{"path":"research/how-independent-reasons-add/swedish-objections/README.md","content":"# Can two Swedish objections be weighed together?\n\nThe source check found an error in both public maps. Four of the seven responding parties scored the proposal at or below **2.33**. Only two scored it at or below **2.30**. Correcting that figure still leaves the central inference open. The survey asked how important the proposal was. Even all seven parties could agree Sweden needs an institute and give its establishment different priorities. The model passed 14 of 14 invented controls, then only four of eight checks on the actual sources. I would use it to identify what needs checking before a person reviews the two links. [Premiss survey](https://premiss.se/rapport/ai-risker) · [Real cases](survey-controls.jsonl)\n\n## First iteration · ask the existing experimental question\n\nOn 13 November 2026 Codex took seven claims verbatim from the public Swedish map and their seven English translations. They make four pairs across two conclusions. The existing experimental prompt asks whether both objections bear directly on the claim, both undermine evidence, or the answer is mixed or unclear. Run each pair in both orders. The expectations for reading the explanations were fixed before the calls. [Criteria](criteria.md) · [Inputs](published.json) · [Swedish map](https://deliberus.com/extractions/sverige-och-ett-ai-sakerhetsinstitut) · [English map](https://deliberus.com/extractions/sweden-and-an-ai-safety-institute)\n\nThe model labeled 15 of 16 pairs direct. Seven of eight pair-and-language combinations kept their label when reversed. Reversing the order in Swedish changed the answer on the party count and RISE. The apparent consistency hides a material change of meaning. In both languages, when the TechSverige statement came first, uncertainty about what the institute would add became an assertion that it adds nothing. Several answers inferred the absence of near-unanimity from the range alone. A population of a hundred people with ninety-nine identical maximum scores and one minimum score can span the full scale. [Baseline readings and receipts](baseline.jsonl)\n\n**Decision after iteration 1.** Record what each objection establishes and what could still hold. A model can recover those limits from the wording in some cases. It cannot recover the exact counts from reported extrema. Compare the same open questions with and without primary-source notes, keeping the map sentences, order and model fixed. The historic map sentences keep their original decimal precision throughout this experiment.\n\n## Second iteration · ask what each objection leaves possible\n\nWith the map sentences alone, both Swedish and English answers now allowed for TechSverige’s concern to be satisfied by a better justification. All four answers that included RISE allowed wide agreement to coexist with one dissenting voice. An important error remains. Both languages sometimes called the seven respondents divided on whether the institute was needed, even while recognizing that a low importance score need not express opposition. The English answer also weakened FOI's recommendation that existing actors *should* do the work into a claim that they *can*. All eight answers recorded the origin of the two observations as unknown. The input strips the source names, so this abstention is defensible. [Readings and receipts](bare.jsonl)\n\n**Decision after iteration 2.** Supply a short account of each source, with the publisher’s actual table where applicable. Use the same questions. Check whether the model preserves the distinction between a source saying an inference is weak and the mapper deciding how much force that should carry.\n\n## Third iteration · supply the sources\n\nThe provenance answer matched the source records on all eight language-and-pair combinations. Both readings of the FOI and TechSverige pair preserved TechSverige's express lack of opposition. On the four readings that received the erroneous cutoff and the table, three explicitly flagged the rounding issue. The fourth silently changed the threshold to 2.33 in its explanation. Even a flagged answer could contradict itself. One Swedish reading said only three parties scored higher than 2.3. Five did. Every answer for the consensus target still made some move from the parties' ratings to a statement about agreement among them. In several cases the same answer later acknowledged that rating importance and agreeing about need can come apart. Provenance was right while the inference remained questionable. [Readings and receipts](sources.jsonl) · [Source notes](source-notes.json)\n\n**Decision after iteration 3.** Use an unseen, invented example to distinguish a mismatch in the measured property from weak evidence that genuinely bears on the target. Use ten named-as-a-group respondents, a precise threshold and a fixed distribution. Change the survey question or the target while keeping the count fixed. Also check unanimous versus near-unanimous agreement when one dissenter is observed. For this contrast ask only what the recorded answers logically settle; evidence may still affect belief where it leaves the conclusion open. This is an evaluation question about scope, separate from the deployed independence microtask.\n\n## Fourth iteration · isolate the inference about agreement\n\nA different, invented local-archive survey supplied seven paired tests per language. The polarity of the evidence, the threshold and the membership of the group were explicit. All 14 labels matched the frozen answers. In both languages the model gave compatible worlds for the case that matters here. All ten delegates could see a need, yet four could prefer to spend this year’s money elsewhere. Alternatively, the same four could believe the town has no need for an archive. Both groups could give exactly the same priority answers. For the reverse case, prioritizing something one believes unnecessary is unusual, but the record does allow such a preference. These results test logical compatibility. The actual independence microtask asks how much two reasons add under plausible rival explanations. [Controls and receipts](contrast.jsonl)\n\n**Decision after iteration 4.** Return the same entailment question to the real survey, with the referenced population taken from each target. Test the campaign’s exact target against the table and against the one consultation response. Keep two numerical targets as controls, one correct at source precision and one which crosses the cutoff. This checks whether the narrow question transfers before proposing any further development.\n\n## Fifth iteration · return to the actual propositions\n\nFour of eight answers matched, with the same outcomes in both languages. The model correctly left near-unanimity open given RISE’s objection and accepted four scores at or below 2.33. It said the party survey contradicted near-unanimity about need. Even among those same seven parties, every answer is compatible with every party saying an institute is needed. Their priorities could differ for reasons of cost, timing or competing demands. The model also accepted four scores at or below 2.30, rounding the observations despite the explicit instruction to use the table’s precision. Three explanations for English inputs came back in Swedish. This is a small transfer check with one model, mapper-supplied bilingual descriptions and previously examined material. The two versions reflect one set of underlying observations. [Real-case readings](survey-controls.jsonl) · [Exact questions](survey-controls.json)\n\n**What I would use now.** Let a reader identify the source of each objection and offer the specific inference for inspection. For a quantitative assertion, check the operator, published values and denominator arithmetically. For a claim about agreement, ask which people were counted, what those people were actually asked and what further assumption turns their answers into the proposed conclusion. An uncertainty about the justification can be a genuine objection. One respondent can supply some evidence about a population. The present results give no calibrated strength for either, and the two survey summaries reuse the same respondents. I would have a human read these two links before letting them affect the institute map’s displayed strength. For the prospective coordination map, the analogous question is whether agreement that a risk matters establishes agreement about which action to take. That would make a useful next real-case evaluation.\n\n**Map correction.** The local Swedish and English map builders and exports now give M and L as 2.33 and four parties at or below 2.33. Two link explanations now state what the survey range and low scores can support. The rebuild yielded the same 250 claim keys and 334 relationships per language. The published pages supplied the earlier snapshot and were not updated during the experiment. The Swedish builder also now emits the current microtask fixture field name. This makes regeneration reproduce the existing checked-in fixture byte for byte.\n\n**Cost and reproducibility.** On the direct Anthropic route Claude Sonnet 5.5 served 54 calls, one attempt each. The router’s computed charges by iteration were $0.0435840, $0.0351498, $0.0525592, $0.0391080 and $0.0264260, totaling **$0.1968270**. Combined with the preceding polarity study, the recorded total is $0.4033613. The production independence prompt and graph received no changes in this study. Run `python3 verify.py` from this directory for an offline check of the two counts and all saved outcomes. From the repository root, the two runners accept fresh output paths with `PYTHONPATH=.` and the project Python environment. The rubric is Codex’s source reading. Fredrik has not yet assessed these judgments. [Sources and rubric](criteria.md) · [Verification](verify.py) · [Previous study](../polarity-eval/README.md)\n"}