How close is the reconstruction?

Our rebuild of RP’s GCR fund model against RP’s published numbers, window by window, with Monte Carlo error

Help improve this draftComment with Hypothes.is; flag @daaronr for a response.

Select any text, chart or number and annotate it with Hypothes.is (the tab at the top right) to flag errors, missing sources, or a clearer way to show something. David plans to respond, especially where a comment mentions @daaronr. Public Hypothes.is annotations can be read by anyone who opens this page, so don't post anything confidential.

How this was made: an AI agent (Claude, working for The Unjournal) rebuilt Rethink Priorities' GCR fund model from RP's public write-up, update log and Donor Compass data, and built these pages. David Reinstein set the task. Neither he nor RP has checked the numbers yet, and RP hasn't seen this page.

How to use Hypothes.is Read comments on this page

In short. At RP’s stated values, the reconstruction lands within about 5% of RP’s published life-years per $1M in every window up to 500 years, for all three funds, and 12-18% low in the 500+ window, which RP’s own datasets move around by far more than that. The ranking of the funds matches RP’s in every window except where RP’s bio and nuclear numbers are within a few percent of each other. We had to infer or assume several things RP’s write-up doesn’t state; they are listed at the bottom.

D = import(new URL("./engine/dashboard.mjs", window.location.href).href)
UI = import(new URL("./panel.mjs", window.location.href).href)
Plot = import("https://cdn.jsdelivr.net/npm/@observablehq/plot@0.6.17/+esm")
rep = fetch(new URL("./data/replication.json", window.location.href).href).then(r => { if (!r.ok) throw new Error(`couldn't load the replication results (${r.status})`); return r.json(); })
tipEl = UI.tipEl
chartW = Math.max(300, Math.min(900, width))
narrow = chartW < 600
fundName = k => D.FUND_BY_KEY[k].short

Ours versus RP’s, window by window

Each dot is our value divided by RP’s: 1 means an exact match. The shaded band is within 10%. The bars show how far our value moves across ten further runs of a million draws each (Monte Carlo noise).

ratioChart = {
  const rows = D.FUND_KEYS.flatMap(fk => rep.funds[fk].windows.map((w, k) => {
    const rb = k === 5;
    const ours = rb ? w.oursRB : w.ours, lo = rb ? w.seedRBMin : w.seedMin, hi = rb ? w.seedRBMax : w.seedMax;
    return {fk, fund: fundName(fk), window: w.window, ratio: ours / w.rpData, lo: lo / w.rpData, hi: hi / w.rpData, ours, rp: w.rpData};
  }));
  return Plot.plot({
    width: chartW, height: 300, marginLeft: 50, marginBottom: narrow ? 50 : 34, marginTop: 26,
    ariaLabel: "Ratio of this reconstruction's values to RP's published values, by window and fund",
    ariaDescription: rows.map(d => `${d.fund}, ${d.window}: ${d.ratio.toFixed(2)}`).join("; "),
    fx: {domain: D.WINDOWS, label: null, tickFormat: w => `${w} yrs`, tickRotate: narrow ? -35 : 0, padding: 0.1},
    x: {domain: D.FUND_KEYS.map(fundName), axis: null, padding: 0.3},
    y: {domain: [0.7, 1.15], grid: true, label: "↑ Ours / RP", tickFormat: d => d.toFixed(2)},
    color: {domain: D.FUNDS.map(f => f.short), range: D.FUNDS.map(f => f.color), legend: true},
    marks: [
      Plot.rectY([0], {y1: 0.9, y2: 1.1, fill: "#e6fffa"}),
      Plot.ruleY([1], {stroke: "#1a202c"}),
      Plot.ruleX(rows, {fx: "window", x: "fund", y1: "lo", y2: "hi", stroke: "fund", strokeWidth: 3, strokeOpacity: 0.5}),
      Plot.dot(rows, {fx: "window", x: "fund", y: "ratio", fill: "fund", r: 5, stroke: "white"}),
      Plot.tip(rows, Plot.pointer({fx: "window", x: "fund", y: "ratio", title: d => `${d.fund}, ${d.window} years\nOurs: ${D.fmt(d.ours)}\nRP: ${D.fmt(d.rp)}\nRatio: ${d.ratio.toFixed(3)} (seed range ${d.lo.toFixed(2)} to ${d.hi.toFixed(2)})`}))
    ]
  });
}
html`<p class="rg-fig-note">Main run: ${rep.mainN.toLocaleString("en-US")} draws (seed ${rep.options.seed}). Bars: ${rep.seedRuns} further runs of ${rep.seedN.toLocaleString("en-US")} draws. RP: Donor Compass dataset ${rep.rpDataset.name}, risk-neutral column. The 500+ window uses our lower-noise estimate (it averages over whether humanity settles the stars instead of sampling it; same expected value).</p>`

{
  const cls = r => Math.abs(Math.log(r)) < Math.log(1.1) ? "rg-ok" : Math.abs(Math.log(r)) < Math.log(1.25) ? "rg-warn" : "rg-bad";
  const rows = D.FUND_KEYS.flatMap(fk => rep.funds[fk].windows.map((w, k) => {
    const rb = k === 5, ours = rb ? w.oursRB : w.ours, r = ours / w.rpData;
    const lo = rb ? w.seedRBMin : w.seedMin, hi = rb ? w.seedRBMax : w.seedMax;
    return html`<tr><td>${k === 0 ? html`<strong>${fundName(fk)}</strong>` : ""}</td><td>${w.window}</td>
      <td class="num">${D.fmt(w.rpData)}</td><td class="num">${D.fmt(w.rpWriteup, 2)}</td><td class="num" title=${D.fmtFull(ours)}>${D.fmt(ours)}</td>
      <td class=${"num " + cls(r)}>${r.toFixed(2)}</td><td class="num">${rb ? "" : `±${D.fmt(2 * w.batchSE, 2)}`}</td><td class="num">${D.fmt(lo, 2)} to ${D.fmt(hi, 2)}</td></tr>`;
  }));
  return html`<details><summary>The numbers</summary><table class="table rg-table"><thead><tr><th>Fund</th><th>Years</th><th class="num">RP's data</th><th class="num">RP's write-up</th><th class="num">Ours</th><th class="num">Ours / RP</th><th class="num">± (main run)</th><th class="num">Seed range</th></tr></thead><tbody>${rows}</tbody>
    <tfoot><tr><td colspan="8">Expected life-years per $1M, existential plus non-existential. "±": two standard errors from ten batches of the main run (not shown for 500+, where it understates the noise). Shading: within 10% (green), 10-25% (yellow).</td></tr></tfoot></table></details>`;
}

Monte Carlo error

Up to 500 years, our numbers are steady: two standard errors are about 0.3% of the value at 10 million draws, and the ten runs of a million draws agree to within 1-4%. So the 2-6% gap to RP in the 10-500 year windows is mostly a real difference between our rebuild and RP’s model, not noise.
The 500+ window is noisy for everyone. Its average rests on a few hundred draws in which humanity settles the stars and long-run risk is very low. Our runs of a million draws spread over about ±7% even with the lower-noise estimate. RP’s own 500+ values moved by 3 to 4 times between its May and June 2026 datasets, and RP says it can’t separate the effect of its June fix from resampling noise.
On the model page the default is 100,000 draws, so expect roughly ±15-25% in the 500+ window and about ±3% in each window up to 500 years; “Run more draws” narrows it.

Other checks

{
  const pf = ["p1", "p5", "p10", "p50", "p90", "p95", "p99"], ix = [0, 1, 2, 3, 4, 5, 6];
  const nonRows = D.FUND_KEYS.flatMap(fk => {
    const f = rep.funds[fk];
    return [html`<tr><td><strong>${fundName(fk)}</strong></td><td>RP</td><td class="num">${D.fmt(f.rpNonext.mean, 2)}</td>${pf.map(k => html`<td class="num">${D.fmt(f.rpNonext[k], 2)}</td>`)}</tr>`,
      html`<tr><td></td><td>ours</td><td class="num">${D.fmt(f.nonMean)}</td>${ix.map(i => html`<td class="num">${D.fmt(f.pct.non[i])}</td>`)}</tr>`];
  });
  const head = html`<tr><th>Fund</th><th></th><th class="num">Mean</th>${["1st", "5th", "10th", "50th", "90th", "95th", "99th"].map(h => html`<th class="num">${h}</th>`)}</tr>`;
  const v = rep.valueOfFuture, R = v.rp;
  const pv = [["0.1", "p0_1"], ["1", "p1"], ["5", "p5"], ["50", "p50"], ["95", "p95"], ["99", "p99"], ["99.9", "p99_9"]];
  return html`<p><strong>Non-existential effects</strong> (smaller catastrophes, next 100 years only), life-years per $1M. We fitted one number (life-years lost per death) to Sentinel Bio's mean; everything else follows.</p>
    <table class="table rg-table"><thead>${head}</thead><tbody>${nonRows}</tbody></table>
    <p><strong>The value of the future</strong> (life-years, before any fund acts), percentiles across draws.</p>
    <table class="table rg-table"><thead><tr><th></th><th class="num">Mean</th>${pv.map(([p]) => html`<th class="num">${D.ordinal(Number(p))}</th>`)}</tr></thead>
    <tbody><tr><td>RP</td><td class="num">${D.fmt(R.mean, 1)}</td>${pv.map(([, k]) => html`<td class="num">${D.fmt(R[k], 1)}</td>`)}</tr>
    <tr><td>Ours</td><td class="num">${D.fmt(v.meanRB, 2)}</td>${pv.map(([p]) => html`<td class="num">${D.fmt(v.pct[p], 2)}</td>`)}</tr></tbody></table>`;
}

What doesn’t match

Every window from 10 to 500 years is 2-6% low, getting worse with time, and the 500+ window is 12-18% low. We couldn’t remove this with the risk-curve shapes we tried without making the early windows too high. It is probably a detail of RP’s risk curve or value curve that the write-up doesn’t give.
The 99.9th percentile of the value of the future is 1.7 × 10²⁹ in ours and 2 × 10²⁸ in RP’s table, while the mean and every other percentile match. Our extreme tail has more draws in which the Milky Way is fully settled and long-run risk is very low. None of the variants we tried moves it. Either RP’s tail is thinner at that point, or the table is from a different run.
The AI fund’s non-existential value is 5% high, and its 0-5 year window 4% high.
The growth rate’s mean is 1.17% in ours; RP states 1%. No two-parameter distribution we tried gives a 1% mean with RP’s 0.6% to 2% range. It barely matters.

What we inferred and what we assumed

“Inferred” means RP’s published numbers point to it. “Assumed” means we chose it, sometimes by fitting RP’s numbers, and it could be wrong. On the model page, controls carry the same tags.

What	What RP says	What we did	Status
How uncertain inputs are spread	A mean and a 90% range for each	Probabilities: a beta distribution through the mean and range. Positive quantities: a lognormal through the range. Means then match RP’s to within 2-3%	Assumed; checks well
Shares of risk by cause	AI 90%, bio 3%, nuclear 3%, other 4%, with ranges	Drawn together, from a Dirichlet(9, 0.3, 0.3, 0.4); it reproduces all eight of RP’s percentiles	Inferred, strong
No effect / raises risk / lowers risk	Means and ranges for the first two	Drawn together, from a Dirichlet with concentration 5; it matches RP’s AI-to-bio ratio and RP’s negative percentiles	Inferred, strong
Long-run risk and settlement speed	Mean and range	Log-uniform through RP’s range. A lognormal through the same range gives the wrong mean and an unstable average	Inferred
When effects start and how long they last	Means and ranges	Lognormal; matches RP’s non-existential window shares to within 0.002	Inferred, strong
The three funds’ random draws	Not stated	Shared: the same world and the same luck for each fund. RP’s bio and nuclear non-existential totals agree to five digits	Inferred
Shape of the risk curve	Peaks within decades, stays high about 90 years	A straight rise from 80% of the peak’s extra risk to the peak, then a straight fall to the long-run level, fitted to RP’s numbers	Assumed, fitted
How a fund’s cut applies	“relative reductions in their cause’s peak risk levels during the time of perils”	A share of the cause’s risk in each year while the effect lasts	Assumed, fitted
When a fund backfires	Not stated	The same size of effect, with the opposite sign	Assumed; matches RP’s percentiles
Value of settling the stars	Each star worth Earth at carrying capacity (June 2026 fix); Milky Way and Virgo densities	Tarsney’s (2020) cubic growth model as written	Assumed (from Tarsney)
Smaller catastrophes	Chances in the next 10 years, for three sizes	A constant yearly chance for 100 years; deaths in the middle of each size range (on a log scale); 25.6 life-years per death, fitted	Assumed, one number fitted
Time	Not stated	Continuous; the choice changes results by well under 1% after the first five years	Assumed

Details, and the sources, are in the Technical notes.