How the Donor Compass
model works

A step-by-step walkthrough with real examples from the Cross-Cause Prioritization Model

This document walks through the model underlying the Rethink Priorities Donor Compass and Cross-Cause Fund. It uses a series of concrete examples to show how worldview inputs translate into allocation recommendations, and how different modeling choices change the outcome.

Here, we assume you have read the methodology overview. This walkthrough shows the model in action.

Step 1: A single worldview, one allocation

We start with the simplest case: a single worldview with 100% credence and $10M to allocate. This worldview cares about humans across all time horizons (including the far future), is risk-neutral, and places no weight on animal welfare.

Scenario 1

$10M, human-focused longtermist, risk-neutral

View the full worldview inputs

Moral weights
Human life-years	1.0
Human YLDs	0.8
Human income doublings	0.3
Chickens / Birds	0
Fish	0
Shrimp	0
Non-shrimp invertebrates	0
Mammals	0

Time period	Discount factor
0 to 5 years	1.00
5 to 10 years	0.95
10 to 20 years	0.85
20 to 100 years	0.70
100 to 500 years	0.50
500+ years	0.10

Risk profile	Neutral (0)
Discount factor for non-AI risk interventions	0

With these inputs, the model scores each fund. The biosecurity fund scores highest because its modeled effects include averting existential-scale catastrophes, which, when valued across all future generations, produce very large expected values. Even with a 0.10 discount over the 500+-year period, the scale of potential future lives preserved outweighs all near-term effects.

How does the model score a fund?

Each fund's data contains one or more effects (e.g., human life-years saved, disability reduced, income doubled). Each effect has estimated values broken down by time period (0–5 years, 5–10 years, etc.) and risk profile column.

For a single effect, the score is:

score = moral_weight × Σ_t(value[t] × discount_factor[t])

The model looks up the value for the chosen risk profile column and each time period, multiplies by the corresponding discount factor, sums across all time periods, and then multiplies by the moral weight for that effect type. A fund's total score is the sum across all its effects.

For example, if GiveWell has an effect "lives saved" with values [11,706; 910; 325; 65; 0; 0] across the six time periods, and the discount factors are [1.0, 0.95, 0.85, 0.70, 0.50, 0.10], the contribution from that effect alone is:

1.0 × (11706×1.0 + 910×0.95 + 325×0.85 + 65×0.70 + 0 + 0) = 12,895

Add the contributions from its other effects (disability reduction, income improvement), and you get GiveWell's total score of about 20,024 under this worldview.

Funds focused on existential risk reduction have effects in the 500+ year column on the order of 10³² human life-years (representing the value of all future generations if extinction is averted). Even multiplied by a discount of 0.10, this produces scores around 10³¹, which is why those funds dominate under risk-neutral assumptions.

Sentinel Bio $10M (100%)

The entire $10M goes to a single fund: the one with the highest marginal value under this worldview.

Step 2: Diminishing returns kick in

What happens when there is more money to allocate? Each fund has a diminishing returns curve built into the data. The idea is simple: the most impactful opportunities within a fund get funded first. As more money flows in, the next dollar achieves somewhat less. Eventually, a second fund becomes the better marginal use of money.

We keep the exact same worldview but raise the budget to $100M.

Scenario 2

$100M, same worldview

The model allocates in $10M increments. At each step, it recalculates marginal values accounting for how much each fund has already received.

View the round-by-round allocation

Round	$10M →
1	Sentinel Bio
2	Nuclear fund (Longview)
3	Sentinel Bio
4	Nuclear fund (Longview)
5	Sentinel Bio
6	Nuclear fund (Longview)
7	Sentinel Bio
8	Sentinel Bio
9	Nuclear fund (Longview)
10	Sentinel Bio

Sentinel Bio $60M (60%)

Nuclear $40M (40%)

After the first $10M, diminishing returns make the biosecurity fund slightly less attractive. The nuclear safety fund, which has a similar base score, starts receiving allocations. The money alternates between the two as each one's marginal value decreases.

Step 3: Changing the risk profile

So far, we have used a risk-neutral profile: the model simply takes the expected value of each fund's impact distribution. But what if you are skeptical of the most optimistic scenarios?

The "Continuous Upside Skeptical" risk profile downweights the upper tail of each fund's impact distribution. For global catastrophic risk funds, the upside scenarios are where most of the expected value lives. When you discount those, the picture changes dramatically.

What does "Continuous Upside Skeptical" do, exactly?

We could just say "don't count effects above the 99th percentile." Instead, we use a continuous function that starts dropping off at the 97.5th percentile and decays to zero at the 99.9th percentile, to avoid an arbitrary discontinuous jump in which outcomes are and aren't valued.

Specifically, outcomes up to the 97.5th percentile of the impact distribution receive full weight. Above that, the weight decays exponentially, falling to 1% at the 99th percentile and reaching zero by the 99.9th percentile. The formula for the weight between the 97.5th and 99.9th percentile is:

weight = e^{−ln(100) / 1.5 × (p − 97.5)}

The average weight of all values standardizes this weight. In practice, this nearly eliminates the contribution of extremely high-end outcomes while leaving the bulk of the distribution untouched.

For GCR funds, the enormous expected values (10³¹+) come almost entirely from the upper tail. Once those are discounted, the remaining expected value can turn negative. For GiveWell, whose impact distribution is much tighter, the upper tail barely differs from the median, so discounting it has almost no effect on the score.

Scenario 3

$10M, same moral weights and discounts, but "Continuous Upside Skeptical"

View the changed input

Everything is the same as Scenario 1 except:

Risk profile

Continuous Upside Skeptical (7)

This single change flips the sign on the GCR funds. Under this risk profile, the far-future extinction effects become negative (the model interprets the most likely outcome as slightly harmful rather than hugely beneficial). Meanwhile, GiveWell's near-term, evidence-backed effects are largely unaffected.

GiveWell $10M (100%)

View the fund scores: Neutral vs. Continuous Upside Skeptical

Fund	Neutral	Upside Skeptical
Sentinel Bio	1.12 × 10³¹	−5.18 × 10²⁹
Nuclear fund	1.11 × 10³¹	−5.21 × 10²⁹
AI fund	1.10 × 10³⁰	−2.71 × 10³¹
GiveWell	20,024	19,796
LEAF	18,689	16,994

Animal welfare funds score zero because this worldview assigns zero weight to animal effects. GiveWell's score barely changes between risk profiles because its impact distribution is tight (the upper tail is close to the median, so discounting it makes little difference).

Step 4: Merging worldviews

In practice, you might not be fully committed to either risk profile. Perhaps you think there is an 80% chance the upside-skeptical view is right and a 20% chance the risk-neutral view is right. The model handles this by treating each view as a separate worldview with its own credence.

There are different ways to combine these views. We show two aggregation methods here: credence-weighted and Maximize Expected Choiceworthiness (MEC).

Scenario 4a

$10M, credence-weighted aggregation (80% upside skeptical / 20% longtermist)

View the worldview setup

Worldview	Credence	Risk profile	Other inputs
Human-focused longtermist	20%	Neutral (0)	Same as Scenario 1
Human-focused, upside skeptical	80%	Continuous Upside Skeptical (7)	Same moral weights & discounts

With credence-weighted aggregation, each worldview allocates its share of each $10M increment. At 80/20 credence, the upside-skeptical worldview directs $8M, and the longtermist worldview directs $2M.

GiveWell $8M (80%)

Sentinel Bio $2M (20%)

The budget splits in proportion to the credences. Each worldview picks its own top fund independently.

Scenario 4b

$10M, MEC aggregation (same 80/20 worldviews)

MEC (Maximize Expected Choiceworthiness) works differently. Instead of letting each worldview allocate independently, it computes a single weighted-average score across all worldviews and puts the entire increment into whichever fund scores highest overall.

Sentinel Bio $10M (100%)

Despite holding only 20% credence, the longtermist worldview's scores are so astronomically large (10³¹) that even after multiplying by 0.2, they still overwhelm the skeptical worldview's scores (on the order of 10⁴). The 80% credence on the skeptical side makes no practical difference.

Step 5: Merging at a larger scale

Now we scale up to $100M with a 30/70 credence split (30% upside-skeptical, 70% longtermist), using credence-weighted aggregation. Diminishing returns interact with the credence split.

Scenario 5

$100M, credence-weighted, 30% upside skeptical / 70% longtermist

Sentinel Bio $40M (40%)
Nuclear $30M (30%)
GiveWell $30M (30%)

The longtermist worldview (70% credence) splits its $70M between biosecurity and nuclear safety as diminishing returns force diversification. The upside-skeptical worldview (30% credence) sends all of its $30M to GiveWell, which remains the robustly best option under skeptical assumptions even at scale. The result: a portfolio dominated by GCR giving with a substantial near-term hedge, driven by genuine moral uncertainty rather than arbitrary splitting.

Step 6: Adding animal welfare weights

So far, every worldview assigned zero moral weight to animals. Now we include worldviews that value animal welfare. We keep $100M and use credence-weighted aggregation with three worldview groups:

Scenario 6

$100M, credence-weighted, with animal welfare weights

View the worldview setup

Worldview group	Credence	Risk profile	Animal weights
Human-focused longtermist	45%	Neutral (0)	Zero
Animal welfare advocate	40%	Upside Skeptical (7)	Chickens 0.4, Fish 0.3, Shrimp 0.1, Invertebrates 0.05
GiveWell-focused, upside skeptical	15%	Upside Skeptical (7)	Zero

Animal Welfare Funds $40M (40%)

Sentinel Bio $27M (27%)

Nuclear $18M (18%)

GiveWell $15M (15%)

Sentinel Bio (27%)
Animal Welfare funds (40%)
Nuclear (18%)
GiveWell (15%)

The animal welfare advocate's worldviews (40% total credence) direct their entire share to animal welfare funds. Within that share, cage-free campaigns receive the most ($20M), followed by the EA Animal Welfare Fund ($12M) and other animal welfare programs ($8M). The longtermist worldviews (45%) split between biosecurity and nuclear as before. The GiveWell-focused worldview (15%) sends everything to GiveWell.

The result is a genuinely diversified portfolio spanning all three cause areas: GCR reduction (45%), animal welfare (40%), and near-term human welfare (15%), each driven by the credences assigned to the underlying worldviews.

Summary

These scenarios illustrate the core mechanics of the model:

Scoring

Each fund receives a score based on its modeled effects, multiplied by the worldview's moral weights and discount factors, then collapsed according to the chosen risk profile.

Diminishing returns

The model allocates in increments. Each increment goes to whichever fund has the highest marginal value at that point. As a fund receives more, its marginal value declines, eventually allowing other funds to compete.

Risk profiles

The choice of risk profile can completely change which funds score highest. Risk-neutral views tend to favor GCR funds (whose upside is enormous but uncertain); upside-skeptical views tend to favor GiveWell (whose impact is smaller but more certain).

Worldview aggregation

When you hold multiple worldviews, the aggregation method matters. Credence-weighted aggregation gives each worldview proportional control over the allocation. Other methods, like MEC, combine scores into a single average, which can allow worldviews with extreme value estimates to dominate even when credences are low.

All numbers in this document were generated by the Donor Compass Advanced Mode model using the March 31, 2026, dataset. The primary aggregation method shown is credence-weighted; eight other methods are available and can produce different results when multiple worldviews disagree.