is: which sources were reachable, selected, ignored, flagged, and institutionally endorsed?

Sources

Return to Blog