The mechanism: a Scale-Up Fidelity Stress Test (SFST) — a short, standardized, publicly scored instrument that a government or implementing agency must complete and register before a donor releases financing to scale a pilot program past a defined budget threshold. It is not another readiness checklist. It is built the way credit-risk scoring was built: retroactively, from actual outcomes, so its thresholds are calibrated against real voltage-drop cases rather than expert intuition.

Concretely: a consortium of the organizations that already hold this data — J-PAL, the Center for Global Development, 3ie, and MERL teams at major funders — pools existing pilot-to-scale-up pairs where both an original effect size and a post-scale-up effect size exist (Kenya's Teacher Internship Programme, TaRL's multi-country rollouts, Action Schools!

BC, and comparable health and cash-transfer scale-ups).

For each case, it also codes measurable, checkable pre-scale-up capacity indicators: civil-service supervisory visit frequency, frontline staff turnover rate in the prior 12 months, whether hiring/firing authority sits with the original NGO or transfers to government, and the agency's track record on any prior scale-up in the same sector.

That paired dataset lets you fit a model of effect-size retention as a function of capacity indicators instead of guessing.

The output is a short instrument — target under 20 measurable items, completable from administrative data plus a site-visit checklist, no self-reported 'readiness' scores — that produces a single retention-probability estimate: given this program's original effect size and this implementer's measured capacity profile, what fraction of the effect is likely to survive at scale, with a confidence interval.

Donors (GPE, Gavi, World Bank scale-up loans, bilateral aid agencies) adopt it as a disclosure requirement, not a veto: financing can still proceed on a low score, but the projected benefit-cost ratio used to justify the budget must use the SFST-adjusted effect size, not the original trial's, and the score gets registered publicly alongside the funding decision so it can be checked against outcomes later.

This does two things existing frameworks don't. First, it forces the conversation about implementation capacity to happen with numbers attached before the money moves, rather than as a post-hoc explanation for a failed evaluation. Second, because every registered score gets checked against the eventual outcome, the instrument keeps improving — it's a living model, not a one-time checklist, and its calibration data becomes a public good the next country can use.