432 Matching Annotations
  1. Mar 2026
    1. Vignette exercises: respondents rate hypothetical people's life satisfaction based on descriptions, revealing how individuals anchor the scale and enabling cross-person calibration.

      Do they actually do this in the paper? doublecheck

    2. Calibration questions ask respondents to rate well-defined scenarios (e.g., "How satisfied would you be if you won $1,000?"). By observing how people rate the same reference points, researchers can estimate individual differences in scale use.

      Is this a reasonable examlpe? Do they ask questions like that in the exercises mentioend in the paper?

    3. Cost-effectiveness estimates vary by an order of magnitude depending on how WELLBYs are valued relative to DALYs.

      What's the source for this OOM claim?? Find and link it with a verbatim quote . #implement

      Also it's not in our 'evaluation summary as far as I know'

    1. Each scale point represents an equal welfare increment. If violated, summing is invalid and interventions targeting different baselines become incomparable.

      David Reinstein --- personally, this is the one I find least plauslible and most important.

    2. nterpersonal Comparability LSA = 7 ≈ LSB = 7 implies UA ≈ UB When two people report the same score, they experience similar welfare. Scale-use heterogeneity violates this assumption.

      I don't think this one is necessary if we can (instead) assume that differences are equivalent. For example, if we assume that person A is actually experiencing higher welfare at all levels of reported score, but the differences between the scores are comparable, then compared to interventions for measured differences in well-being, that shouldn't matter.

      I think it could also still be reliable if the distribution between the two populations is the same, even though we don't have specific inter-person comparability between any two compared individuals.

    3. 1 WELLBY = 1-point increase on a 0-10 life satisfaction scale × 1 person × 1 year W = Σi Σt LSit

      Those are not clearly defined here, nor the indexing

    1. We'll produce a practitioner-focused summary document, belief elicitation results with confidence intervals, and structured notes.

      Change this to "we hope to" and "We will share outputs". -- I can't guarantee right now that we'll get enough input or have bandwidth to produce this. #implement

    2. (Note: QALYs may be more directly comparable than DALYs for this purpose.)

      Leave out the QALYs parentheses bit here. Add "(or QALYs)" after "~1 SD in DALYs". #implement

    3. scale?

      Add "is a move from 1-3 for one person as good as a move from 1-2 for 2 people"? At the end of this paragraph... "even if these don't hold, does the linear WELLBY aggregation yield 'nearly as much value' for decisionmaking as other potential measures"? #adjust #implement

    4. When comparing a mental health intervention (measured in WELLBYs) to a physical health intervention (measured in DALYs)

      Either of these, especially the physical health intervention, could be measured either way. This overstates it a bit. Perhaps, just to give this as an example, suppose there is a case... #adjust #implement

    5. but more work is needed.

      "more work is neeeded" That's very much vague -- we nIt would be nice to have at least one specific point suggesting that the difference in scale means potentially matters and merits more study

    6. Each has strengths and limitations—and how they relate to each other, and whether either reliably captures what matters for human welfare, directly affects which interventions get prioritized.

      I'm allergic to platitudes. IIRC you should have some notes somewhere providing at least one case where this matters .

  2. Feb 2026
    1. adversarial manipulation.

      I don't think we discussed adversarial manipulation or have any results on it, so I'm a little worried that whatever generated this discussion is doing a sort of generic pandering and putting in what it generally expects to see in papers like this.

    2. Our results support AI as structured screening and decision support rather than full automation,

      This seems like a sort of milquetoast generic caveat. In what sense is this what our AI results support? This seems a bit pandering.

    3. xhibiting consistent failure modes: compressed rating scales, uneven criterion coverage, and variable identification of expert-flagged concerns.

      I'm guessing this is a bit premature/too much rounding up a few observations to general conclusions, but let me look at the results a bit more carefully.

    4. often approach the ceiling implied by human inter-rater variability on several criteria,

      This is interesting and strong. It comes across maybe a little bit overstated, so we just need to be careful about how we're framing this result.

    5. high-quality but noisy reference signal

      I think this is right, but the term "reference signal" sounds technical in an information theoretic sense, and we want to make sure we're not misapplying it.

    6. narrative critiques

      Yes, we focus on the critiques here, but the on journal evaluations do more than just critique. They discuss, they offer suggestions, implications, et cetera.

    7. overing economics and social-science working papers

      "covering ... working papers" Is mostly accurate but not quite right. We don't cover all working papers, and we have a specific focus on research relevant to global priorities. We can also evaluate post-journal publication, but I'm not sure how to best summarize this in a simple way in the abstract.

      The idea of "open evaluation platform" also could be a bit confusing here because it's not mainly about crowd sourcing. Yes, the "paid expert review packages" cover this, but I don't quite think this is worded in the best possible way.

    8. Peer review is strained, and AI tools generating referee-like feedback are already adopted by researchers and commercial services—yet field evidence on how reliably frontier LLMs can evaluate research remains scarce.

      This is a decent first sentence, although it bears the marks of AI-generated text. But also I'm not sure if it's really in line with our newest spin on this.

  3. Nov 2025
    1. returned file id keyed by path, size, and modification time.

      what does this mean? "Keyed by" ?

      This implies it is kept on the server and won't need a later upload.

  4. Sep 2025
    1. Zhang and Abernethy (2025) propose deploying LLMs as quality checkers to surface critical problems instead of

      Is this the only empirical work? I thought there were others underway. Worth our digging into. Fwiw I can do an elicit.org query.