D = import(new URL("./donor-compass.mjs", window.location.href).href)
M = import(new URL("./xrisk-model.mjs", window.location.href).href)
Plot = import("https://cdn.jsdelivr.net/npm/@observablehq/plot@0.6.17/+esm")
snap = fetch(new URL("./rp-donor-compass.json", window.location.href).href).then(r => { if (!r.ok) throw new Error(`couldn't load the data file (${r.status})`); return r.json(); })
P = snap.current.projects
opts = ({budget: snap.current.budget, increment: snap.current.incrementSize, step: snap.current.drStepSize})
profiles = snap.current.riskProfileOptions.map(o => o.label)
// "?" markers. site.js shows the text on hover, keyboard focus or tap.
tipEl = text => { const s = document.createElement("span"); s.className = "tip"; s.tabIndex = 0; s.setAttribute("role", "button"); s.setAttribute("aria-label", "Explanation: " + text); s.dataset.tip = text; s.textContent = "?"; return s; }
lab = (text, tip) => { const s = document.createElement("span"); s.append(text, " ", tipEl(tip)); return s; }
FUNDS = ["givewell", "leaf", "ea_awf", "navigation_fund_general", "navigation_fund_cagefree", "sentinel_bio", "longview_nuclear", "longview_ai"]
COLORS = ({givewell: "#2b6cb0", leaf: "#319795", ea_awf: "#68a357", navigation_fund_general: "#2f855a", navigation_fund_cagefree: "#9ac16f",
sentinel_bio: "#dd6b20", longview_nuclear: "#975a16", longview_ai: "#c53030"})
chartW = Math.max(300, Math.min(980, width))
narrow = chartW < 560
SHORT = ({givewell: "GiveWell", leaf: "LEAF", ea_awf: "EA AWF", navigation_fund_general: "Navigation, general",
navigation_fund_cagefree: "Navigation, cage-free", sentinel_bio: "Sentinel Bio", longview_nuclear: "Longview nuclear", longview_ai: "Longview AI"})
labelOf = k => narrow ? SHORT[k] : D.FUND_LABELS[k]
fundColor = keys => ({domain: keys.map(labelOf), range: keys.map(k => COLORS[k])})
barMargins = ({marginLeft: narrow ? 128 : 190, marginRight: narrow ? 78 : 90})
pct = (x, d = 0) => `${(100 * x).toFixed(d)}%`
money = x => `$${Math.round(x)}M`
versionName = id => snap.versions.find(v => v.id === id).name.replace(/ at .*$/, "")
RISK_TIP = "How Donor Compass turns a fund's spread of simulated outcomes into one number. 'Neutral' is the plain average. 'Upside Sceptical' and 'Continuous Upside Sceptical' discount very large, unlikely gains; 'Downside Critical' counts harmful outcomes five times; 'Combined' does both. WLU (weighted linear utility) gives large gains less weight and losses more, more so as c rises. 'Upside Sceptical' and 'Ambiguity Bilateral' are in the data but not described in the methodology we read."From the report to Donor Compass
How Rethink Priorities’ Donor Compass and Cross-Cause Fund draw on Muñoz Morán (2023), shown with RP’s own published numbers
Rethink Priorities’ Donor Compass suggests how to split a donation across eight funds, and RP’s Cross-Cause Fund uses the same model for its own allocation. Three of the funds work on global catastrophic risks (GCR): Sentinel Bio, and Longview’s nuclear and AI funds. Their numbers come from an RP model that lists this report as an influence. This page follows that link step by step, then uses RP’s published data to show what the report’s kind of reasoning does to the recommendations.
It’s our reading of RP’s public documents, data and code, checked on 19 October 2026. RP hasn’t reviewed it. Corrections are welcome as a Hypothes.is annotation or by email.
NoteIn short
- Donor Compass doesn’t run the report’s model, but its GCR numbers come from a close relative. RP’s GCR fund write-up says its model “was based on work by Tarsney (2020)… It was also influenced by Thorstad (2022) and Munoz Moran (2023).” It keeps the report’s core (a time of perils, a proportional cut in risk that lasts a few years, a path for the value of the future) and adds things the report left for later: Monte Carlo, effects that may be zero or harmful, and a branch where humanity settles space.
- Almost all of the GCR funds’ expected value lands after year 500. In RP’s data, the 500+ year window holds about \(10^{14}\) times as much as all the earlier windows put together: about \(6 \times 10^{19}\) life-years per $1M for the bio and nuclear funds, \(1.6 \times 10^{20}\) for the AI fund.
- So in Donor Compass the far-future number works almost like a switch. Its sign matters far more than its size. In RP’s worldviews that use plain expected value, a factor above about \(10^{-13}\) on the 500+ window is enough for a GCR fund to outscore every other fund. A risk profile that makes that window negative (five of RP’s nine do, for the bio fund) sinks them. With RP’s 14 default worldviews and its mix of six aggregation methods, the GCR funds get 35.5% of a $200M budget, close to the 35% RP gave them in the Cross-Cause Fund’s first distribution. Multiply every far-future value by 10, or divide it by a billion, and none of the six methods changes its split, including Nash bargaining and maximizing expected choiceworthiness, which compare values across worldviews. Give the 500+ window no weight and the GCR share falls to 25%; give every worldview the same risk profile and it ranges from 0% to 90%.
- The evaluation’s three points carry over, in a new form (below). Whether an effect could be harmful becomes a choice of risk profile. The value-growth question becomes RP’s 6% chance of cubic growth in space, which RP says dominates the average. And the horizon becomes a single factor on everything after year 500.
The chain from the report to a recommendation
RP uses “cross-cause model” for two things. One is the 2023 Cross-Cause Cost-Effectiveness Model, a Monte Carlo tool for single projects. The other is the 2026 model behind Donor Compass and the Cross-Cause Fund, whose default inputs are headed “Rethink Priorities Cross-Cause Model Inputs”. The report’s ideas reach the second through RP’s GCR fund model:
- 1. The reportOctober 2023the report's kind of model
What a small cut in extinction risk, lasting a few years, is worth under different paths for risk and for the value of the future. This site re-implements it.
- 2. Cross-Cause Cost-Effectiveness ModelNovember 2023the report's kind of model
Its x-risk module has the report's structure: "We measure mitigation effectiveness by the proportional reduction of yearly risk." Its overview thanks the report's author for "conceptual guidance".
- 3. RP's GCR fund model2026, "a working draft"the report's kind of model
A time of perils; each fund cuts its cause's peak risk for about 11 years; a 50% chance of no effect and a 5% to 15% chance of raising risk; value that grows logistically on Earth, or cubically if humanity settles the stars (6% chance); 10 million Monte Carlo draws. The results are split into six time windows. Write-up.
- 4. Published fund datawhere the two meet
For each fund: impact per $1M by type of effect, time window (0-5, 5-10, 10-20, 20-100, 100-500 and 500+ years) and risk profile, plus a schedule of diminishing returns. The GCR funds' impact is in human life-years, from cutting extinction risk and three sizes of non-extinction catastrophe. Data viewer.
- 5. Scoring under each worldviewDonor Compass
Each worldview weights the types of effect (moral weights), weights the six windows, and picks a risk profile. RP's defaults also cut the score of every fund except the AI fund by 40%, for the chance that a near-term AI catastrophe makes their work moot. Methodology.
- 6. Combining worldviewsDonor Compass
Aggregation methods combine the worldviews, each handing out the budget $2M at a time as diminishing returns set in. A donor's answers set the worldviews in Donor Compass; for the Cross-Cause Fund's default allocation RP uses its own 14, and averages the splits from six methods (below).
What carried over, and what changed
| The report | RP’s GCR fund model | What Donor Compass sees | |
|---|---|---|---|
| Risk | A time of perils (20% per century for 100 years, then low) is one of four paths | A time of perils with uncertain timing: 20% (5% to 40%) “Chance of existential catastrophe in the next 100 years”, peaking in year 17, lasting 90 years | Nothing directly; it’s inside the fund numbers |
| The action | Risk cut by a share \(f\) for \(P\) years | “relative reductions in their cause’s peak risk levels during the time of perils”, lasting 11 years (2.5 to 30) | Nothing directly |
| Sign | Non-negative in every analysis | 50% chance of no effect; “Probability of raising risk” 5% (bio, nuclear) or 15% (AI) | The risk profile decides how much the harmful draws count |
| Value of the future | Five paths, from constant to cubic or logistic | Logistic on Earth; with stellar settlement (6% chance), “The total value of civilization grows cubically” | Almost all of it lands in the 500+ window |
| Horizon | Up to 120,000 years | No end date | One factor for everything after year 500 |
| Uncertainty | Fixed values; means only | Monte Carlo over every parameter (10 million draws since June 2026) | Nine risk profiles, each a different summary of the draws |
| Non-extinction catastrophes | Not modeled | Three sizes, from 10 million deaths up, in the next 100 years | Added to the extinction figures in the early windows |
Sources for each row are on the Departures page, along with our approximation of RP’s central values in the report’s model.
Where the value sits
These are RP’s numbers: expected human life-years per $1M for each fund, by time window. The GCR funds’ bars add up their extinction and non-extinction effects. For GiveWell and LEAF we show only their life-year effects; Donor Compass also counts their effects on disability and income. The scale is logarithmic, so each gridline is a factor of 100,000, and negative values hang below the line.
The numbers in the chart
Under “Neutral” (plain expected value), the 500+ window holds all but about one part in \(10^{14}\) of each GCR fund’s value. Switch to either of the stronger WLU profiles, “Combined” or either upside-skeptical profile, and the bio and nuclear funds’ far-future values turn hugely negative; for the AI fund, every profile except “Neutral” and “Ambiguity Bilateral” does that. As we read it, once the largest gains are discounted, the draws in which a fund raises risk dominate.
RP’s write-up shows why the profiles matter so much. Its expected life-years per $1M, over all time windows:
| Fund | Mean | 1st percentile | 10th | 50th | 90th | 99th |
|---|---|---|---|---|---|---|
| Sentinel Bio | \(7 \times 10^{19}\) | \(-1 \times 10^{10}\) | 0 | 0 | \(1 \times 10^{10}\) | \(1 \times 10^{16}\) |
| Longview nuclear | \(6 \times 10^{19}\) | \(-1 \times 10^{10}\) | 0 | 0 | \(1 \times 10^{10}\) | \(1 \times 10^{16}\) |
| Longview AI | \(2 \times 10^{20}\) | \(-2 \times 10^{12}\) | \(-2 \times 10^{9}\) | 0 | \(8 \times 10^{10}\) | \(1 \times 10^{16}\) |
The mean is 6,000 to 20,000 times the 99th percentile, so almost all of it comes from the top 1% of draws. RP puts it this way: “the tails of the distribution – those in which the intervention has a large impact, and wherein the future has rapid star settlement – dominate the average value of the funds’ activities. The vast majority of the simulations have zero or lower values. As such, risk attitudes will be determinative in evaluating GCR funds.”
The version menu shows how much the far-future cell can move. Between the May and June 2026 versions, the 500+ values rose four- to fivefold while the earlier windows barely changed. RP’s update log records two June changes: a fix to the value of the future under stellar expansion, and a move from 1 million to 10 million draws. It says the funds’ cost-effectiveness rose “by about 3-4x, but it’s not quite clear how much of that increase was due to the fixes vs. resampling from the very heavy right tailed distributions”, and that “these fixes had a negligible impact on the recommended allocations between funds.” The next two sections show why both statements can be true.
One worldview at a time
Donor Compass scores each fund under each worldview. It multiplies each effect by the worldview’s moral weight and by the factor for its time window, reads the values for the worldview’s risk profile, and adds everything up (methodology, section 4). Pick a starting point, then change the factors to see what moves the ranking. The bottom chart shows what this worldview alone would do with $400M, the default budget in RP’s data.
With a single worldview, RP’s aggregation methods agree, apart from how they break exact ties: each $2M goes to the fund with the highest score at the margin, after diminishing returns, and a fund gets nothing more once its schedule reaches zero. In this data that happens at $30M for Longview nuclear, $38M for Sentinel Bio, $352M for Longview AI, between $32M and $84M for the animal funds and $302M for LEAF; GiveWell never fills up.
RP’s 14 default worldviews
For the Cross-Cause Fund, RP weighs 14 worldviews, from total utilitarianism to Kantianism (default inputs). The table shows each one’s settings for the far future, and what it would do alone with $400M.
Three patterns stand out. Worldviews that give the 500+ window any weight, at a risk profile where it’s positive, fund the GCR funds first, until those funds can’t absorb more. Worldviews whose risk profile makes that window negative give them nothing, however small the factor. And worldviews that ignore the far future can still fund them through the 20-100 year window, where the AI fund’s estimated impact is large and most other funds have little or none.
Combining them
For the Cross-Cause Fund’s default allocation, RP mixes six aggregation methods: Nash bargaining (35%), credence-weighted (30%), maximizing expected choiceworthiness (MEC, 15%), maximizing expected truthlikeness (MET, 10%), split-cycle voting (8%) and the Borda count (2%). Since 30 April 2026, each method splits the whole budget on its own, and the allocation is the weighted average of the six splits (update log). All six hand out the budget $2M at a time, valuing each fund at the margin after diminishing returns. As we read the app’s code:
- Credence-weighted: each worldview spends its credence’s share of every $2M on the fund it rates highest.
- Nash bargaining starts from a fall-back point, where each worldview spends its own share on its favorite fund. It gives the whole $2M to one fund only if every worldview likes that at least as well as the fall-back point (picking the fund with the largest product of gains); otherwise it splits the $2M as at the fall-back point. With RP’s 14 worldviews, as RP set them, no fund ever passes that test, so Nash bargaining gives the same split as the credence-weighted method.
- MEC adds up the worldviews’ values for each fund, weighted by credence, and gives the $2M to the fund with the highest total.
- MET acts on the most credible worldview. When none has at least half the credence (RP’s largest has 10%), it places the worldviews on a map by how alike they value and rank the funds, and acts on the one nearest the credence-weighted center.
- Split-cycle voting and the Borda count treat each worldview as a voter whose vote counts by its credence. They use only each worldview’s ranking of the funds, not the size of its values. Borda gives the $2M to the fund with the most points; split cycle shares it among the funds that no other fund beats head to head, after dropping the weakest win in each cycle.
RP’s update log says four of these methods now “cap the amount of funding that any fund can receive at 5x its current budget”. In the app’s code, that cap is the point where a fund’s diminishing-returns schedule reaches zero, and the other methods stop there too. The app sets no other caps unless a user adds them.
We wrote our own versions of the six methods and of the averaging, working from the app’s code, then ran RP’s own functions on the same data and worldviews. In every case in the table further down, at both budgets, the two agree on every fund’s share to within \(10^{-13}\) percentage points.
The budget matters, because funds fill up. The Cross-Cause Fund page splits $200M; RP’s Donor Compass data set $400M, which the sections above use. At $200M, our version of RP’s mix lands close to what RP published:
Every fund is within a quarter of a point of RP’s first distribution round. RP’s “recommended allocations” table, in whole percents, differs from ours by one point for two funds; we don’t know why. This is still not RP’s recommendation. It’s our re-implementation of RP’s published method on RP’s published data, and RP lists “Our best judgment” (“Where our models may be incomplete or wrong”) among the factors its allocation considers.
Try changing every worldview at once:
So the answer to the question we started with is no. Under every one of the six methods, multiplying all of RP’s far-future values by 10, or dividing them by as much as a billion, leaves the split exactly where it was. That includes the two methods that compare values across worldviews. Nash bargaining reduces here to the credence-weighted split. MEC does add up the worldviews’ values, so for each GCR fund it weighs huge positive far-future values (mostly under “Neutral”) against huge negative ones (under the WLU and “Combined” profiles). Scaling them all together doesn’t change which side wins: for the bio and nuclear funds the positive side does, and they get all they can absorb; for the AI fund the negative side does, and MEC gives it nothing. Only when the far-future values shrink by something like a trillion, to the size of the near-term values, do the splits start to move.
What does move the allocation is whether the 500+ window counts at all, and which risk profiles the worldviews use. With no weight on that window, RP’s mix gives the GCR funds about 25% of $200M rather than 35.5%; with every worldview risk-neutral it gives them 90%, and with every worldview on the WLU (c=0.05) profile, almost nothing. For the two voting methods, the far future as RP set it cuts the other way: they give the GCR funds less than they would with no weight on the 500+ window, because under the profiles that make it negative it puts those funds at the bottom of those worldviews’ rankings.
That fits RP’s note that a three- to fourfold rise in the GCR funds’ cost-effectiveness had “a negligible impact on the recommended allocations”. With numbers this large, the far future decides whether a GCR fund comes first, and its sign decides which way.
The same pattern in the report’s model
The report’s model, run on this site, also puts most of a mitigation’s value after year 500, just far less extremely:
The space-settlement branch is what pushes RP’s numbers so far beyond the report’s. Our RP preset leaves it out, which is why it understates RP’s average (Departures). To apply Donor Compass-style window factors to the report’s model, turn on switch 4 on the model page; this link opens the RP preset with the factors of RP’s first worldview.
What this means for the evaluation’s points
The Unjournal’s evaluation of the report made three main points. Each has a counterpart in Donor Compass:
- Sign. The evaluation noted that the report’s analyses treat the effect of an action as known and positive. RP’s GCR model doesn’t: it gives each fund a 50% chance of no effect and a 5% to 15% chance of raising risk. But its average is so dominated by the best draws that the harmful ones only show up once a risk profile discounts those draws. In Donor Compass, then, the sign question becomes a question about attitudes to risk, and that’s the setting the GCR share is most sensitive to.
- Value growth. The evaluation argued that the value-growth paths, up to cubic with no ceiling, do much of the work. RP’s GCR model includes a cubic path, in its 6% space-settlement branch, and RP says draws “wherein the future has rapid star settlement” dominate the funds’ average. So the concern applies to RP’s numbers with more force, even though, as above, none of RP’s six aggregation methods is sensitive to their size.
- Horizon. The evaluation argued that with any floor on risk, the choice of a finite horizon matters more than whether an infinite sum converges. RP’s 500+ window has no end date. A donor’s only control over the far future is one factor applied to all of it, and since that window is about \(10^{14}\) times the rest, under plain expected value any factor above about \(10^{-13}\) gives it the deciding vote. That’s a blunt way to state a view about the far future.
Sources, checks and caveats
- RP’s fund data. The Donor Compass data viewer, version “June 4, 2026” (the one it opens on, as of 19 October 2026), and the three earlier versions in the same file; RP’s overview explains the format. We copied the parts this page uses into rp-donor-compass.json. The data are RP’s. Our tests check that the GCR figures match the table in RP’s write-up.
- RP’s default worldviews: the spreadsheet linked from the methodology page, checked 19 October 2026.
- Scoring and allocation: the methodology, RP’s update log, and the scoring and aggregation code in the Donor Compass app, read on 19 October 2026. Our version of the scoring, the six aggregation methods RP mixes and the mix itself is in donor-compass.mjs. Tests compare its splits with those from RP’s code on the same inputs.
- RP’s published allocation: the Cross-Cause Fund page, checked 19 October 2026, for the $200M budget, the “recommended allocations” table and the July 2026 distribution round.
- RP’s GCR fund model: the write-up, “a working draft”, unchanged between 12 and 19 October 2026; and the update log. We use the model’s outputs; we haven’t checked the model itself.
- Not done: the way Donor Compass turns a donor’s answers into worldviews, and three methods the app offers that RP’s mix doesn’t use (“my favorite theory”, MSA and lexicographic maximin).
- One number we can’t place. RP’s walkthrough describes GCR effects “in the 500+ year column on the order of \(10^{32}\) human life-years”. None of the four data versions comes close: the largest 500+ value in any of them is \(1.6 \times 10^{20}\).
Quotations are verbatim. Everything else is our summary and may be out of date. The model page and its other companions: Model · Departures · Context.