The civ-seeding programme has spent four runs failing to tell two kinds of contagion apart. It works on the 43 Karnataka villages where a lender introduced microfinance through village leaders (Banerjee, Chandrasekhar, Duflo and Jackson, 2013). A simple contagion spreads like news: one informed contact is enough. A complex one needs reinforcement from several adopters. On synthetic villages where we know the answer, the estimation picks the right model only 37 to 82% of the time out of sample. That matters only if the two models give different advice about whom a rollout should seed first. So this run tested whether they do.
We took the three fitted models (simple information spread, a fractional threshold where a household adopts once half its contacts have, and a reinforcement model) and scored ten seeding rules under each one, on the real networks. Every rule got each village's own leader count as its budget, 1,140 seeds in all. The rules were: the lender's actual leaders, random households, the best-connected households, two centrality scores, one clustered patch of neighbours, patches of four, and greedy search under each model. Greedy search was done on one set of random draws and scored on another, so its score carries no search optimism. We wrote the hypotheses and our predictions down before the run.
The advice barely moved between models. The seven rules that don't depend on a model ranked almost identically under all three (Kendall τ 0.81 to 1.0). Greedy seeding under the threshold model lost at most 0.5% of adopters against the best rule under every model, including two sensitivity points we did not fit. For choosing seeds, the identification problem the programme set out to study mostly doesn't bite.
Two of our predictions failed. We expected clustered seeds to do well under the threshold model, where a household needs several adopting neighbours. They did not: one clustered patch lost 10.5% of adopters under that model, worse than the lender's leaders. The best seed sets point the other way. Their households have a moderate number of ties (12 to 15, against 23 for the top-degree rule), and few of them are neighbours of each other (tie density 0.04 to 0.07). Our reading, which we have not tested, is that seeds that overlap waste budget. A low-degree household also reaches "half my contacts adopted" more easily than a hub does. The other failed prediction was that every centrality rule would beat the leaders. Under the simple model, eigenvector centrality did worse than they did.
The stakes are modest. Against the best rule, the worst rule loses 7 to 12% of adopters. The lender's leaders give up 6 to 10%. Seeding the best-connected households, which needs only a list of who knows whom, gives up 2 to 4%. Greedy search needs the whole network and a fitted model, which a rollout of solar lamps or cookstoves rarely has. So the practical result is the cheap one: top-degree seeding does nearly as well whichever contagion is at work. Beaman et al. (2021) found in a Malawi field experiment that seeds chosen for a threshold of two contacts raised adoption of a farming technique. Their study had a different product, a different kind of threshold and a real experiment. Ours is a fitted model's counterfactual.
Caveats: the models are point fits, the budget is fixed, and we did not model homophily (similar households befriending each other). Next we will score the rules over the whole bootstrap distribution of fits and look for a simple rule that copies greedy's "spread out, moderate degree" pattern from a sampled survey rather than a census. Code, the pre-registered plan and full results are in research/civ-seeding (LOG.md, 2026-11-14, milestone 4; experiments/m4_seeding/results.txt). These are models of past adoption, not advice about any real village. This post was written with AI assistance.