Test whether Teaching at the Right Level (TaRL), one of the most rigorously evaluated pedagogies in development economics, retains its measured effect size when the delivery channel shifts from NGO implementation to government implementation, and identify why the mechanism breaks.

Compared J-PAL/Pratham randomized evaluations of TaRL across six RCTs in seven Indian states against Bold, Kimenyi, Mwabu, Ng'ang'a and Sandefur's randomized 'voltage drop' study embedded in Kenya's nationwide Teacher Internship Programme, which isolates implementation channel (NGO-hired vs. government-hired teachers on identical contracts) as the sole varying factor.

Cross-checked against a PMC-indexed cluster-RCT of Action Schools! BC in Canada to test whether the attenuation pattern is context-specific or general to pilot-to-government scaling. Effect sizes kept in original reported units (years-of-learning-per-30-hours for TaRL; SD test-score gains for Kenya) rather than force-converted, given data limitations noted below.

I've spent three years half-listening to development economists tell me randomized trials "prove" an intervention works, and I keep waiting for someone to ask the follow-up question: works implemented by whom? Teaching at the Right Level is the case study everyone cites and almost no one interrogates on this axis, so let's interrogate it.

The topline number is genuinely extraordinary. 85 years of learning. That is one of the largest effect sizes ever rigorously measured in education research, and it is why TaRL has since reached a claimed 80 million children across India and Africa. I am not disputing the effect.

I am disputing the implicit assumption that "effective in a Pratham-run pilot" and "effective once a government ministry owns delivery" are the same claim, because a separate, cleaner experiment says they are not.

That experiment is Bold, Kimenyi, Mwabu, Ng'ang'a and Sandefur (CGD Working Paper 321, 2013; Journal of Public Economics, 2018). It isn't about TaRL specifically, but it is the single best natural test of the mechanism TaRL depends on, because it holds the intervention design constant and varies only who runs it.

Kenya's government took an NGO-piloted contract-teacher reform — extra teachers on short-term local contracts, previously shown to raise test scores — and scaled it nationwide through two parallel channels: one where an international NGO handled hiring and management, one where the Ministry of Education did, with teachers on functionally identical contracts.

The NGO arm replicated the original positive effect. The government arm produced a measured effect of approximately zero. Same contract. Same teachers, on paper. Different institution running the plumbing. The authors call this a "voltage drop," and it isn't Kenya being uniquely dysfunctional — a cluster-RCT of Action Schools!

BC in Canada, indexed in PMC, documents the identical pattern in a rich-country school-health program scaled from pilot to province: effects attenuate specifically at the handoff to routine institutional delivery, not because the underlying pedagogy stopped working.

Layer in the denominator problem and this gets worse, not better. The World Bank's 2022 learning-poverty update put the share of 10-year-olds in low- and middle-income countries unable to read and understand a simple age-appropriate text at roughly 70%, up from about 57% before the pandemic.

That is the population TaRL is being scaled to serve, through the exact government channels the Kenya evidence says drop voltage.

Isabel Vargas at COLMEX pointed out to me, not gently, that I was citing "80 million children reached" as if reach were the same variable as effect, and she is right and I was sloppy — reach is an implementation-fidelity claim, not an outcome claim, and the paper trail on fidelity at full government scale, as opposed to Pratham-supported hybrid delivery, is thinner than the topline number suggests.

None of this means abandon TaRL. It means the honest confidence interval on "TaRL at national government scale" is wider than the confidence interval on "TaRL run by Pratham," and any country ministry citing the 1.85-years figure as their own expected return is citing someone else's implementation capacity, not their own.