Analysis 01Shallow evaluation · draft v0.2 · 5 September 2026
Should a philanthropist fund child nutrition through South Africa’s crèches?
A draft funding assessment of child nutrition through South Africa’s early learning programmes. It brings published research and an interactive cost-effectiveness model alongside questions raised by my work at Ubhejane and Mzomoyethu Crèches. Model inputs and source checks still awaiting review are identified below.
Question
Fund, don’t fund, or narrow the thesis
Unit of account
GiveWell units of value, against GiveWell’s benchmark and its 10x bar
Effort
Shallow review: days, not weeks; no expert interviews yet
Author
Rory Clark, with the AI workflow logged in Appendix B
In a nutshell
The age window matters
Crèche gardens do not fix stunting in the children who attend. Stunting is largely set by age two, and the one randomised trial of a crèche-platform model found no growth effect on enrolled preschoolers.
Reach beyond the classroom
That trial found a 17-point fall in stunting among the enrolled children’s younger siblings, reached through a parenting component for infants’ caregivers. The crèche works as a door to the under-twos and the household, not as a feeding point for four-year-olds.
The platform costs too much
At mid-range assumptions the platform thesis returns about 0.2 times GiveWell’s benchmark and clears the 10x bar in none of 2,000 draws. Under this model it fails on cost alone: it would need to cost about one dollar per child-year.
Public funding has a cost
The narrower idea, paying for registration so that unregistered centres receive the subsidy already budgeted, looked like 5x in the first draft. Charging the public money it unlocks at any plausible shadow value, and pricing registration as infrastructure rather than paperwork, takes it to about 0.1x at mid-range, with a sign that flips on what a Treasury rand is worth. It is an ECD-access thesis, not a nutrition thesis.
Test the narrower options first
Recommendation: do not fund either as a nutrition fund. Score the cheap uses of the platform first: vitamin A, deworming and micronutrient powders through 42,000 centres. If the crèche is still the idea, frame it as ECD access with a nutrition standard attached, and value school readiness and maternal employment, which this evaluation does not.
1 · The thesis
As its proponents would state it, and as I once stated it.
Early learning programmes, the crèches, reach 1.66 million South African children every weekday. Give each centre a productive food garden, a nutrition standard for the meals it serves, training for the practitioners, and a growth chart on the wall, and you have a delivery platform for child nutrition that the state already pays to keep open. Stunting falls, diets improve, and the garden keeps producing after the project leaves.
I would have pitched it that way in 2019, and did, in the Mandlakazi proposal. The three crèche gardens we built near Manyoni were this model. The question this memo asks is whether a philanthropist should put serious money behind it, and the honest answer needs the evidence rather than the memory of the onion harvest.
2 · The problem
Big, unchanged, and with money on the table that did not move.
Children under five who are stunted
more than 1.5 million, 27% in the 2016 survey, unchanged for two decades
Early learning programmes counted in 2021
42,420, enrolling 1,660,316 children, about 39 a programme
Registration status in 2021 (Department of Social Development definition)
40% registered, 16% in process, 42% unregistered; about one third receiving the subsidy
Subsidy per child per day
R17 until 2025; R24 from Budget 2025, the first rise since 2019; cost-modelled target about R32; paid at silver level under the bronze, silver and gold system that replaced the 2021 definition
New money, Budget 2025
R10 billion added, to take subsidised children from 800,000 to 1.5 million by 2027/28
Implementation, late 2025
only 30% of registered centres receive the subsidy (Daily Maverick, 21 October 2025); the R10 billion had not reached centres (Daily News, 11 November 2025)
Budget 2026
about 100,000 children added against the 700,000 planned; the 2026 subsidy allocation cut by R800 million, from R5.1 billion to R4.3 billion, with slow spending given as the reason and no rollover (Ilifa Labantwana; Legal Resources Centre)
Sources: SADHS 2016 via Children Count; ECD Census 2021; Parliament and Ilifa Labantwana on Budget 2025; Daily Maverick, October 2025; Daily News, November 2025; Ilifa Labantwana and the Legal Resources Centre on Budget 2026. “Registered” meant Department of Social Development registration in 2021 and means bronze, silver or gold under Bana Pele now, which is why “one third subsidised” and “30% of registered centres” are not the same denominator.
3 · The evidence
One trial of a crèche-platform model, and what the rest of the literature adds.
Study
Design
Finding
Gelli et al. 2018, Journal of Nutrition (NEEP-IE, Malawi)
Cluster RCT, 60 community childcare centres, 1,248 preschoolers aged 3 to 6 and 304 younger siblings aged 6 to 24 months; one year; the centre as the entry point for household food production, nutrition behaviour change, and parenting groups for caregivers of infants.
No anthropometric effect on the enrolled children. Younger siblings: height-for-age z-score up 0.44, stunting down 17 percentage points (P below 0.05), a secondary outcome in about 150 children per arm.
Gelli et al. 2020, Journal of Nutrition (same trial)
Dietary adequacy analysis of the preschoolers: eleven micronutrients, multipass 24-hour recall.
Mean probability of adequacy of nutrient intake up 5 percentage points (standard error 1).
Gelli, Kemp, Margolies et al. 2022, Food Security (economic evaluation of NEEP-IE)
Micro-costing and expenditure analysis, societal perspective, 12-month costing horizon; deaths and stunting cases averted via the Lives Saved Tool; 10% discount rate.
Total cost $197,377 for 4,806 beneficiaries: $41 per beneficiary, $595 per stunting case averted, $18,310 per death averted, $516 per DALY averted. Benefit-cost ratio 3.6 to 24.7, depending on the value placed on a life.
Olney et al. 2015, Journal of Nutrition (Helen Keller International, Burkina Faso)
Cluster RCT, 55 villages; homestead food production with health-committee delivery.
Diarrhoea down 15.9 points in children aged 3 to 12.9 months; wasting down 8.8 points (P = 0.08); anaemia down 14.6 points in the 3 to 5.9-month subgroup (P = 0.03).
Masset et al. 2012, BMJ
Systematic review, 23 studies, mostly home gardens.
Positive effect on production of the promoted foods, none on household income; very little evidence on stunting, wasting or underweight; some vitamin A effect, none on iron.
Wendt et al. 2025, Maternal and Child Nutrition (FAARM, Bangladesh)
Cluster RCT, 96 settlements, three years.
No reduction in anaemia or micronutrient deficiencies in women or children.
Blakstad et al. 2021 and 2022, Tanzania
Cluster RCT of home gardens, 1,006 women, followed to year three.
Dietary diversity and food security improved at one year; the dietary diversity gain was gone by year three.
Waddington et al. 2014, Campbell Collaboration
Systematic review of Farmer Field Schools, 92 evaluations, 15 of usable quality.
Yields up 13% and income up 20% in pilots; no positive outcomes in national-scale programmes evaluated more than two years on, with facilitator recruitment the stated problem.
Thrive by Five Index 2021 and 2024, South Africa
National survey of children aged 50 to 59 months in early learning programmes.
5.65% stunted (2021, revised) and 7% moderately or severely stunted with 25% mildly stunted (2024), against about 16% for the same age nationally and 27% of all under-fives (SADHS 2016). Enrolled children are the less-stunted children; stunted children in programmes are five months behind their peers.
Nourishing Beginnings 2025, IJERPH (Makhanda)
Mixed methods, 8 centres, 307 children.
Three meals offered but starch-dominated; cost-driven substitutions; little protein, dairy or vitamin-A food.
Read together: the platform can change what young children eat, and in Malawi it changed the growth of the under-twos in the household, through a component that spoke to their caregivers directly. It did not change the growth of the children who sat in the centre. Home-garden effects elsewhere are modest, often fade, and have not shown up as income. The South African numbers add a selection problem: the children in programmes are the less-stunted children of their age.
4 · The costs
Programme costs and working estimates.
The Malawi economic evaluation is the only costing of the centre-platform model itself: $41 per beneficiary for the year, $595 per case of stunting averted and $516 per DALY averted, with a benefit-cost ratio between 3.6 and 24.7 depending on the value placed on a life. Against GiveWell’s roughly $50 to $100 per DALY, that is five to ten times too expensive before any discount for transfer to South Africa, where stunting is lower and delivery dearer. Comparable homestead programmes cost more: a poultry-based programme in Burkina Faso cost about $796 per household-year; other nutrition-sensitive agriculture programmes are costed at $132 to $160 per beneficiary-year. The state subsidy of R24 per child per day is worth roughly $330 a child-year, of which only part is food on the plate.
Author note · to complete
Per-child and per-centre costs from the Mandlakazi records (three crèches, 41 household baseline records, 2019 to 2022) and the RVCC close-out go here, with the cost slider in the model re-set to them and its flag flipped from “AI-proposed” to “author-set”.
5 · The model
Move the assumptions. Watch the verdict.
A back-of-envelope model in GiveWell’s unit of value: one unit is one person’s consumption doubled for a year, the benchmark gives about three units per $1,000, and GiveWell funds at about ten times the benchmark. Two theses share the yardstick. Every parameter has a range, a source and a flag saying who set it. The tornado shows which assumptions the answer rests on, and the Monte Carlo band shows how wide the honest answer is. One caution on the yardstick: GiveWell called this benchmark “cash” until November 2025, and since October 2024 it has scored GiveDirectly’s own programme at three to four times it, so beating the benchmark is not beating GiveDirectly.
This is version 0.2. Version 0.1 scored Thesis B at about 5x by counting the state subsidy a registration unlocks as pure benefit and pricing registration as paperwork. The red-team pass in Appendix B caught both. B now carries a charge for the public money it displaces, a success probability, and an acceleration rather than a creation of subsidy years.
Multiple of the benchmark0.21xbar 10x
Assumptions
$15$120
AI-proposed · author review pending The Malawi NEEP-IE programme cost $41 per beneficiary for its year (Gelli, Kemp, Margolies et al. 2022), with a quarter contributed in kind by communities. Placeholder until the author enters per-child costs from the Mandlakazi and RVCC records.
1.0 year3 years
AI-proposed · author review pending Mandlakazi ran about two and a half years; NEEP-IE one. The post-exit food tail is credited once per programme, not once per year.
0.050.40
AI-proposed · author review pending NEEP-IE sampled 304 younger siblings aged 6 to 24 months alongside 1,248 preschoolers, a ratio of 0.24, and reached their caregivers through parenting groups. That is a sampling ratio from a trial with an explicit infant component, not a reach rate a centre garden produces; South African fertility is lower, so the mid is set below it.
2 pp17 pp
AI-proposed · author review pending NEEP-IE: stunting down 17 points among younger siblings (P < 0.05), a secondary outcome in about 150 children per arm. Mid value discounts for a single trial, transfer to South Africa, and scale.
0.0 pp2 pp
AI-proposed · author review pending NEEP-IE found no anthropometric effect on the enrolled preschoolers; stunting is largely set by age two.
3.0%25%
AI-proposed · author review pending Galasso and Wagstaff 2019: the 9 to 10% Sub-Saharan penalty is a population aggregate; per stunted person their inputs (about 1.6 fewer school years, a 0.625 SD cognitive deficit) imply 20 to 25%, before any replicability discount and before South African youth unemployment near 50% halves what reaches consumption.
30 years40 years
AI-proposed · author review pending Convention; South African working life from about 18 to 55.
16 years20 years
AI-proposed · author review pending Child reached at 6 to 24 months; earning from the late teens.
3.0%5.0%
AI-proposed · author review pending GiveWell convention of 4% a year.
0.000.60
AI-proposed · author review pending South African under-five mortality is about 33 per 1,000, a third of it neonatal, leaving perhaps 7 per 1,000 between one and five. With Olofin et al. 2013 hazard ratios for moderate stunting, a fully causal reading gives about 0.4 per 100 cases; marginal cases near the threshold carry less. The Malawi evaluation’s LiST modelling implies about 3.3 per 100 in a far higher-mortality setting.
$0.50$25
AI-proposed · author review pending The Social Employment Fund reported 38.6 tonnes of produce across 1,995 gardens: about 19 kg a garden, roughly $0.55 a child at 39 children. Centre output also displaces food the centre would have bought. The author’s Ubhejane yield records will replace this.
$600$1,300
AI-proposed · author review pending Per-person consumption of bottom-two-quintile households with young children, grants included: roughly R1,200 to R2,000 a month, at about R18 per dollar. The food poverty line (about $530 a year) is a floor, not the mean, and a low baseline flatters both theses.
0.0 years2.0 years
AI-proposed · author review pending Tanzania home-garden dietary diversity gains were gone by year three (Blakstad et al. 2022). The Farmer Field School review finds an absence of evidence at national scale rather than measured fade-out. Credited once per programme.
40%100%
AI-proposed · author review pending Counterfactual and credit share: gardens and behaviour change at these centres would mostly not happen without the fund, so the counterfactual is close to one; credit is shared with co-funders. Government spending on practitioners and the subsidy exists either way and is not a reason to discount.
At these settings, per enrolled child-year, the thesis falls below the benchmark. The bar is about 10x the benchmark.
Multiple of the benchmark0.21xMonte Carlo middle 80%: 0.14x to 0.64x · 0% of draws clear the bar
Cost per unit of value$1,567one person’s consumption doubled for a year
Units per $1,0000.64the benchmark gives 3.0
Open Philanthropy bridge$4,477per DALY-equivalent, converting the consumption result at Open Philanthropy’s exchange rate
Stunting averted among under-twos0.016 units
Food and diet for enrolled children0.007 units
Stunting averted among enrolled children0.006 units
What would have to be true
To reach the bar with everything else as set, the philanthropic cost per enrolled child-year would have to be $0.96. It is set to $45.
To merely match the benchmark it would have to be $9.6.
With 2,000 draws across the stated ranges, the median is 0.29x, the middle 80% runs from 0.14x to 0.64x, 2% of draws beat the benchmark and 0% clear the bar. The width is the finding.
Which assumptions move the answer. Each bar swings one assumption from its low to its high with the rest as set. Clay runs toward worse than the current setting, green toward better. Log scale; results at or below zero sit at the left edge.Table view: every assumption and its effect
Assumption
Low
Now
High
Multiple at low
Multiple at high
Set by
Philanthropic cost per enrolled child-year
$15
$45
$120
0.64x
0.08x
AI-proposed, review pending
Adult income gain per stunting case averted
3.0%
12%
25%
0.11x
0.35x
AI-proposed, review pending
Under-twos reached per enrolled child
0.05
0.15
0.40
0.14x
0.41x
AI-proposed, review pending
Stunting reduction among reached under-twos
2 pp
8 pp
17 pp
0.13x
0.34x
AI-proposed, review pending
Share of the benefit the fund can claim
40%
70%
100%
0.12x
0.30x
AI-proposed, review pending
Consumption value of extra food per enrolled child-year
$0.50
$5
$25
0.17x
0.41x
AI-proposed, review pending
Stunting reduction among enrolled children
0.0 pp
0.5 pp
2 pp
0.16x
0.36x
AI-proposed, review pending
Discount rate
3.0%
4.0%
5.0%
0.26x
0.18x
AI-proposed, review pending
Under-five deaths averted per 100 stunting cases averted
0.00
0.25
0.60
0.19x
0.25x
AI-proposed, review pending
Baseline consumption per person in target households
$600
$950
$1,300
0.24x
0.20x
AI-proposed, review pending
Years the food effect persists after the programme ends
0.0 years
0.5 years
2.0 years
0.20x
0.24x
AI-proposed, review pending
Years until the child starts earning
16 years
17 years
20 years
0.22x
0.20x
AI-proposed, review pending
Working years over which the gain accrues
30 years
35 years
40 years
0.20x
0.22x
AI-proposed, review pending
Years the programme runs at a centre
1.0 year
2.0 years
3 years
0.22x
0.21x
AI-proposed, review pending
Benchmark: about 0.003 units per dollar, the cost of doubling consumption for someone at the $2.15 a day line. GiveWell called this "cash" until November 2025; since October 2024 it has scored GiveDirectly’s own programme at roughly 3 to 4 times this benchmark.
Bar: about ten times the benchmark, GiveWell criteria. About $33 per unit.
Under-five death averted: 116 units, GiveWell moral weights, within an age curve that peaks at 134 for ages 5 to 9.
Open Philanthropy bridge: a DALY at about $100,000 and a doubled-consumption year at about $35,000. An exchange rate for a consumption result, not a health measurement.
Ranges are hand-set judgements, not calibrated intervals; the Monte Carlo draws each parameter from a triangle between its low and high with your setting as the mode, and treats parameters as independent.
6 · Who else is working on this
A crowded field, and a cheaper thesis nobody in it has scored.
Government: the Department of Basic Education, which took over early childhood development in 2022, runs the subsidy, the registration drive and the Bana Pele programme. Philanthropy: Ilifa Labantwana (ELMA, UBS Optimus and the DG Murray Trust), the DG Murray Trust itself, Innovation Edge, the Do More Foundation, SmartStart and Grow Great, with the South African Medical Research Council on the evidence. A November 2023 synthesis report on nutrition support for ECD and a public call for a national early childhood feeding scheme set the policy agenda. GiveWell’s livelihoods team, formed in 2025, is looking at agriculture and graduation programmes but has recommended nothing in this space. Renaissance Philanthropy’s own nutrition work sits in cardiovascular nutrition and crop resilience.
The gap is not gardens, meals or training, all of which have owners. Two things do not. First, the cheapest uses of a platform that holds 1.66 million two-to-five-year-olds every weekday: vitamin A supplementation, deworming and micronutrient powders, each a few dollars per child-year in GiveWell’s programme reviews and none of them scored for South African centres. Call that Thesis C; it is the one a fund leader should be asked about first. Second, the pipeline between budgeted subsidy and an unregistered centre, which is worth pursuing as an ECD-access thesis, valued for school readiness, fee relief and mothers’ employment, and not as a nutrition fund.
7 · The cruxes
Seven questions that decide the recommendation.
The 1,000-day window
Centres enrol two-to-five-year-olds; stunting is largely set by age two. In Malawi the platform worked through younger siblings, reached by parenting groups for their caregivers, not through the children in the room.
Parameter: under-twos reached per enrolled child. At mid-range, zero takes A from 0.21x to about 0.1x; the trial’s own ratio with its full 17-point effect takes A to about 0.75x. No setting of this channel reaches the bar.
Selection
Enrolled four-year-olds are 6 to 7% stunted against about 16% for their age nationally. The more stunted children are outside the programmes, in unregistered or no centres.
The claim that unregistered centres hold the more stunted children is asserted, not measured. A survey of stunting by registration status would settle it; without one, a fund that improves subsidised centres reaches the already-reached.
The binding constraint
Is it money, the registration and subsidy pipeline, or meal quality? Money exists on paper and was clawed back for not flowing; a third of centres are covered; meals are starch.
Parameters: cost per child registered and the share of subsidy that becomes consumption. B at $5 a child with 90% reaching poor people is a different thesis from B at $430 a child with 30%. The first is paperwork; the second is building compliance, and the evidence says the second is what unregistered centres face.
The shadow price of public money
B’s benefit is the state subsidy it causes to flow. That money is not free: the 2026 Budget returned R800 million of it to general spending.
Parameter: value of the displaced public rand. B’s sign flips at about 0.28 times the benchmark; at zero it is 0.8x, at one it destroys value. A donor who believes Treasury money is worth nothing to the poor can back B; nobody else can, on this model.
Cost, all in
Practitioner time, water, fencing, inputs, facilitation and follow-up. Malawi cost $41 per beneficiary-year; Burkina Faso’s poultry programme $796 per household-year.
Parameter: philanthropic cost per enrolled child-year. At mid-range effects, A reaches the bar only below about $1 and matches the benchmark only below about $10. Nothing that delivers gardens and behaviour change costs a dollar a child-year, so A fails on cost before the effect sizes are argued.
Persistence and quality
Home-garden dietary gains in Tanzania were gone by year three; national-scale training programmes show no measured effect two years on.
Parameter: years the food effect persists after exit. Even at two years A moves from 0.21x to about 0.23x, because the food channel is small once centre output is valued at what the Social Employment Fund gardens actually produced. The Ubhejane record is the author’s own test of persistence.
Crowdedness and the real gap
Ilifa Labantwana, the DG Murray Trust, Innovation Edge, SmartStart and Grow Great occupy nutrition, registration support and advocacy. The cheapest uses of a platform that holds 1.66 million two-to-five-year-olds have not been scored here at all.
If vitamin A supplementation, deworming and micronutrient powders through centres cost a few dollars per child-year, as GiveWell’s programme reviews suggest, that Thesis C dominates both A and B and should be the first thing a fund leader is asked about.
8 · Verdict
Don’t fund either as a nutrition fund. Score Thesis C first.
Thesis A. As a stunting intervention for enrolled children the platform thesis fails on the evidence: no growth effect, wrong age window, and a cost that puts it at about a fifth of GiveWell’s benchmark at mid-range assumptions, with no draw in two thousand clearing the bar. It would need to cost about a dollar per child-year. Narrowed to the crèche as a door to the under-twos it becomes plausible in principle, but the trial that showed the sibling effect did so with a parenting component the platform thesis does not describe or cost.
Thesis B. The first draft scored it near 5x. It did so by counting state subsidy as benefit and charging nothing for it. Charge the displaced public money at any plausible value, price registration as compliance rather than paperwork, and treat the fund’s effect as bringing the subsidy forward rather than creating it, and B sits near 0.1x with a sign that flips on the shadow value of a Treasury rand. It is not a nutrition thesis. It may be a good ECD-access thesis; this evaluation has not valued school readiness or mothers’ employment, and a fund leader who wants B should be asked to.
Team, technique and track record. Not assessable for a fund that does not yet exist. What I would look for: a leader who has moved a centre from bronze to silver and can say what it cost; a pre-registered endpoint at 24 months; a costing of vitamin A, deworming and micronutrient powders through the same centres before any garden is planted.
What would change my mind. A South African trial showing under-two growth effects through the centre platform; a costing near a dollar per child-year; evidence that Treasury money returned from ECD is worth close to nothing to poor households, which would rescue B; or a Thesis C costing that turns out worse than either.
9 · For the fund leader
Ten questions I would ask before the second meeting.
Which children do you reach who are not already in a subsidised centre, and how do you know?
How do you reach the under-twos, and what did that cost per household in your last programme?
What did your gardens look like 24 months after exit? Show me the site visits, not the closeout report.
What is your all-in cost per enrolled child-year, including practitioner time, water and follow-up?
If the subsidy pipeline is the binding constraint, why are you buying gardens? And if it is the pipeline, what does a bronze-to-silver upgrade cost per centre?
What result would make you stop?
Which endpoint would you pre-register: height-for-age at 24 months, dietary diversity, or attendance?
The 2025 plan was 700,000 more subsidised children by 2027/28; about 100,000 were added and the 2026 Budget cut R800 million. What does that do to your model?
Who owns the growth-monitoring data, and who acts on a child who is falling off the curve?
Have you costed vitamin A, deworming and micronutrient powders through the same centres, and why is your thesis better than that?
10 · Field reality
Questions for field follow-up.
Author note · to write from the record
Water. What a crèche garden needs through a KwaZulu-Natal dry season, who carries it, and what the holding pond at Gudlintaba actually held.
The practitioner. Who tends the garden at seven in the morning, what she is paid, and what happens when she leaves.
The committee. Which decisions took three months, and which took one conversation with the right person.
After the contract. Garden beds, water storage and protective shadecloth shown in the June 2022 aerial of Ubhejane.
The Social Employment Fund number. Why 38.6 tonnes across 1,995 gardens is the right number to be angry about, and what it says about wage programmes that grow food as a side effect.
Renaissance Philanthropy’s Africa Jobs Fund, launched in May 2026, sets out to double the lifetime incomes of at least 250,000 people with $100 million over five years, at under $10 of philanthropic subsidy per doubled-consumption year against its own stated bar of $40. Its published method discounts each deal for counterfactual impact, partial attribution, probability of success and timing; the deal-level inputs are not public, and in a real evaluation I would ask for the spreadsheet. Dividing the headline targets gives about ten million person-years and about $400 of philanthropic cost per person durably doubled. The rows below apply the same formula as above, cost per doubled-consumption year equal to cost per worker divided by attribution, counterfactual, doublings, persistence and venture success, to show which inputs the headline depends on. The first row is what the targets imply if no discounts were applied; it is a reconstruction, not the fund’s claim.
Scenario
Assumptions
Cost per doubled-consumption year
Mobility, if no discounts were applied
$400 per worker, full attribution, 13 years
about $10 per doubled-consumption year, roughly 33x the benchmark
Mobility, optimistic
$1,500 per worker, half attributed, 10 years, 60% venture success
about $290, roughly 1.1x
Mobility, two-year Gulf contract
$3,000 per worker, 30% attributed, 2 years
about $10,000, roughly 0.03x
Manufacturing, Kenya EPZ type
$1,000 per job, 30% attributed, 8 years
about $990, roughly 0.3x
Manufacturing, if jobs resemble Ethiopia’s industrial parks
$2,000 per job, one year of tenure
about $89,000, near zero
The result is dominated by leverage, attribution and persistence, not by the well-evidenced wage gap. Three cruxes for the fund’s evaluators: what share of the gain a minority catalytic investor can claim once competitors and government corridors are netted out; how long the income lift lasts when factory tenure in the region is measured in months and most mobility contracts in one or two years; and whose consumption doubles once living costs abroad and remittance shares are counted.
Appendix B
How AI was used, what it got wrong, and what the red team changed.
Research. AI-assisted research on 5 September 2026 covered philanthropic fund models, evaluation standards and the evidence for child nutrition through crèches. Source access was limited to search excerpts and abstracts for several publishers and organisations; the full texts were not reviewed.
Citation review. A second review checked about 45 claims against available abstracts and search excerpts. It corrected an author attribution and publication year, narrowed a claim about the Tanzania follow-up, and added the Malawi cost estimates. Full-text verification remains necessary.
Critical review. An independent AI review produced 28 critiques, listed below. The main correction was to include the public cost of subsidies unlocked by centre registration. With that cost, registration costs and the probability of success included, the model’s mid-range estimate for Thesis B fell from about 5x to about 0.1x the benchmark.
Earlier errors. A cited GiveWell blog post could not be verified and was removed. The review also found differing statements of Open Philanthropy’s benchmark on the source page; that assumption remains listed for verification.
Model status. Claude assisted with drafting the model and subsequent revisions. Its assumptions remain marked AI-proposed pending author review. A calculation-check script accompanies the model in the repository.
All 28 red-team critiques and what was done
Critique
Severity
Label
What was done
Thesis B treated the public money it unlocks as free
high
changed
Added a shadow-value charge for displaced public spending; B moved from about 5x to about 0.1x at mid-range.
B priced registration as paperwork, not bronze-to-silver compliance
high
changed
Cost range widened to $5 to $430 per child, mid $60.
B had no counterfactual structure
high
changed
Replaced one attribution slider with years brought forward, success probability and credit share.
“Cash” is no longer GiveDirectly as GiveWell scores it
high
changed
Benchmark note added; verdict wording changed; GiveDirectly now scores 3 to 4x the benchmark.
The largest ECD channels are missing
high
noted
School readiness and maternal employment named as omitted; B relabelled an ECD-access thesis. Not modelled.
The Malawi economic evaluation’s numbers were available
high
changed
Entered: $41 per beneficiary, $595 per case, $516 per DALY, ratio 3.6 to 24.7.
NEEP-IE is not “this exact model”
high
changed
Parenting component described; under-two reach re-sourced as a sampling ratio and lowered.
The $16 flip condition was the benchmark, not the bar
medium-high
changed
Crux 5 and verdict now use the bar cost, about $1 per child-year.
Attribution logic inverted between theses
medium-high
changed
A’s counterfactual raised to 0.7; B charged for the money it causes.
Baseline consumption set at the poverty floor
medium-high
changed
Baseline raised to bottom-two-quintile consumption, mid $950.
Income penalty misread as a per-person effect
medium
changed
Range widened, source rewritten, employment caveat added.
Selection compared unlike age groups
medium
changed
Age-matched 16% figure added beside the 27% under-five figure.
Policy picture a year stale
medium-high
changed
2026 Budget row added after verifying the R800 million cut and the 100,000 figure.
Food channel contradicted its own source by 25 times
medium
changed
Range cut to $0.50 to $25 per child-year, mid $5, with displacement noted.
Mortality high end indefensible for South Africa
medium-low
changed
Range cut to 0 to 0.6 deaths per 100 cases, mid 0.25, derivation shown.
B’s basis should be the centre, not the child
medium
noted
Kept per child for readability; noted that cost and benefit both scale with the centre.
Subsidy increment mis-specified
medium
changed
Replaced with a share of subsidy reaching poor people: meals, fee relief, wages.
No Thesis C for the platform’s cheapest uses
medium
noted
Vitamin A, deworming and micronutrient powders added to the landscape, verdict and questions. Not modelled.
Every prior is AI-proposed and the field layer is empty
high
noted
True. This is the author’s work before submission.
Appendix A attributed assumptions the fund does not make
medium-high
changed
Row relabelled, the fund’s own discount method and $40 bar stated, Ethiopia row reworded.
Most cruxes failed the page’s own definition of a crux
medium
changed
Each crux now names a parameter and the value at which the verdict moves.
Monte Carlo band uncalibrated and moving with the reader
medium-low
partly
Share of draws clearing the bar now printed; ranges labelled as judgements; mode still follows the slider by design.
Persistence credited once per year rather than once per exit
low
changed
Programme length added; the tail is credited once.
Log, page and model disagreed on their own numbers
medium
changed
Check script committed to the repository; numbers reconciled after the revision.
“Cost per DALY-equivalent” mislabels a consumption result
low-medium
changed
Tile relabelled as the Open Philanthropy bridge; Malawi’s own $516 per DALY shown separately.
Persistence and income citations stretched
low
changed
Sources re-quoted to the abstracts.
Income channel assumes a labour market South Africa lacks
low-medium
noted
Employment caveat in the source line; no separate parameter.
“One third” and “30%” not reconciled
low-medium
changed
Registration definitions stated in the problem table.
Still to verify or complete
The source review relied on search excerpts and abstracts; full articles and several publisher and news sites were not accessible. Check the original publications before relying on the findings.
The cost to move a centre from bronze to silver registration has no published figure; the $55 to $430 per child range is a construction from R40,000 to R300,000 a centre.
Olofin et al.’s hazard ratio for moderate stunting was not retrievable; only the severe figure (5.48) is confirmed.
Whether Open Philanthropy’s practical bar is nearer $50 or $100 per DALY.
Exchange rate and the consumption baseline for bottom-two-quintile households with young children.
