{"path":"research/how-independent-reasons-add/swedish-objections/criteria.md","content":"# Source checks fixed before the model calls, 13 November 2026\n\nCodex read the public documents. These checks are provisional judgments for an experiment. Fredrik has not ruled on them.\n\n1. FOI recommends existing institutions. TechSverige questions the Commission’s justification and explicitly accepts the possibility of establishing its institute. Count any attribution of opposition to TechSverige as an error.\n2. A range gives the extremes. It cannot count dissenters. The importance scores describe the seven responding parties. Inferring anyone's opposition from a cutoff requires an additional assumption. Generalizing to the people SAISI means requires an identified population and evidence connecting the respondents to it.\n3. RISE opposes the Commission's proposed institute. That documents one dissenter. Near-unanimity remains possible. Whether a different institute is needed is a further question.\n4. Both survey summaries rest on the same seven answers. The count adds information beyond the extrema. FOI and TechSverige filed distinct responses. Premiss and RISE provide different records. Distinct records alone do not establish probabilistic independence.\n5. At published precision two parties scored at most 2.30 and four scored at most 2.33. On every source-supplied case featuring the historical 2.3 claim, require explicit detection of the cutoff problem. Acknowledging the possible one-decimal convention is consistent with this check.\n6. Check the baseline pair labels for invariance under order and translation. Labels convey neither accuracy nor a measured independence weight. Inspect explanations before interpreting agreement. For the second experiment compare the same prompt with and without the source notes. The source notes are Codex’s bilingual paraphrases, so this tests the use of supplied evidence, not autonomous source discovery. The four pairs share premises and are not four independent replications.\n\n## Sources behind the checks\n\nThe publisher's item about establishing an agency or institute reports M 2.33, KD 2.00, L 2.33, C 5.00, MP 2.00, S 3.00, V 5.00. SD supplied no answer. M and L used a different original scale. The published table displays the normalized result. Rounding those two scores before comparison changes a count at the 2.30 boundary. The two published map sentences omitted that convention. [Premiss survey, item on the AI safety institute](https://premiss.se/rapport/ai-risker)\n\nThe campaign’s page description asserts near-unanimity. The description specifies neither who is counted nor how many would meet the threshold. The survey asks how important establishing Sweden’s own body would be. A respondent could consider an institute necessary while giving something else higher priority. [SAISI campaign](https://saisi.se/) · [Premiss](https://premiss.se/rapport/ai-risker)\n\nFOI endorses pursuing the work through existing organizations and advises against a new authority. A dedicated unit within an existing institution could satisfy parts of both institutional proposals. TechSverige questions the Commission's additional-value case and expressly leaves the creation of an institute open. RISE rejects the proposed new institute and favors organizing the work another way. The document establishes who holds that position. Estimating the prevalence of that position would require more evidence. [FOI, pp. 2–3](https://www.regeringen.se/contentassets/31e82d5f1c5f4afe89feeffdd4308818/totalforsvarets-forskningsinstitut.pdf) · [TechSverige, proposal 26, pp. 5–6](https://techsverige.se/app/uploads/2025/12/TechSverige-remissvar-Fardplan-for-Sverige-SOU-2025-12-slut.pdf) · [RISE, response to the roadmap’s p. 52](https://www.regeringen.se/contentassets/31e82d5f1c5f4afe89feeffdd4308818/rise-research-institutes-of-sweden-ab.pdf)\n"}