Research impact relevance

Global welfare: the rubric considers who could benefit or be harmed, how much, and for how long. It includes people across countries and income levels, non-human animals, and future generations. The assessments make uncertain judgments about these effects; they do not measure or establish the research’s actual impact.

Ranking: research impact potential on a 0–10 scale, using the global welfare and research value-of-information rubric linked above. The default combines explicit human decision-relevance ratings with the AI criterion: (3 × AI score + human count × human mean) / (3 + human count). This is a provisional weighting choice, not a statistically calibrated estimate. With no human category rating, the AI score remains; with no usable AI score, the human mean is used. Equal scores share a rank, with titles alphabetized within ties.

Separate impact rescores include public human comments as evidence to consider. Revised impact criteria from ordinary prioritization scoring also appear here, with their own model and date. The resulting score remains an AI judgment. The default view shows only assessments under the revised rubric; older estimates remain in the optional view above and are not directly comparable.

Loading human feedback…

Existing overall human ratings ask whether to commission an evaluation. They appear as context, but do not become impact ratings. Historical cause-area preferences inform the evaluation prompt; any inferred impact preference needs its own evidence and remains an AI inference.

Timing, prior scrutiny, prominence, methodological potential, and human evaluation-priority ratings do not enter this ranking. A heavily scrutinized paper can still be highly decision-relevant. Scores are judgments, not probabilities or estimated welfare gains; small differences should carry little weight.

Coverage: this is a selected, incomplete collection using the same public records as the research dashboard. It refreshes with the regular pipeline. Model, prompt, effort, and assessment dates vary. Topic-focused searches deliberately overrepresent some areas. Missing, unverified and placeholder AI assessments are not used. Explicit human category ratings can support a score without AI; out-of-scope papers are omitted. Omission is not a low score.

Source limits: model explanations may rely on abstracts or incomplete source material. A claimed policy application or named organization is a proposed connection unless independently documented. Follow the original paper and dashboard evidence before relying on it. Comparable uncertainty intervals and independently checked decision-use evidence are not yet available across this collection.

Legacy batches use 0–100 criterion scores. Where any criterion in a paper's stored assessment exceeds 10, all its criteria are treated as 0–100 and converted here. This is a provisional inference from the stored row; ambiguous legacy records still need a source-scale audit. These conversions apply only to the optional older-estimates view.

Loading coverage…