structured measurement schemas (Asirvatham, Mokski, and Shleifer 2026), iterative quality-checking workflows (Zhang and Abernethy 2025), or the kind of prompt-robustness engineering motivated by specification-search concerns (Asher et al. 2026)—should improve further.

can achieve

with no iteration, retrieval augmentation, chain-of-thought scaffolding, or multi-step agentic loop.

deliberately

deliberately

strong case that frontier LLMs can serve as additional expert raters in structured evaluation pipelines,

Our headline finding is that the best-performing model (GPT-5 Pro) matches or exceeds pairwise human inter-rater rank agreement on overall quality,

against expert evaluations for 60 economics and social-science working papers

multi-dimensional

These developments make the evidentiary gap salient: funders, editors, and policymakers need to know when AI evaluation outputs are trustworthy enough to use, and when they are unstable, biased, or manipulable. Recent work highlights all three concerns. First, reproducibility can be “jagged”: repeated runs of the same models on the same corpus over time can be highly consistent for some tasks and models, but much less so for others (Thomas, Romasanta, and Pujol Priego 2026); robustness may require separating scientific judgment from computational execution (Xu and Yang 2026); and even without overt adversarial intent, subtle reframings of the same task can induce systematic shifts in outputs—a form of LLM “specification search”—raising concerns about frame-sensitive biases when models serve as measurement instruments (Asher et al. 2026). Second, adversarial manipulation is not hypothetical: invisible-text “prompt injection” can substantially inflate LLM-assigned review scores and acceptance recommendations in simulated peer review (Choi et al. 2026), and prompt-injection vulnerabilities are also documented in other high-stakes advice settings (Lee et al. 2025). Third, even when outputs look fluent and plausible, it remains unclear whether AI models approximate expert judgment: AI-generated reviews tend to cover more surface-level sections while being less thematically diverse and less focused on interpretation, originality, and applicability than human reviews (Rajakumar et al. 2026); LLMs used as manuscript quality checkers identify only a small fraction of confirmed critical errors even with the strongest reasoning models (Zhang and Abernethy 2025); and LLM scoring exhibits systematic range restriction and halo effects that can distort agreement metrics (Wang et al. 2025).

Meanwhile, publishers are formalizing policies that treat manuscripts and reviews as confidential and prohibit reviewers from uploading them into general-purpose generative AI tools

under strain

At 20 kTA reference scale

CAPEX scaling

Scalable GF technology 50% Switches to “cheap” GF prices Pivotal uncertaint

yments:

The slider complements this by letting users explore “what if progress is partial?” scenarios.

If any one of these succeeds at commercial scale, the “cheap” price regime applies

$0.20 – $1.20

Consider correlated scenarios via the maturity slider

Use for relative comparisons rather than absolute predictions

Equipment depreciation period

Weighted average cost of capital

Breakthrough technologies that could trigger the “cheap” scenario: - Autocrine cell lines (cells produce own FGF2) - Plant molecular farming ($1-10/g target) - Precision fermentation at scale - Polyphenol substitution (reduces GF requirements by 80%)

30-200 g/L Final biomass at harvest Cycle time 0.5-5 days Time per production batch Media turnover 1-10 ratio 1=batch, >1=perfusion

Distribution

Why correlate? In “good worlds” for cultured chicken: - Technologies are more likely adopted (higher P) - Custom reactors are more common (lower CAPEX) - Financing is cheaper (lower WACC) This prevents unrealistic scenarios where technology succeeds but financing remains prohibitively expensive.

The model uses a latent maturity factor (0–1) to correlate technology adoption, reactor costs, and financing: Padopted=bound(Pbase+k⋅(m−0.5),0,1) What does “bound” mean? bound(x, 0, 1) ensures the result stays between 0 and 1. Also c

The GF progress slider interpolates between current and target prices: PGF=Pcurrent×(0.01)progress At 0% progress: current prices ($5,000–500,000/g) At 100% progress: target prices ($1–100/g for cheap scenario)

Example calculation: - Cell density: 50 g/L → need 1000/50 = 20 L per kg - Media turnover: 3× (perfusion system) → 20 × 3 = 60 L/kg - Media price: $0.50/L (hydrolysates) → 60 × 0.50 = $30/kg

uld the resulting adjustments change the cost-effectiven

WELLBY is a reasonably useful measure in this context

Given the available collected data [...], how should [funders] measure the impact on wellbeing? [...] What measures of well-being should charities, NGOs, and RCTs collect for impact analysis?

How reliable is the WELLBY measure [...] relative to other available measures in the 'wellbeing space'? How much insight is lost by using WELLBY and when will it steer us wrong?

More detailed questions on WELLBY reliability

"Meaningful change" = at least one intervention currently in the top 5 moves out of the top 5, OR the #1 ranked intervention changes. This assumes future RCTs incorporate these methods and Founders Pledge updates their CEA accordingly.

If you propose a measure other than linear WELLBY in your answer above, how much more would it cost to achieve the same welfare improvement using linear WELLBY instead?

Currently, standard practice (used by HLI and Founders Pledge) treats SDs on different mental health instruments as interconvertible with WELLBY SDs on a roughly 1:1 basis.

calibration questions can partially adjust for it.

Key assumptions

If it's unreliable or systematically misleading, billions of dollars in funding decisions could be poorly directed.

Is the WELLBY (linear, 0–10 life satisfaction) a useful and reliable measure for comparing interventions—particularly those involving mental health, consumption, and life-saving—in the context that organizations like Founders Pledge use it?

Benjamin et al. show that calibration questions

effective altruism-aligned

How the workshop is structured

3. Could calibration questions improve things?

"calibration questions"

If "7 out of 10" means something very different to different people, that's a fundamental challenge for the WELLBY as a tool for comparing interventions.

This is one of the most important recent papers

Two measures dominate these analyses. The DALY (disability-adjusted life year) comes from health economics and captures years of healthy life lost to disease or disability.

Founders Pledge, GiveWell, and Open Philanthropy

Our highest priority will be to avoid wasting author time. We’re very cognizant from first-hand experience that poor conversion quality, perhaps requiring back-and-forth with the author, is very unpleasant and a huge time suck.

Following acceptance, authors may pass their manuscript to the journal in any reasonable format (LaTeX or markdown preferred; Word and PDF acceptable).The document will be published in a “web-first” format, such as the Distill version of R Markdown.This allows reflowable text and mobile readability.We currently do not plan to support interactive content, as we do not think the large effort is worth the modest benefit.

The review process will be done using a manuscript in PDF format, which can be generated by the authors using whatever software they prefer (e.g., LaTeX). This avoids wasting the time of authors of papers that are later rejected.

The journal Alignment will be a fast and rigorous venue for theoretical AI alignment—research on agency, understanding, and asymptotic behavior of advanced and potentially self-modifying synthetic agents

Many potential criticisms of papers are “NP” (can be checked easily), so credentials of reviewer should be irrelevant

If confidential: Massive reviewer effort (the report) boiled down to a single bit (!)

. Our bet is that we can could unify and expand the field of alignment by establishing a legitimate academic journal with an unorthodox review pipeline.

Since they are regarded as informal by institutional academia, time spent on such outputs is dead time, from the perspective of institutional research performances indicators and career progressio

Our experimental solution to address this problem is to publish each accepted paper with a “reviewer abstract”.  Its main goal is to help a potential reader decide — on the paper’s merits — if the paper is worth reading.

Reviewer Compensation

We think it’s very reasonable to spend an average of ~$3k per paper on reviewer payments.

We intend to experiment with LLM recommendations to surface candidates that might not be salient to the editors.

Author identity known to reviewers

If a submission is published in the journal, the AF post is updated to reflect this, and the reviewer abstract is added. The reviewer abstract can be upvoted on AF, with the reviewers with AF accounts who sign the abstract receiving karma as appropriate.

Journal Not Conference

We are tentatively planning on making the journal archival, meaning that publication there constitutes the “version of record”, in contrast to a workshop publication. (Preprints of course are allowed.)

Public review avoids this, but introduces additional problems due to lack of confidentiality: less honest, more combative and defensive conversations between authors and reviewers. Public review also produces an artifact that is poorly suited to a reader because the c

Paper-by-Paper Comparison

Full LLM evaluations, human adjudication, new human evaluations

Research Goals and Questions (overview)

Table 3.1: Token usage and estimated cost per model

To assess whether different LLMs produce systematically different evaluations, we collected ratings from multiple providers: OpenAI (GPT-4o-mini), Anthropic (Claude Sonnet 4), and Google (Gemini 2.0 Flash)

?tbl-llm-token-cost-summary

We first use the earlier GPT‑5 Pro evaluation run that covered all papers in our Unjournal sample with a simpler JSON‑schema prompt

temporal “data leakage” between the multi‑year regrowth label and contemporaneous predictors;

We start by examining selected evaluations in detail. In the next step we will juxtapose these LLM assessments with the human evaluators’ written reports.

sensitivity of return‑on‑investment calculations to assumptions about donor lifetime value and unobserved costs.

potential spillovers and spatial correlation across postal codes

Case study: Williams et al. (2024)

not fully propagate uncertainty

additionality and opportunity costs in the policy framing,

the treatment of “biophysical potential” versus business‑as‑usual regrowth,

edictors; incomplete or optimistic treatment of uncertainty around the headline 215 Mha estimate; a broad and permissive definition of land “available for natural regeneration”; limitations of the carbon overlay and permanence assumptions; and only partial openness of code and workflows, which increases barriers to full replication.

relatively

This October 2025 run asked the model only for numeric ratings and journal‑tier scores (no diagnostic summary or reasoning trace);

To understand what GPT‑5 Pro is actually responding to, we re‑ran the model on four focal papers (Adena and Hager 2024; Peterman et al. 2024; Williams et al. 2024; Green, Smith, and Mathur 2025) using a refined prompt (as shown in the previous section).

temporal leakage from contemporaneous predictors,

ourth, when sufficient information to compute a standardized mean difference (SMD) was lacking and the text reported a “null,” outcomes were set to an “unspecified null” of 0.01. This imputation is transparent but ad hoc; it could bias pooled estimates upward (relative to zero) and may not reflect the true variance of those effects. The manuscript would benefit from sensitivity checks setting these to 0, excluding them, or modeling them with conservative variances.

Model assessment summary

For “tier_will,” given its status as a WZB discussion paper and the need to disregard actual publication knowledge, I might predict it will land around 3.2 to 4.0.

clearer pre-analysis plan deviation trackin

Heterogeneity analyses suggest stronger effects in urban areas and in PLZs with higher employment, more children, and more Catholics, and with higher predicted giving potential. These patterns can guide targeting but also indicate that the ITT estimates average over meaningful heterogeneity.

Adena and Hager 2024

Data construction choices appear reasonable but introduce some judgment calls. Winsorizing PLZ-day donations at €1,000 reduces variance from heavy tails; the authors show that results are directionally robust, but precision trades off.

The most important methodological limitations concern exposure heterogeneity and spillovers. Treatment is assigned at the PLZ level, but impressions are probabilistic and sparse (roughly one in ten Facebook users in treated PLZs received at least one impression), so the estimates are ITT and likely attenuated relative to the effect of actually seeing the ad; the TOT is not estimated. The allocation strategy partly allows Facebook to endogenously concentrate impressions, creating within-treatment variation in exposure that is not exploited for causal TOT analysis (e.g., using randomized budgets as an instrument in a dose–response framework). Spillovers across PLZs are plausible (algorithmic leakage of geotargeting and social diffusion). The authors document positive “share of treated neighbors” effects and argue the main estimates are lower bounds, but the neighbor-treatment share is not itself randomized, and spatial correlation or common shocks could inflate these coefficients; the spillover analysis should be interpreted cautiously. Robustness to spatial correlation in errors is only partly addressed by robust standard errors and randomization inference; alternative SEs (e.g., spatial HAC or clustering at larger administrative units) and placebo geographies would further strengthen inference.

claims & evidence, methods, logic & communication, open science, global relevance, and an overall

For each of 47 such papers,

In this project, we test whether current large language models (LLMs) can generate research evaluations that are comparable, in structure and content, to expert human reviews.

Furthermore, a key promise of AI is to directly improve science and research.

works

a high‑stakes, policy‑relevant domain, and as the first step toward a broader benchmark and set of tools for comparing and combining human and AI research evaluations.

model reliably identifies many of the same methodological and interpretive issues

and to produce a narrative assessment anchored in the PDF of each paper.

Comparing LLM and human reviews of social science research using data from Unjournal.org

Strong pantropical mapping, but several methodological and interpretive risks remain. Training data on natural regrowth include substantial omission error in humid biomes; pseudo-absences and misclassification may bias the model. Random forests were trained on class-balanced points and probabilities are treated as calibrated; no prevalence correction or probability calibration is shown, yet expected areas/carbon rely on these values. Validation is not fully spatial; accuracy likely inflated by autocorrelation (declines at greater distances); no formal spatial block cross-validation or predictive uncertainty mapping. Reported “confidence intervals” for total area/carbon are effectively deterministic sums, not uncertainty; overall uncertainty is understated. Predictions at 30 m depend on several coarser predictors (300 m–1 km), so effective map resolution is coarser and may mislead fine-scale planning. Final maps omit socioeconomic predictors (despite similar accuracy), assuming stationarity from 2000–2016 to 2030 and potentially overstating practical feasibility. Carbon estimates exclude permanence/leakage dynamics and use coarse downscaled inputs. Data products are open, but code is only “on request,” limiting full reproducibility.

78

Show cod

(only partially available)

Inter-rater reliability (Cohen’s κ): We treated the AI as one “reviewer” and the average human as another, and asked: how often do they effectively give similar ratings, beyond what chance alignment would predict? For overall scores, Cohen’s κ (with quadratic weighting for partial agreement) came out around 0.25–0.30. This would be characterized as “fair” agreement at best – a significant gap remains between AI and human judgments. For most specific criteria, the weighted κ was lower, often in the 0.1–0.2 range (and for one category, effectively 0). A κ of 0 would mean no more agreement than random chance (given the distribution of scores), so some categories are bordering on that. In contrast, typical inter-human agreement on these kinds of scoring tasks can also be low, but usually one would hope for κ in the 0.3–0.6 range among trained reviewers on well-defined criteria. Our finding of κ < 0.3 in all cases suggests that the AI’s ratings are not interchangeable with a human reviewer’s – at least not without further calibration.

Correlation (Pearson’s r) between the AI’s and human scores across papers: This tells us, for example, if a paper that humans gave a high score also tended to get a high score from AI (regardless of absolute difference). For the Overall scores, Pearson r ≈ 0.30, indicating a weak-to-moderate positive correlation.

r ~ 0

Table 3.1 show agreement metrics across rating criteria. To quantify the agreements and differences observed, we calculated several statistics comparing LLM scores to the human scores, aggregated by criterion:

has a distinct “taste,” elevating some work and devaluing other work differently than human referees.

or example, Aghion et al. 2017 was among the top few for human reviewers, but the LLM overall score put it notably lower relative to others, hence a downward green curve.

high quality due to its robust modeling and accessible methodologies, making it relevant for climate mitigation and biodiversity

Comparing its midpoint percentiles to a reference group of serious research from the last three years will help inform this.

Shared notes

on the right, the papers are ordered by the AI’s overall score (rank 1 = highest rated by AI)

For example, in the Logic & Communication column, we see many light-orange cells – the AI often thought papers were a bit clearer or better argued (by its judgment) than the human evaluators did.

deed, GPT often noted lack of code or data sharing in papers and penalized for it, whereas some human reviewers may have been more forgiving or did not emphasize open-science practices as strongly (especially if they focused more on content quality). As a result, for many papers the AI’s Open Science score is 5–10 points below the human average.

Figure 3.2: Relative ranking (overall) by LLM and Human evaluators

he brief rationales clarify what evidence in the paper drove each score.

- tier_should = where the paper deserves to publish if quality-only decides. - tier_will = realistic prediction given status/noise/connections.

Round all scores to the nearest 0.5

We use a single‑step call with a reasoning model that supports file input. One step avoids hand‑offs and summary loss from a separate “ingestion” stage. The model reads the whole PDF and produces the JSON defined above. We do not retrieve external sources or cross‑paper material for these scores; the evaluation is anchored in the manuscript itself.

• Default = arithmetic mean of the other six midpoints (rounded).

plus a short rationale

The credible intervals communicate uncertainty rather than false precision

# Environment setup

Direct ingestion preserves tables, figures, equations, and sectioning, which ad‑hoc text scraping can mangle. It also avoids silent trimming or segmentation choices that would bias what the model sees.

he sample includes r dim(research_done)[1] papers

49

per criterion (and noting the range of individual scores).

Please turn on the hypothes.is plugin, and view public annotations to see Latex math representations

model for now)

Figure 3.4: Human uncertainty (CI width) vs |LLM − Human|. Spearman correlation reported.

Table 3.3: CI coverage: does one interval contain the other’s point estimate?

coverage

Table 3.1: Agreement metrics by metric (Pearson, Spearman, mean/median bias, κ unweighted/weighted).

metric

mean_bias

metric

Calibration

calibration

As another indicator of agreement

monotonic relationship

he horizontal radius covers the union of all human evidence for that paper — combining individual raters’ point scores and their CIs

For each paper (selected metric), the ellipse is centered at the pair of midpoints (Human, LLM).

(unless a point lies outside its own stated CI, which is flagged in a diagnostic check).

robust CI ellipses

Uncertainty fields. Where available, we carry lower/upper bounds for both LLM and humans; these are used in optional uncertainty checks but do not affect the mid‑point comparisons below.

optional

we fold these into the single LLM metric global_relevance.

Human criteria are recoded to the LLM schema (e.g., claims → claims_evidence, adv_knowledge → advancing_knowledge, etc.)

Sources. LLM ratings come from results/metrics_long.csv (rendered in the previous chapter). Human ratings are imported from your hand‑coded spreadsheet and mapped to LLM paper IDs via UJ_map.csv.

compute paired scores per (paper, metric)

We (i) harmonize the two sources to a common set of metrics and paper IDs

compares LLM‑generated ratings

human evaluations across the Unjournal criteria.

Limited International Scope: The focus on US data and markets (acknowledged in footnote 20) may limit generalizability, particularly given different regulatory environments and consumer preferences globally.

Supply Chain Complexity: The analysis doesn't fully account for how restaurant vs. grocery substitution patterns might differ, despite acknowledging this limitation.

Importance:5/10

The explicit recognition that 'given the range of unknowns involved' makes direct comparison infeasible demonstrates strategic research thinking that EA organizations should emulate.

Score: 78/100

Comment 1

Temporal Dynamics: The document doesn't adequately address how substitution patterns might change as plant-based products become more mainstream or as consumer familiarity increases.

Meta-Analysis Limitations: While the document presents cross-price elasticity estimates, it could better address how traditional meta-analysis approaches might be inappropriate given the fundamental methodological concerns raised.

the document could better explain how these predictions will be weighted against expert evaluations.

This nuanced view prevents oversimplified dismissal of conflicting results.

However, the document could strengthen its analysis by exploring potential explanations for these inconsistencies more systematically.

The progression from broad goal-focused questions to specific operationalized questions follows sound research design principles.

supermarkets raising PBA prices during economic windfalls effectively illustrates how naive analysis could overstate substitution effects.

This represents sophisticated epistemic humility—recognizing that documenting limitations transparently provides significant value even when definitive answers remain elusive

Audience Analysis and Calibration

This document presents The Unjournal's approach to investigating whether plant-based meat alternatives effectively displace animal product consumption—a critical question for animal welfare funding decisions. The analysis demonstrates sophisticated awareness of methodological limitations while proposing a structured approach to synthesize conflicting evidence.

This document presents The Unjournal's Pivotal Questions initiative investigating plant-based meat substitution effects on animal product consumption. The analysis reveals significant methodological challenges in the existing research, with conflicting findings across studies using similar data sources. The document proposes a focused operationalized question about Impossible/Beyond Beef price changes and chicken consumption, while acknowledging fundamental limitations in current estimation approaches. Key insights include the recognition of endogeneity problems, the value of transparent uncertainty quantification, and the need for more rigorous experimental designs to inform animal welfare funding decisions.

acob Schmiess; we ac

Extracted forecast: Will Impossible Beef reach price-parity with conventional ground beef by December 31, 2030? Quality scores: Precision: 85 | Verifiable: 90 | Important: 75 | Robustness: 65 Our reasoning: Based on 2 independent analyses, the estimated probability is 72.3%. There is high consensus among the forecasts. Individual model forecasts: Model 1: 72.3% - "Plant-based meat costs have declined significantly due to scale economies and R&D investments, while conventional beef prices face upward pressure from environmental regulations and feed costs, making price parity likely within the 5+ year timeframe."

Based on 2 independent analyses, the estimated probability is 7.5%. There is high consensus among the forecasts.

1. "What will be the price of IB+ on January 1, 2030?"

op 3 Most Overconfident Predictions 1. "Will most cultured meat (by volume) be produced wi..." Reality check: ~32.6% | Gap: 17.4%

Figure 2.2: Per-paper mid-point scores across all metrics. Darker green → higher percentile. Columns ordered by each paper’s overall average. #altair-viz-41a3d5b933fe425c841b312f24a54cc6.vega-embed { width: 100%; display: flex; } #altair-viz-41a3d5b933fe425c841b312f24a54cc6.vega-embed details, #altair-viz-41a3d5b933fe425c841b312f24a54cc6.vega-embed details summary { position: relative; } Save as SVGSave as PNGView SourceView Compiled VegaOpen in Vega Editor var VEGA_DEBUG = (typeof VEGA_DEBUG == "undefined") ? {} : VEGA_DEBUG; (function(spec, embedOpt){ let outputDiv = document.currentScript.previousElementSibling; if (outputDiv.id !== "altair-viz-41a3d5b933fe425c841b312f24a54cc6") { outputDiv = document.getElementById("altair-viz-41a3d5b933fe425c841b312f24a54cc6"); } const paths = { "vega": "https://cdn.jsdelivr.net/npm/vega@5?noext", "vega-lib": "https://cdn.jsdelivr.net/npm/vega-lib?noext", "vega-lite": "https://cdn.jsdelivr.net/npm/vega-lite@5.20.1?noext", "vega-embed": "https://cdn.jsdelivr.net/npm/vega-embed@6?noext", }; function maybeLoadScript(lib, version) { var key = `${lib.replace("-", "")}_version`; return (VEGA_DEBUG[key] == version) ? Promise.resolve(paths[lib]) : new Promise(function(resolve, reject) { var s = document.createElement('script'); document.getElementsByTagName("head")[0].appendChild(s); s.async = true; s.onload = () => { VEGA_DEBUG[key] = version; return resolve(paths[lib]); }; s.onerror = () => reject(`Error loading script: ${paths[lib]}`); s.src = paths[lib]; }); } function showError(err) { outputDiv.innerHTML = `<div class="error" style="color:red;">${err}</div>`; throw err; } function displayChart(vegaEmbed) { vegaEmbed(outputDiv, spec, embedOpt) .catch(err => showError(`Javascript Error: ${err.message}<br>This usually means there's a typo in your chart specification. See the javascript console for the full traceback.`)); } if(typeof define === "function" && define.amd) { requirejs.config({paths}); let deps = ["vega-embed"]; require(deps, displayChart, err => showError(`Error loading script: ${err.message}`)); } else { maybeLoadScript("vega", "5") .then(() => maybeLoadScript("vega-lite", "5.20.1")) .then(() => maybeLoadScript("vega-embed", "6")) .catch(showError) .then(() => displayChart(vegaEmbed)); } })({"config": {"view": {"stroke": "transparent"}, "background": "white", "title": {"font": "Source Sans Pro", "fontSize": 16, "color": "#222"}, "axis": {"labelFont": "Source Sans Pro", "titleFont": "Source Sans Pro", "labelColor": "#222", "titleColor": "#222", "gridOpacity": 0.15}, "legend": {"labelFont": "Source Sans Pro", "titleFont": "Source Sans Pro"}, "range": {"heatmap": ["#f6fbf3", "#e2f1d7", "#cfe7ba", "#badc9c", "#a6d27f", "#92c861", "#7dbd43", "#69b325", "#55a807", "#477b13"], "category": ["#99bb66", "#f19e4b", "#6bb0f3", "#d9534f", "#636363", "#ffb400", "#53354a", "#2780e3", "#3fb618", "#8e6c8a"]}}, "data": {"name": "data-b64186ba8bd194e866931748b3fda927"}, "mark": {"type": "rect"}, "encoding": {"color": {"field": "midpoint", "legend": {"title": "Score"}, "scale": {"domain": [0, 100]}, "type": "quantitative"}, "tooltip": [{"field": "short", "type": "nominal"}, {"field": "metric", "type": "nominal"}, {"field": "midpoint", "type": "quantitative"}], "x": {"field": "short", "sort": ["Banerjee (2023)", "Acemoglu (2024)", "Buntaine (2023)", "Kremer (2022)", "Akram (2017)", "Walker (2023)", "Jack (2022)", "Alcott (2024)", "Crawfurd (2023)", "Bhat (2022)", "Liang (2021)", "Haushofer (2020)", "Epperson (2024)", "Arora (2023)", "Fangwa (2023)", "Trammel (2025)", "Alatas (2019)", "Hill (2024)", "Barker (2021)", "Aghion (2017)", "Bahar (2022)", "Carson (2023)", "Clancy (2024)", "Kubo (2023)", "Schuett (2023)", "Barberio (2022)", "Chuard (2022)", "Bettle (2023)", "Kremer (2020)", "Bhattacharya (2020)", "Bruers (2021)"], "title": null, "type": "nominal"}, "y": {"field": "metric", "sort": null, "title": null, "type": "nominal"}}, "height": 280, "width": 434, "$schema": "https://vega.github.io/schema/vega-lite/v5.20.1.json", "datasets": {"data-b64186ba8bd194e866931748b3fda927": [{"metric": "global_relevance", "paper": "Acemoglu et al. 2024", "midpoint": 92, "short": "Acemoglu (2024)"}, {"metric": "open_science", "paper": "Acemoglu et al. 2024", "midpoint": 75, "short": "Acemoglu (2024)"}, {"metric": "logic_communication", "paper": "Acemoglu et al. 2024", "midpoint": 90, "short": "Acemoglu (2024)"}, {"metric": "advancing_knowledge", "paper": "Acemoglu et al. 2024", "midpoint": 88, "short": "Acemoglu (2024)"}, {"metric": "methods", "paper": "Acemoglu et al. 2024", "midpoint": 83, "short": "Acemoglu (2024)"}, {"metric": "claims_evidence", "paper": "Acemoglu et al. 2024", "midpoint": 85, "short": "Acemoglu (2024)"}, {"metric": "overall", "paper": "Acemoglu et al. 2024", "midpoint": 86, "short": "Acemoglu (2024)"}, {"metric": "global_relevance", "paper": "Aghion et al. 2017", "midpoint": 70, "short": "Aghion (2017)"}, {"metric": "open_science", "paper": "Aghion et al. 2017", "midpoint": 40, "short": "Aghion (2017)"}, {"metric": "logic_communication", "paper": "Aghion et al. 2017", "midpoint": 85, "short": "Aghion (2017)"}, {"metric": "advancing_knowledge", "paper": "Aghion et al. 2017", "midpoint": 80, "short": "Aghion (2017)"}, {"metric": "methods", "paper": "Aghion et al. 2017", "midpoint": 72, "short": "Aghion (2017)"}, {"metric": "claims_evidence", "paper": "Aghion et al. 2017", "midpoint": 65, "short": "Aghion (2017)"}, {"metric": "overall", "paper": "Aghion et al. 2017", "midpoint": 69, "short": "Aghion (2017)"}, {"metric": "global_relevance", "paper": "Akram et al. 2017", "midpoint": 85, "short": "Akram (2017)"}, {"metric": "open_science", "paper": "Akram et al. 2017", "midpoint": 60, "short": "Akram (2017)"}, {"metric": "logic_communication", "paper": "Akram et al. 2017", "midpoint": 83, "short": "Akram (2017)"}, {"metric": "advancing_knowledge", "paper": "Akram et al. 2017", "midpoint": 80, "short": "Akram (2017)"}, {"metric": "methods", "paper": "Akram et al. 2017", "midpoint": 88, "short": "Akram (2017)"}, {"metric": "claims_evidence", "paper": "Akram et al. 2017", "midpoint": 85, "short": "Akram (2017)"}, {"metric": "overall", "paper": "Akram et al. 2017", "midpoint": 83, "short": "Akram (2017)"}, {"metric": "global_relevance", "paper": "Alatas et al. 2019", "midpoint": 70, "short": "Alatas (2019)"}, {"metric": "open_science", "paper": "Alatas et al. 2019", "midpoint": 50, "short": "Alatas (2019)"}, {"metric": "logic_communication", "paper": "Alatas et al. 2019", "midpoint": 80, "short": "Alatas (2019)"}, {"metric": "advancing_knowledge", "paper": "Alatas et al. 2019", "midpoint": 75, "short": "Alatas (2019)"}, {"metric": "methods", "paper": "Alatas et al. 2019", "midpoint": 85, "short": "Alatas (2019)"}, {"metric": "claims_evidence", "paper": "Alatas et al. 2019", "midpoint": 80, "short": "Alatas (2019)"}, {"metric": "overall", "paper": "Alatas et al. 2019", "midpoint": 75, "short": "Alatas (2019)"}, {"metric": "global_relevance", "paper": "Alcott et al. 2024", "midpoint": 65, "short": "Alcott (2024)"}, {"metric": "open_science", "paper": "Alcott et al. 2024", "midpoint": 75, "short": "Alcott (2024)"}, {"metric": "logic_communication", "paper": "Alcott et al. 2024", "midpoint": 90, "short": "Alcott (2024)"}, {"metric": "advancing_knowledge", "paper": "Alcott et al. 2024", "midpoint": 85, "short": "Alcott (2024)"}, {"metric": "methods", "paper": "Alcott et al. 2024", "midpoint": 77, "short": "Alcott (2024)"}, {"metric": "claims_evidence", "paper": "Alcott et al. 2024", "midpoint": 80, "short": "Alcott (2024)"}, {"metric": "overall", "paper": "Alcott et al. 2024", "midpoint": 80, "short": "Alcott (2024)"}, {"metric": "global_relevance", "paper": "Arora et al. 2023", "midpoint": 60, "short": "Arora (2023)"}, {"metric": "open_science", "paper": "Arora et al. 2023", "midpoint": 70, "short": "Arora (2023)"}, {"metric": "logic_communication", "paper": "Arora et al. 2023", "midpoint": 78, "short": "Arora (2023)"}, {"metric": "advancing_knowledge", "paper": "Arora et al. 2023", "midpoint": 85, "short": "Arora (2023)"}, {"metric": "methods", "paper": "Arora et al. 2023", "midpoint": 80, "short": "Arora (2023)"}, {"metric": "claims_evidence", "paper": "Arora et al. 2023", "midpoint": 75, "short": "Arora (2023)"}, {"metric": "overall", "paper": "Arora et al. 2023", "midpoint": 77, "short": "Arora (2023)"}, {"metric": "global_relevance", "paper": "Bahar et al. 2022", "midpoint": 80, "short": "Bahar (2022)"}, {"metric": "open_science", "paper": "Bahar et al. 2022", "midpoint": 40, "short": "Bahar (2022)"}, {"metric": "logic_communication", "paper": "Bahar et al. 2022", "midpoint": 70, "short": "Bahar (2022)"}, {"metric": "advancing_knowledge", "paper": "Bahar et al. 2022", "midpoint": 75, "short": "Bahar (2022)"}, {"metric": "methods", "paper": "Bahar et al. 2022", "midpoint": 72, "short": "Bahar (2022)"}, {"metric": "claims_evidence", "paper": "Bahar et al. 2022", "midpoint": 70, "short": "Bahar (2022)"}, {"metric": "overall", "paper": "Bahar et al. 2022", "midpoint": 68, "short": "Bahar (2022)"}, {"metric": "global_relevance", "paper": "Banerjee et al. 2023", "midpoint": 95, "short": "Banerjee (2023)"}, {"metric": "open_science", "paper": "Banerjee et al. 2023", "midpoint": 80, "short": "Banerjee (2023)"}, {"metric": "logic_communication", "paper": "Banerjee et al. 2023", "midpoint": 85, "short": "Banerjee (2023)"}, {"metric": "advancing_knowledge", "paper": "Banerjee et al. 2023", "midpoint": 95, "short": "Banerjee (2023)"}, {"metric": "methods", "paper": "Banerjee et al. 2023", "midpoint": 90, "short": "Banerjee (2023)"}, {"metric": "claims_evidence", "paper": "Banerjee et al. 2023", "midpoint": 88, "short": "Banerjee (2023)"}, {"metric": "overall", "paper": "Banerjee et al. 2023", "midpoint": 90, "short": "Banerjee (2023)"}, {"metric": "global_relevance", "paper": "Barberio et al. 2022", "midpoint": 85, "short": "Barberio (2022)"}, {"metric": "open_science", "paper": "Barberio et al. 2022", "midpoint": 40, "short": "Barberio (2022)"}, {"metric": "logic_communication", "paper": "Barberio et al. 2022", "midpoint": 80, "short": "Barberio (2022)"}, {"metric": "advancing_knowledge", "paper": "Barberio et al. 2022", "midpoint": 60, "short": "Barberio (2022)"}, {"metric": "methods", "paper": "Barberio et al. 2022", "midpoint": 70, "short": "Barberio (2022)"}, {"metric": "claims_evidence", "paper": "Barberio et al. 2022", "midpoint": 65, "short": "Barberio (2022)"}, {"metric": "overall", "paper": "Barberio et al. 2022", "midpoint": 67, "short": "Barberio (2022)"}, {"metric": "global_relevance", "paper": "Barker et al. 2021", "midpoint": 85, "short": "Barker (2021)"}, {"metric": "open_science", "paper": "Barker et al. 2021", "midpoint": 55, "short": "Barker (2021)"}, {"metric": "logic_communication", "paper": "Barker et al. 2021", "midpoint": 80, "short": "Barker (2021)"}, {"metric": "advancing_knowledge", "paper": "Barker et al. 2021", "midpoint": 75, "short": "Barker (2021)"}, {"metric": "methods", "paper": "Barker et al. 2021", "midpoint": 70, "short": "Barker (2021)"}, {"metric": "claims_evidence", "paper": "Barker et al. 2021", "midpoint": 65, "short": "Barker (2021)"}, {"metric": "overall", "paper": "Barker et al. 2021", "midpoint": 72, "short": "Barker (2021)"}, {"metric": "global_relevance", "paper": "Bettle 2023", "midpoint": 75, "short": "Bettle (2023)"}, {"metric": "open_science", "paper": "Bettle 2023", "midpoint": 40, "short": "Bettle (2023)"}, {"metric": "logic_communication", "paper": "Bettle 2023", "midpoint": 70, "short": "Bettle (2023)"}, {"metric": "advancing_knowledge", "paper": "Bettle 2023", "midpoint": 65, "short": "Bettle (2023)"}, {"metric": "methods", "paper": "Bettle 2023", "midpoint": 60, "short": "Bettle (2023)"}, {"metric": "claims_evidence", "paper": "Bettle 2023", "midpoint": 55, "short": "Bettle (2023)"}, {"metric": "overall", "paper": "Bettle 2023", "midpoint": 61, "short": "Bettle (2023)"}, {"metric": "global_relevance", "paper": "Bhat et al. 2022", "midpoint": 85, "short": "Bhat (2022)"}, {"metric": "open_science", "paper": "Bhat et al. 2022", "midpoint": 60, "short": "Bhat (2022)"}, {"metric": "logic_communication", "paper": "Bhat et al. 2022", "midpoint": 80, "short": "Bhat (2022)"}, {"metric": "advancing_knowledge", "paper": "Bhat et al. 2022", "midpoint": 85, "short": "Bhat (2022)"}, {"metric": "methods", "paper": "Bhat et al. 2022", "midpoint": 80, "short": "Bhat (2022)"}, {"metric": "claims_evidence", "paper": "Bhat et al. 2022", "midpoint": 75, "short": "Bhat (2022)"}, {"metric": "overall", "paper": "Bhat et al. 2022", "midpoint": 78, "short": "Bhat (2022)"}, {"metric": "global_relevance", "paper": "Bhattacharya and Packalen 2020", "midpoint": 55, "short": "Bhattacharya (2020)"}, {"metric": "open_science", "paper": "Bhattacharya and Packalen 2020", "midpoint": 40, "short": "Bhattacharya (2020)"}, {"metric": "logic_communication", "paper": "Bhattacharya and Packalen 2020", "midpoint": 70, "short": "Bhattacharya (2020)"}, {"metric": "advancing_knowledge", "paper": "Bhattacharya and Packalen 2020", "midpoint": 60, "short": "Bhattacharya (2020)"}, {"metric": "methods", "paper": "Bhattacharya and Packalen 2020", "midpoint": 45, "short": "Bhattacharya (2020)"}, {"metric": "claims_evidence", "paper": "Bhattacharya and Packalen 2020", "midpoint": 55, "short": "Bhattacharya (2020)"}, {"metric": "overall", "paper": "Bhattacharya and Packalen 2020", "midpoint": 54, "short": "Bhattacharya (2020)"}, {"metric": "global_relevance", "paper": "Bruers 2021", "midpoint": 55, "short": "Bruers (2021)"}, {"metric": "open_science", "paper": "Bruers 2021", "midpoint": 25, "short": "Bruers (2021)"}, {"metric": "logic_communication", "paper": "Bruers 2021", "midpoint": 60, "short": "Bruers (2021)"}, {"metric": "advancing_knowledge", "paper": "Bruers 2021", "midpoint": 50, "short": "Bruers (2021)"}, {"metric": "methods", "paper": "Bruers 2021", "midpoint": 35, "short": "Bruers (2021)"}, {"metric": "claims_evidence", "paper": "Bruers 2021", "midpoint": 40, "short": "Bruers (2021)"}, {"metric": "overall", "paper": "Bruers 2021", "midpoint": 45, "short": "Bruers (2021)"}, {"metric": "global_relevance", "paper": "Buntaine et al. 2023", "midpoint": 90, "short": "Buntaine (2023)"}, {"metric": "open_science", "paper": "Buntaine et al. 2023", "midpoint": 70, "short": "Buntaine (2023)"}, {"metric": "logic_communication", "paper": "Buntaine et al. 2023", "midpoint": 85, "short": "Buntaine (2023)"}, {"metric": "advancing_knowledge", "paper": "Buntaine et al. 2023", "midpoint": 85, "short": "Buntaine (2023)"}, {"metric": "methods", "paper": "Buntaine et al. 2023", "midpoint": 90, "short": "Buntaine (2023)"}, {"metric": "claims_evidence", "paper": "Buntaine et al. 2023", "midpoint": 90, "short": "Buntaine (2023)"}, {"metric": "overall", "paper": "Buntaine et al. 2023", "midpoint": 85, "short": "Buntaine (2023)"}, {"metric": "global_relevance", "paper": "Carson et al. 2023", "midpoint": 60, "short": "Carson (2023)"}, {"metric": "open_science", "paper": "Carson et al. 2023", "midpoint": 55, "short": "Carson (2023)"}, {"metric": "logic_communication", "paper": "Carson et al. 2023", "midpoint": 80, "short": "Carson (2023)"}, {"metric": "advancing_knowledge", "paper": "Carson et al. 2023", "midpoint": 65, "short": "Carson (2023)"}, {"metric": "methods", "paper": "Carson et al. 2023", "midpoint": 70, "short": "Carson (2023)"}, {"metric": "claims_evidence", "paper": "Carson et al. 2023", "midpoint": 75, "short": "Carson (2023)"}, {"metric": "overall", "paper": "Carson et al. 2023", "midpoint": 68, "short": "Carson (2023)"}, {"metric": "global_relevance", "paper": "Chuard et al. 2022", "midpoint": 65, "short": "Chuard (2022)"}, {"metric": "open_science", "paper": "Chuard et al. 2022", "midpoint": 50, "short": "Chuard (2022)"}, {"metric": "logic_communication", "paper": "Chuard et al. 2022", "midpoint": 75, "short": "Chuard (2022)"}, {"metric": "advancing_knowledge", "paper": "Chuard et al. 2022", "midpoint": 60, "short": "Chuard (2022)"}, {"metric": "methods", "paper": "Chuard et al. 2022", "midpoint": 65, "short": "Chuard (2022)"}, {"metric": "claims_evidence", "paper": "Chuard et al. 2022", "midpoint": 70, "short": "Chuard (2022)"}, {"metric": "overall", "paper": "Chuard et al. 2022", "midpoint": 64, "short": "Chuard (2022)"}, {"metric": "global_relevance", "paper": "Clancy 2024", "midpoint": 80, "short": "Clancy (2024)"}, {"metric": "open_science", "paper": "Clancy 2024", "midpoint": 55, "short": "Clancy (2024)"}, {"metric": "logic_communication", "paper": "Clancy 2024", "midpoint": 70, "short": "Clancy (2024)"}, {"metric": "advancing_knowledge", "paper": "Clancy 2024", "midpoint": 75, "short": "Clancy (2024)"}, {"metric": "methods", "paper": "Clancy 2024", "midpoint": 60, "short": "Clancy (2024)"}, {"metric": "claims_evidence", "paper": "Clancy 2024", "midpoint": 65, "short": "Clancy (2024)"}, {"metric": "overall", "paper": "Clancy 2024", "midpoint": 68, "short": "Clancy (2024)"}, {"metric": "global_relevance", "paper": "Crawfurd et al. 2023", "midpoint": 90, "short": "Crawfurd (2023)"}, {"metric": "open_science", "paper": "Crawfurd et al. 2023", "midpoint": 70, "short": "Crawfurd (2023)"}, {"metric": "logic_communication", "paper": "Crawfurd et al. 2023", "midpoint": 85, "short": "Crawfurd (2023)"}, {"metric": "advancing_knowledge", "paper": "Crawfurd et al. 2023", "midpoint": 80, "short": "Crawfurd (2023)"}, {"metric": "methods", "paper": "Crawfurd et al. 2023", "midpoint": 75, "short": "Crawfurd (2023)"}, {"metric": "claims_evidence", "paper": "Crawfurd et al. 2023", "midpoint": 70, "short": "Crawfurd (2023)"}, {"metric": "overall", "paper": "Crawfurd et al. 2023", "midpoint": 78, "short": "Crawfurd (2023)"}, {"metric": "global_relevance", "paper": "Epperson and Gerster 2024", "midpoint": 60, "short": "Epperson (2024)"}, {"metric": "open_science", "paper": "Epperson and Gerster 2024", "midpoint": 85, "short": "Epperson (2024)"}, {"metric": "logic_communication", "paper": "Epperson and Gerster 2024", "midpoint": 80, "short": "Epperson (2024)"}, {"metric": "advancing_knowledge", "paper": "Epperson and Gerster 2024", "midpoint": 70, "short": "Epperson (2024)"}, {"metric": "methods", "paper": "Epperson and Gerster 2024", "midpoint": 78, "short": "Epperson (2024)"}, {"metric": "claims_evidence", "paper": "Epperson and Gerster 2024", "midpoint": 80, "short": "Epperson (2024)"}, {"metric": "overall", "paper": "Epperson and Gerster 2024", "midpoint": 75, "short": "Epperson (2024)"}, {"metric": "global_relevance", "paper": "Fangwa et al. 2023", "midpoint": 85, "short": "Fangwa (2023)"}, {"metric": "open_science", "paper": "Fangwa et al. 2023", "midpoint": 45, "short": "Fangwa (2023)"}, {"metric": "logic_communication", "paper": "Fangwa et al. 2023", "midpoint": 80, "short": "Fangwa (2023)"}, {"metric": "advancing_knowledge", "paper": "Fangwa et al. 2023", "midpoint": 75, "short": "Fangwa (2023)"}, {"metric": "methods", "paper": "Fangwa et al. 2023", "midpoint": 80, "short": "Fangwa (2023)"}, {"metric": "claims_evidence", "paper": "Fangwa et al. 2023", "midpoint": 78, "short": "Fangwa (2023)"}, {"metric": "overall", "paper": "Fangwa et al. 2023", "midpoint": 75, "short": "Fangwa (2023)"}, {"metric": "global_relevance", "paper": "Haushofer et al. 2020", "midpoint": 75, "short": "Haushofer (2020)"}, {"metric": "open_science", "paper": "Haushofer et al. 2020", "midpoint": 65, "short": "Haushofer (2020)"}, {"metric": "logic_communication", "paper": "Haushofer et al. 2020", "midpoint": 80, "short": "Haushofer (2020)"}, {"metric": "advancing_knowledge", "paper": "Haushofer et al. 2020", "midpoint": 70, "short": "Haushofer (2020)"}, {"metric": "methods", "paper": "Haushofer et al. 2020", "midpoint": 85, "short": "Haushofer (2020)"}, {"metric": "claims_evidence", "paper": "Haushofer et al. 2020", "midpoint": 80, "short": "Haushofer (2020)"}, {"metric": "overall", "paper": "Haushofer et al. 2020", "midpoint": 76, "short": "Haushofer (2020)"}, {"metric": "global_relevance", "paper": "Hill et al. 2024", "midpoint": 65, "short": "Hill (2024)"}, {"metric": "open_science", "paper": "Hill et al. 2024", "midpoint": 60, "short": "Hill (2024)"}, {"metric": "logic_communication", "paper": "Hill et al. 2024", "midpoint": 85, "short": "Hill (2024)"}, {"metric": "advancing_knowledge", "paper": "Hill et al. 2024", "midpoint": 70, "short": "Hill (2024)"}, {"metric": "methods", "paper": "Hill et al. 2024", "midpoint": 78, "short": "Hill (2024)"}, {"metric": "claims_evidence", "paper": "Hill et al. 2024", "midpoint": 75, "short": "Hill (2024)"}, {"metric": "overall", "paper": "Hill et al. 2024", "midpoint": 73, "short": "Hill (2024)"}, {"metric": "global_relevance", "paper": "Jack et al. 2022", "midpoint": 88, "short": "Jack (2022)"}, {"metric": "open_science", "paper": "Jack et al. 2022", "midpoint": 60, "short": "Jack (2022)"}, {"metric": "logic_communication", "paper": "Jack et al. 2022", "midpoint": 85, "short": "Jack (2022)"}, {"metric": "advancing_knowledge", "paper": "Jack et al. 2022", "midpoint": 78, "short": "Jack (2022)"}, {"metric": "methods", "paper": "Jack et al. 2022", "midpoint": 80, "short": "Jack (2022)"}, {"metric": "claims_evidence", "paper": "Jack et al. 2022", "midpoint": 85, "short": "Jack (2022)"}, {"metric": "overall", "paper": "Jack et al. 2022", "midpoint": 79, "short": "Jack (2022)"}, {"metric": "global_relevance", "paper": "Kremer et al. 2020", "midpoint": 85, "short": "Kremer (2020)"}, {"metric": "open_science", "paper": "Kremer et al. 2020", "midpoint": 40, "short": "Kremer (2020)"}, {"metric": "logic_communication", "paper": "Kremer et al. 2020", "midpoint": 80, "short": "Kremer (2020)"}, {"metric": "advancing_knowledge", "paper": "Kremer et al. 2020", "midpoint": 55, "short": "Kremer (2020)"}, {"metric": "methods", "paper": "Kremer et al. 2020", "midpoint": 45, "short": "Kremer (2020)"}, {"metric": "claims_evidence", "paper": "Kremer et al. 2020", "midpoint": 60, "short": "Kremer (2020)"}, {"metric": "overall", "paper": "Kremer et al. 2020", "midpoint": 61, "short": "Kremer (2020)"}, {"metric": "global_relevance", "paper": "Kremer et al. 2022", "midpoint": 95, "short": "Kremer (2022)"}, {"metric": "open_science", "paper": "Kremer et al. 2022", "midpoint": 70, "short": "Kremer (2022)"}, {"metric": "logic_communication", "paper": "Kremer et al. 2022", "midpoint": 85, "short": "Kremer (2022)"}, {"metric": "advancing_knowledge", "paper": "Kremer et al. 2022", "midpoint": 90, "short": "Kremer (2022)"}, {"metric": "methods", "paper": "Kremer et al. 2022", "midpoint": 85, "short": "Kremer (2022)"}, {"metric": "claims_evidence", "paper": "Kremer et al. 2022", "midpoint": 80, "short": "Kremer (2022)"}, {"metric": "overall", "paper": "Kremer et al. 2022", "midpoint": 84, "short": "Kremer (2022)"}, {"metric": "global_relevance", "paper": "Kubo et al. 2023", "midpoint": 60, "short": "Kubo (2023)"}, {"metric": "open_science", "paper": "Kubo et al. 2023", "midpoint": 55, "short": "Kubo (2023)"}, {"metric": "logic_communication", "paper": "Kubo et al. 2023", "midpoint": 78, "short": "Kubo (2023)"}, {"metric": "advancing_knowledge", "paper": "Kubo et al. 2023", "midpoint": 65, "short": "Kubo (2023)"}, {"metric": "methods", "paper": "Kubo et al. 2023", "midpoint": 75, "short": "Kubo (2023)"}, {"metric": "claims_evidence", "paper": "Kubo et al. 2023", "midpoint": 70, "short": "Kubo (2023)"}, {"metric": "overall", "paper": "Kubo et al. 2023", "midpoint": 67, "short": "Kubo (2023)"}, {"metric": "global_relevance", "paper": "Liang et al. 2021", "midpoint": 70, "short": "Liang (2021)"}, {"metric": "open_science", "paper": "Liang et al. 2021", "midpoint": 70, "short": "Liang (2021)"}, {"metric": "logic_communication", "paper": "Liang et al. 2021", "midpoint": 75, "short": "Liang (2021)"}, {"metric": "advancing_knowledge", "paper": "Liang et al. 2021", "midpoint": 80, "short": "Liang (2021)"}, {"metric": "methods", "paper": "Liang et al. 2021", "midpoint": 85, "short": "Liang (2021)"}, {"metric": "claims_evidence", "paper": "Liang et al. 2021", "midpoint": 85, "short": "Liang (2021)"}, {"metric": "overall", "paper": "Liang et al. 2021", "midpoint": 78, "short": "Liang (2021)"}, {"metric": "global_relevance", "paper": "Schuett et al. 2023", "midpoint": 70, "short": "Schuett (2023)"}, {"metric": "open_science", "paper": "Schuett et al. 2023", "midpoint": 80, "short": "Schuett (2023)"}, {"metric": "logic_communication", "paper": "Schuett et al. 2023", "midpoint": 70, "short": "Schuett (2023)"}, {"metric": "advancing_knowledge", "paper": "Schuett et al. 2023", "midpoint": 60, "short": "Schuett (2023)"}, {"metric": "methods", "paper": "Schuett et al. 2023", "midpoint": 55, "short": "Schuett (2023)"}, {"metric": "claims_evidence", "paper": "Schuett et al. 2023", "midpoint": 65, "short": "Schuett (2023)"}, {"metric": "overall", "paper": "Schuett et al. 2023", "midpoint": 67, "short": "Schuett (2023)"}, {"metric": "global_relevance", "paper": "Trammel and Aschenbrenner 2025", "midpoint": 90, "short": "Trammel (2025)"}, {"metric": "open_science", "paper": "Trammel and Aschenbrenner 2025", "midpoint": 40, "short": "Trammel (2025)"}, {"metric": "logic_communication", "paper": "Trammel and Aschenbrenner 2025", "midpoint": 80, "short": "Trammel (2025)"}, {"metric": "advancing_knowledge", "paper": "Trammel and Aschenbrenner 2025", "midpoint": 85, "short": "Trammel (2025)"}, {"metric": "methods", "paper": "Trammel and Aschenbrenner 2025", "midpoint": 75, "short": "Trammel (2025)"}, {"metric": "claims_evidence", "paper": "Trammel and Aschenbrenner 2025", "midpoint": 70, "short": "Trammel (2025)"}, {"metric": "overall", "paper": "Trammel and Aschenbrenner 2025", "midpoint": 75, "short": "Trammel (2025)"}, {"metric": "global_relevance", "paper": "Walker et al. 2023", "midpoint": 90, "short": "Walker (2023)"}, {"metric": "open_science", "paper": "Walker et al. 2023", "midpoint": 60, "short": "Walker (2023)"}, {"metric": "logic_communication", "paper": "Walker et al. 2023", "midpoint": 80, "short": "Walker (2023)"}, {"metric": "advancing_knowledge", "paper": "Walker et al. 2023", "midpoint": 85, "short": "Walker (2023)"}, {"metric": "methods", "paper": "Walker et al. 2023", "midpoint": 80, "short": "Walker (2023)"}, {"metric": "claims_evidence", "paper": "Walker et al. 2023", "midpoint": 85, "short": "Walker (2023)"}, {"metric": "overall", "paper": "Walker et al. 2023", "midpoint": 80, "short": "Walker (2023)"}]}}, {"mode": "vega-lite"});

in-memory Chroma vector DB

(one per page).

International Conference on Learning Representations (ICLR) machine learning conference

5.1 (How) can diagnostic tests make sense? Where is the burden of proof?

o3,

including probability judgement and robustness suggestions.

factual

→ scoring ssrn-4071953.pdf …

# Install the SDK if it isn’t present

key_path = pathlib.Path("key/openai_key.txt")

Project files

Table 2.1: Overall assessments (percentiles)

# 1. Point to the papers you want to score

Following (Garg and Fetzer 2025)

::: {#cell-Averaged ratings .cell tbl-cap=‘Mean ± SD of ten model draws’ execution_count=8}

paper_id

Explore LangChain, OpenAI Evals