Quick start: four questions ? After the dangerous period (the first century), wh

log10

Current basis. Costs are per kg of wet cell biomass, because the code treats the cell-density inputs as grams of wet biomass per litre at harvest. Harvested cells are mostly water, and a wet kg can carry more or less residual medium depending on washing and dewatering: a paste at 15% solids carries about half the protein of one at 30%. So “cost per kg of wet cell biomass” is not a stable unit.

wet-biomass water content is not standardized

the single latent-maturity factor is ad hoc;

Output basis: what “per kg” means here

The May workshop form asked about costs “across all large-scale plants in the world,” without an explicit condition that those plants exist. A June 25 revision added that condition; the later independent respondents received that wording. The new questions separate these interpretations. Earlier answers have not been reclassified as A1 or A2. Wording history and updated questions. Cross-cohort pooling is an exploratory comparison, not a set of answers to an identically worded question.

Privacy and provenance — human-reviewed 31 July 2026.

equation

The function

which the horizon charts plot.

Engine: survival(r), expectedValue(r, v).

Budget ($M)Voting MethodCredence-WeightedMy Favorite TheoryMECMETNash BargainingMSABorda CountSplit CycleLexicographic Maximin

We encourage you to blend your results with RP's preset views, or retake the quiz and combine multiple runs. Global Health & Development20%

I want to spend on whatever pays off most on average, even if the chance of success is small and the work may do harmEvaluate funds using a standard expected value. This is risk-neutral, so when calculating a fund's score, any negative effects can be offset by positive effects of equal probability and scale. A fund that has only a 1% probability of success still looks good if the effect is large enough.75%

If AI poses serious existential risks, the benefits of other interventions could be cut short so their value could reduce. Use the slider to apply a discount: 0% means no adjustment, 100% means only AI interventions are valued.

For substantive or longer-form discussion — please post on 💬 GitHub Discussions. That’s where the conversation can get involved, where others can reply and build on each other, and where everything stays threaded and organized. See the 📖 Discussion Map for where to post what, or jump to a hub:

P(Fed-batch) (?)

Probability of Each Process Mode (?) Code viewof p_fedbatch = Inputs.range([0, 1], { value: urlNum("p_fedbatch", 0.20), step: 0.05, label: html`P(Fed-batch) <abbr style="cursor:help;text-decoration:underline dotted;font-size:0.85em;color:#888;" title="Low density (5–30 g/L), moderate media use (1–2×). Nutrient-concentrated feeds added periodically. Simpler to operate but lower cell density than perfusion.">(?)</abbr>` }) viewof p_perfusion = Inputs.range([0, 1], { value: urlNum("p_perfusion", 0.50), step: 0.05, label: html`P(Perfusion) <abbr style="cursor:help;text-decoration:underline dotted;font-size:0.85em;color:#888;" title="Medium-high density (30–150 g/L), higher media throughput (1–5×). Continuous media exchange with cell retention. Currently the industry standard for high-density CM production.">(?)</abbr>` }) viewof p_continuous = Inputs.range([0, 1], { value: urlNum("p_continuous", 0.30), step: 0.05, label: html`P(Continuous) <abbr style="cursor:help;text-decoration:underline dotted;font-size:0.85em;color:#888;" title="Highest density (50–200 g/L), efficient media use (0.5–3×). Near-steady-state operation; cells grown and harvested continuously with optimized recycling.">(?)</abbr>` })

PILOT / DRAFT — proposal revised July 20, 2026. This is a work-in-progress proposal for a cultivated-meat modeling hack; no event date has been announced. Numbers, scope, and the in-browser simulator are illustrative and will change. Feedback welcome via the propose-a-change form below or by annotating this page with Hypothes.is.

Evidence updates since the May workshop GFI's 2026 industry report records both progress and headwinds: 20,000-L production was reported in 2025, while disclosed sector funding fell to $73.9m from $144m in 2024. Scale and finance should be separate variables, not one maturity score. The FDA's February 2026 inventory lists five completed cultured-cell food consultations, including three completed in 2025. Regulatory progress is real, but it does not validate cost or throughput. A 2025 porcine-cell study found selected plant hydrolysates reduced serum use by up to 70%, while Pasitka's chicken process used $0.63/L animal-component-free medium. These are useful feasibility points, not evidence that full substitution works across cell lines and large reactors. Pablo AMC's July response reframes a central disagreement: a frontier “reasonably achievable” plant is not the same estimand as the average qualifying commercial plant asked about in CM_01. The hack should report both, not blend them.

Cost → demand bridge: shows what a properly harmonized value-chain cost could feed into, while flagging that the current bridge is only illustrative.

We use the public ratings export: per-paper evaluator labels ("Evaluator 1", "Evaluator 2" within each paper), so no one is identifiable and no one can be followed across papers

What this page is. We commissioned an external methodological review of our cost model and belief elicitation form. The reviewer identified several real structural problems. We are sharing the critique openly — and our responses — because we think transparency about limitations is as important as the model itself. (This page reviews the cost model and elicitation design; the subject-matter evaluation of the underlying forecast is Evaluation 2.)

Note on the interaction with cell density: Media cost per liter (/L)andcelldensity(g/L)areconnected—theyarenotparametersyoucanoptimiseindependently.Richermediaformulationsareoftenneededtosustainhighercelldensities,becausecellsindenserculturesconsumeproportionallymoreaminoacids,glucose,andothernutrients.Theeconomicallymeaningfulunitistherefore∗∗mediacostperkgofcellsproduced∗∗(/kg), which reflects both $/L and th

concrete: the public beliefs analysis

Could the action backfire?

scenario ?

0.0001) last? ?5 years, starting now (report)11 years, from year 4 (RP 2026)50 years, starting now (report)Could the action backfire? ?No (report)Its effect is uncertain and may raise riskThe risk it averts comes back later The answers move the detailed settings below; the "Where the settings come from" line on the right tags each value. Presets: Report default · Report, cubic value · Report figures (f = 0.5) · RP 2026 central values ? Scenario viewof riskType = Inputs.select(["top", "constant", "filters", "decay"], { value: U.risk, label: lab("Risk scenario", "How the yearly chance of extinction changes over time. The report's main case is Time of Perils: high risk for a century, then low risk forever. 'What the risk scenarios are' below has the details."), format: d => ({top: "Time of Perils (report section 3)", constant: "Constant risk (Ord's base case)", filters: "Great Filters (report section 3)", decay: "Exponential decay (report section 4)"})[d] }) viewof valueType = Inputs.select(M.PAGE_VALUE_TYPES, { value: U.value, label: lab("Value growth", "How much each future year is worth compared with a year of today's world. With growth, far-future years count for more, so averting risk now is worth more. Options marked 'beyond' aren't in the report."), format: d => ({constant: "V0 Constant (report)", linear: "V1 Linear (report)", quadratic: "V2 Quadratic (report)", cubic: "V3 Cubic (report)", logistic: "V4 Logistic (report)", polynomial: "Polynomial of order n (beyond)", exponential: "Exponential, uncapped (beyond)", capped: "Exponential with hard cap (beyond)", fluctuating: "Fluctuating around a base path (beyond)"})[d] }) Risk scenario ? Time of Perils (report section 3)Constant risk (Ord's base case)Great Filters (report section 3)Exponential decay (report section 4) riskType = "top" Value growth ? V0 Constant (report)V1 Linear (report)V2 Quadratic (report)V3 Cubic (report)V4 Logistic (report)Polynomial of order n (beyond)Exponential, uncapped (beyond)Exponential with hard cap (beyond)Fluctuating around a base path (beyond) valueType = "constant" What the risk scenarios are Constant risk: rt=rhigh forever. Ord’s base case; with constant value, E(w)=vc(1−rc)/rc. Time of Perils: rhigh for the first L years (report: 100), then rlow forever. Great Filters: F high-risk eras of length ℓd separated by low-risk eras of length ℓs, then low risk forever. Report: F=2, ℓ=(100,500,100,…). Exponential decay: rt=r0e−λ(t−1)+r∞. Notebook defaults: λ=10−4, r∞=10−9. The report’s Figure 3 draws a much faster decay than the one used in its tables (see Departures). Per-year risks come from per-century probabilities: 20% per century gives rhigh=1−0.81/100=0.0022289; 1% per century gives rlow≈0.0001. Risk viewof rHighCentury = rangeIn("rh", lab("High risk (% per century)", "Chance of extinction per century during high-risk periods (or always, under constant risk). The report uses 20%, a common estimate for this century. The page converts it to a yearly rate.")) High risk (% per century) ? rHighCentury = 36 viewof rLowCentury = rangeIn("rl", lab("Low risk (% per century)", "Chance of extinction per century once the dangerous period is over. The report uses about 1% (a yearly risk of 0.0001). This number largely sets how long the future is expected to last, so it drives most results.")) Low risk (% per century

High risk

become? ?

persistence P and the horizon T. Only

Its value is E(M)=E(w′)−E(w), the change in the survival-weighted sum of future value.

Each

The purple box, E(M), is the answer, measured in years of today’s world

What this is and isn’t. A research model for seeing how the answer depends on the assumptions. It isn’t a forecast, an estimate of what any real intervention achieves, or a recommendation about where to give. Like the report, it counts extinction only; catastrophes that permanently curtail humanity’s potential without ending it, and partial catastrophes, are left out (the report’s footnote 1, p. 7). Under the results, each setting in use is tagged by where its value comes from: the 2023 report, Rethink Priorities’ 2026 model, an illustrative choice of ours, or you.

Page style:PlainLensTesseraeStudioWe're trying a few looks for this site. Your choice is saved in this browser only. Tell us which you prefer.

Three evaluators assessed this review of the discount rate GiveWell applies in its cost-effectiveness analyses. Two gave quantitative ratings, shown here with their 90% credible intervals.

Recently evaluated

So the concern looks right

back-dating or dropping those predictors barely changes accuracy.

suspected cause was leakage:

Independent computational reproduction; s

hat is an impression to check, rather than an established absence of scrutiny.

overlaps it in what way? I thought that was more about forecswters, and this seems more like a theory model?

80Original AI lens before the Opus shift (not used for the synthesis)

Every way of assessing research encodes someone's goals. In this workshop you choose whose. Each table takes on a persona, from a patient community to a government funder, with $1 billion to distribute. Your group sets its goal and agrees what it wants to know about researchers and their work. Then everyone goes into the research graph: find the people and research that match your persona's interests, decide how much each kind of evidence should count, and name the data the graph can't yet provide. Each persona leaves with its own open, inspectable assessment algorithm, ready for the funding workshop that follows.

Modular science breaks research into separately attributable, reusable components, so every question, claim, piece of evidence, and method can be credited, cited, and built on by itse

AT Protocol, the open protocol behind the Bluesky social network, stores your data in a repository you own; apps read and write it there instead of locking it away in their own silos. A growing ecosystem of research apps already builds on the protocol. You'll claim a researcher ID with Aster, then pick a station and start building. Stand up a personal research page with Sifa, turn your Zotero library into a living literature review with Semble, stream live data with Matadisco or Nebra, connect AI workflows to protocol data, or build Bluesky feeds and starter packs around your research topic. Or bring your own use case instead and work through it with the room's experts. We'll close with quick demos of what everyone built.

Together we build a data network owned by the people in the room: IPFS nodes on ten Raspberry Pis and participants' laptops, joined as equal peers with no cloud server anywhere in the sy

prioritization dashboardThis is a topic-focused working collection within the shared research-prioritization project. Each brief links to its dashboard record for scores, source material, and ratings. The dashboard contains a broader AI collection; these briefs are candidate assessments, not completed Unjournal evaluations or commissioning decisions..

What it says

Transcript

hat are they selling?

Top result

Who was closest?

Robot judgeAI judgeor just tap a name

Funniest guess (bonus, optional)

Cloud software that lets companies store, combine and analyse large amounts of data, charged by how much they use.

technology professionals help each other.It later moved to experts-exchange.com, with a hyphen. Read it again without one.See it for yourselfDraft: not yet re-checked online

Davidelsbeth.geldhof@gmail.coFunniest guess (bonus, optional)Davidelsbeth.geldhof@gmail.co

Who was closest?Robot judgeAI judgeor just tap a name

Give a clueTime!

Search it yourselfChecked 2026-10-06

Wikipedia's article on puppy chow

The inventor and his sister call it 'Tarzan' swinging, and a jungle yell is optional. A 2003 re-examination cancelled all four claims.Read the patentChecked 2026-10-06 · US 6,368,227

The inventor and his sister call it 'Tarzan' swinging, and a jungle yell is optional. A 2003 re-examination cancelled all four claims.

Clue 1: It's a cute reference site for people who build websites.

Marmite

Should each group maintain a short, evolving map of its highest-value pivotal questions and most promising impactful research directions? A proposal for discussion—at and beyond the meeting. The examples are starting points, not settled views.

follow that label to the internal prioritization hub (sign-in required) for an independent follow-up assessment. The label marks a workflow step, not an endorsement.

Needs further prioritization assessment

team

ind research worth evaluating: choose a topic, read the evidence, and add your judgment.

judgment

source title

Cause: Catastrophic risks & forecasting

Get involved — Suggest research for this list ↓ | See how this tool will work (planned workflow) ↓ | Would you sponsor an evaluation? (idea, feedback wanted

They are provisional, not ratings of research quality.

Scoring methodology

hey sometimes disobey instructions, cheat on tasks,misrepresent their work, and are unable to com-plete some tasks at all, necessitating human inter-vention [23–25].

Preliminary evidence suggests that a software-driven intelligence explosion is possible. If onedoes happen, it could be the most consequen-tial technological development in history: AI sys-tems could rapidly eclipse human experts acrossmost domains and radically accelerate technolog-ical progress.

should urgently obtain more visibility into the automation of AI R&D,develop ways to steer and constrain an intelligence explosion, and prepare society to adapt to anintelligence explosion’s impacts

Preliminary evidence suggests that it coul

could AI progress radically accelerate in an “intelligence explosion,” where years of advances arecompressed into months or less?

The Economics of Recursive Self-Improvement

Because it is already receiving media and policy attention, the timing value for independent evaluation is especially high.

No separate written rationale is stored for it.

The main concern is that this may be more of a synthesis and agenda-setting policy paper than an evaluable quantitative social-science contribution,

Governments should urgently

Heavy early media attention

may

Source record Publication-stage and date evidence should be checked in the linked paper record. TARGETED_CURATED

Unjournal public-attention search

The welfare case is strongest under assumptions that future sentient welfare counts substantially, AI R&D automation could materially accelerate dangerous capabilities, and governance preparation can reduce risk without causing major counterproductive acceleration or panic.

An evaluation would focus on how well the empirical premises (the share of AI R&D already automated, the feedback-loop claims) support the policy asks, and on what is claimed versus assumed about speed.

whether and how states should require reporting on R&D automation.

AI R&D could compress years of progress into months and asking governments for visibility, rate limits and emergency plans.

would

Alan Chan, Christoph Winter, Andrew Barto, Jakub Pachocki, Geoffrey Hinton, Eric Horvitz, Yoshua Bengio, Dawn Song, Jack Clark, Hilary Greaves, Anton Korinek, Sam Manning, et al. · Catastrophic risks, the long-term future, forecasting

Unjournal tools The Unjournal's own tools live on their own page.

Econometrics, statistics and data science notes Working notes with a micro, behavioral and experimental focus. Web book Researching and writing for economics students A practical guide to economics dissertations and research projects. Course notes Microeconomics (MSc) Notes for an MSc microeconomics module at Exeter.

Increasing effective giving The puzzle of why people don't give more effectively, what the evidence says, and what we still need to learn. Web book Impact of impact information on giving Field experiments and synthesis on whether telling donors about effectiveness changes what they give. Reports EA Market Testing Public reports from a team running marketing and messaging trials with effective-giving organizations. Data analysis

Games and everyday tools

AI Foundations for Research Judgment How language models work, for researchers who have to judge their output: tokens, measurement, evidence workflows, compute economics and safety.

and site GTD Trio Guitar, tuba and drums jazz trio in Western Massachusetts. Band site K-House Jazz A flexible live jazz group in Exeter, Devon.

Play and study 18 versions of the 12-bar blues in any key, from the basic three-chord form to Charlie Parker's 'Bird' changes.

plus the bands I play tuba and trumpet with

AI-generatedThis is a model-generated prioritization rationale, not source text or a human assessment. Check factual claims against the source abstract and linked paper. (

Stored

(not AI-generated)

source title

scale.

Small score differences are not decisive.

Sensitivity Analysis: Dollar-Swing Metric

Workshop medians span roughly $1 to $100/kg

You are not asked to become a modeler; you are asked to make your expertise modelable. The

visible CM_01 point estimates

Model and planning decision. Retain the numerical forecasts and commitment ledger. Compare scenarios with faster grantmaker expansion and different cause allocations; do not add this ecosystem total to the AI-linked ledger. For an organization considering expansion, the useful next evidence is an eligible funding program, a decision timetable, and a grant it can actually budget against.

Evaluation capacity deserves attention. The overview counts only nine publicly accessible evaluators with at least three full-time staff in evaluation or grantmaking. This supports investigating capacity constraints, but headcount alone cannot calibrate our dollar cap.

Roughly 30% went to global catastrophic risk reduction, versus 20% excluding those two organizations.

evaluator-recommended destination

10 July 2026 revision: OpenAI GPT-5.6 Sol reviewed the Fable branch following the authors’ email discussion. Its additions and edits remain AI-generated and unreviewed by the authors. The revision preserves the earlier source text in Git history, corrects several overstatements, and adds a comparison with Epperson, Diederich, and Goeschl’s unit-donation experiment.

This is the “Fable-adapted” branch of the ICRC paper: consolidated, reorganized, and substantially extended by Claude (Fable 5) in July 2026, under David Reinstein’s direction. Every section carries a provenance tag:

Impact Information

Jan Schmitz University of Münster

Our closest comparison keeps the monetary ask fixed at CHF 150:

that a well-designed unit-donation scheme can increase giving, especially with a large unit size. Their full treatment does more than report a unit cost: it reframes the decision as choosing physical units and can restrict choices to a unit grid

It does not show that private philanthropy can replace government aid, or identify how to build political support for official development assistance

“DONATE 50CHF TODAY: your donation can supply 3 food parcels (ca. 17CHF/parcel for one month) to a Syrian family”

“DONATE TODAY: your donation can supply food parcels (ca. 17CHF/parcel for one month) to a Syrian family”

The mailing went out on April 22, 2021. An unrelated ICRC “door-drop” campaign followed around May 25, and a TV spot ran in German-speaking Switzerland in the same period. This triggered the preregistration’s Case I contingency: the narrow window (gifts through May 31) is the principal sample for hypothesis tests, with the broad three-month window as descriptive robustness. The narrow flag keeps all letters and simply does not count gifts arriving after May 31.

I. Impact of providing information about unit cost in a fundraising solicitation — primary questions.

3  Main results: The impact of impact information

information

anipulation strength and relevance: the cost-per-output treatments must have been meaningful, salient, and representative of what charities actually do and would

Our closest arm-level comparison holds the CHF 150 ask fixed. The control text asks for CHF 150 and mention

scheme with a large unit size substantially increased average giving.

not yet reviewed by the authors.

Table 4.2: Preregistered amount contrasts (SUG-HI-LOW-AMT, H10–H11), narrow sample. Outcome: donation amount in CHF including zeros (revenue per letter).

Know of high-impact research we should consider for evaluation — or other work this tool should be doing? Leave an email if you're open to follow-up discussion; if we later introduce compensation for useful contributions, earlier contributors will be grandfathered in.

Since the workshop: a company-reported production run at 22,000 litres and a new hydrolysate review bear directly on the scale and media questions. Dated updates and their limits.

9, 2026. These developments

kind of food people would buy

generally expected

wide

differed over how far particular substitutions could go,

Rate/discuss (Team) Rate/discuss (Public)

Quick rate

Unjournal commission an evaluation? (-- Strong No … ++ Strong Yes)For aggregation and sorting, the five quick ratings translate to percentile-equivalent scores: -- = 10, - = 30, ~ = 50, + = 70, and ++ = 90. These are deliberately spaced away from 0 and 100..

More filters and source controls

Show/hide scoring rubrics & methodology

About this prioritization tool

Optional uncertainty range (10th and 90th percentiles)

Can you give an A1 cost estimate?

Jump to another question

Full definitions and accounting boundaries

at least 2,000 metric tonnes per plant per calendar year (2 kt/year).

Qualifying

The authors have made the atlas, methods, data, and replication code easy to inspect. We found some research indexing and light public discussion, but no major institution or policy report using the measures yet.

Several plausible routes could lower cultivated meat production costs.

Several plausible routes could lower cultivated meat production costs. The harder question is whether they can work together in a reliable process at commercial scale. Our current reading of the evidence favors testing complete, internally consistent production scenarios before treating lower ingredient prices or higher cell densities as a new cost forecast.

Key Takeaways

Embedded in media discussion

Framing note: Recent industry consolidation — including several company closures in 2024–2025 — is one set of data points, not a verdict on any particular TEA's methodology. Company outcomes depend on funding conditions, market timing, and management decisions that are largely independent of the underlying cost trajectory. The workshop aimed to evaluate the technical evidence on its merits.

PF

Public working summary Workshop held May 8, 2026 Last reviewed July 20, 2026 S2 was off-record and is excluded. Its section below gives only the announced session scope.

Informal Pre-Session Wed May 6, 2026 · 11:00am–12:00pm ET · Zoom · recorded Casual orientation for participants with limited May 8 availability — mostly preparation and broad discussion. Walkthrough of the beliefs form and cost model. David Manheim (Technion/ALTER) and Mirjam Capuder (University of Maribor) participated. It covered introductions, a walkthrough of the interactive cost model dashboard, and early framing questions about key modeling uncertainties — including Manheim's perspective on expert elicitation methodology and how to structure belief-updating around contested TEA assumptions. Substantive points from this session are incorporated where relevant in the S1 and S3 notes below.

These map onto the workshop's structure: cell-line engineering → Crux 1; hydrolysate substitution → Crux 2; albumin and recombinant proteins → beliefs-form question E4; "what share of a hybrid product is cultivated?" → Crux 4.

Regulatory milestones are not directly comparable across jurisdictions. The FDA's February 2026 inventory lists five completed US cultured-cell food consultations, including three completed in 2025; market authorization, inspection, and product scope can still differ.

it

Evaluation 2 of this forecast

in response to this and other evidence.

S2-derived claim,

rigorous

asitka et al. (2024): ~$6/lb under continuous production; Humbird (2021):

S2 was off-record and is excluded. Its section below gives only the announced session scope.

Tool updated: Aug 28, 2026Normal cadence: papers twice weekly; feedback hourlyPaper discovery, scoring, matching, and the full dashboard rebuild normally run on Monday and Thursday. New rating and suggestion submissions are checked up to hourly. The dated chips show the actual latest completed updates.Last paper scan: Aug 31, 2026Last AI scoring: Aug 31, 2026Crux/PQ match update: Jul 23, 2026Human feedback: 17 current ratings across 13 papersPublic anonymous aggregate refreshed Sep 1, 2026. Each rater's latest rating per paper counts once. Discussion and a display name are included only after explicit public-sharing consent; emails, rating keys, and submission ids are never included.Earlier API cost: at least ~$4.06 reconstructed

Normal cadence: papers twice weekly; feedback hourly

A public economics column lists the review among its sources. This suggests some circulation outside the World Bank, although the article does not assess the review’s methods or health-policy conclusions in detail.

Mean absolute score movement?

?

movement?

Agreement alpha?

ecent AI-impact/governance papers Ordinary rerun

AI score stability metrics, with paper cohorts and variation sources shown side by side AI stability metric

17.2 points

cautious

rivate review found that one run treated prior public scrutiny as substantial while the other called it unclear, materially changing the estimated value of a new evaluation

identified publicly.

One same-wording execution pair differed by 30 points.

Prior human-rated comparison papers were less stable on both dimensions: ordinary rerun movement averaged 8.5 points and the wording effect averaged 5.8 points. 2/6 (33%) had wording effects larger than ordinary rerun movement; wording-mean alpha was 0.69.

Recent AI-impact/governance papers: ordinary same-wording reruns differed by 4.1 points on average, while the repeat-averaged wording effect was 1.1 points. 0/6 (0%) had a wording effect larger than their ordinary rerun variation, and wording means produced no broad-tier changes. Prior human-rated comparison papers were less stable on both dimensions: ordinary rerun movement averaged 8.5 points and the wording effect averaged 5.8 points. 2/6 (33%) had wor

Recent AI-impact/governance papers: ordinary same-wording reruns differed by 4.1 points on average, while the repeat-averaged wording effect was 1.1 points. 0/6 (0%) had a wording effect larger than their ordinary rerun variation, and wording means produced no broad-tier changes.

ordinary same-wording reruns differed by 4.1 points on average, while the repeat-averaged wording effect was 1.1 points.

Six recent AI impact/governance papers and six papers with prior Unjournal team ratings

026-08-31T23:54:58.113454+00:00

Held-out check: a third baseline execution is kept outside the balanced headline analysis and used only as a robustness check.

What is being tested: every stability number below describes AI model scores. “Previously human-rated papers” names a paper cohort selected because earlier Unjournal ratings are available; those ratings do not enter the stability calculations. Human and AI scores are compared separately in the amber reference panel.