The approach was set out in a series of papers from projects in India and Nepal. Joshi and Witcombe (1996) described four steps, namely identifying farmers' needs, searching for suitable material, testing it in farmers' fields and disseminating what farmers prefer, and that sequence still organises most programmes. Participatory varietal selection deals with fixed lines, and is distinguished from participatory plant breeding, in which farmers select within segregating material (Sperling et al., 2001). Later writing recast both as forms of client-oriented breeding, the point being that the purpose is a better fit to users and that participation is the means (Witcombe et al., 2005). An early application to finger millet in Karnataka was reported by Gowda et al. (2000).

There is a quantitative argument as well as a social one. Atlin et al. (2001) examined decentralised selection with the tools of selection theory and showed that selecting in the target environment is advantageous when the genetic correlation between station and farm is low, provided that on-farm trials are replicated well enough to keep heritability acceptable. Hill agriculture, with its steep gradients in altitude, aspect and soil depth over short distances, is a setting where that correlation is likely to be weak. Experience with barley in dry areas led Ceccarelli and Grando (2007) to the same position, and the practical methods are collected in a handbook edited by Ceccarelli et al. (2009).

Several formats exist for testing varieties with farmers, and they differ in what they measure and in the burden they place on participants. The mother and baby design pairs a small number of replicated trials containing all entries with many single-entry or few-entry plots managed entirely by farmers (Snapp, 2002). Triadic comparison of technologies, known as tricot, gives each of a large number of farmers three entries drawn from the full set and asks only for rankings (van Etten et al., 2019). Farm walks and post-harvest tests can be attached to either. Table 1 compares the formats. For finger millet in the hills this article proposes the mother and baby design as the core, for reasons given under the third stage below.

Work begins in the season before any trial, with the methods of participatory rural appraisal (Chambers, 1994). Group discussions, held separately with women and with men, establish which landraces and varieties are grown, on what kind of land, in what sequence with other crops, and why. A matrix ranking exercise turns the discussion into a list of traits ordered by importance. Women do much of the transplanting, weeding, harvesting and all of the processing of finger millet in most hill communities, so their ranking may differ from that of men and must be recorded on its own. The team also walks transects from valley floor to upper terraces to define the land types that the trials will have to represent. The stage ends with a short written profile of the variety farmers are looking for in each land type.

The search for entries is guided by the profile from the first stage and not by what happens to be on the shelf. Three sources are worth combining: varieties already released for comparable environments, including those released in a neighbouring state or country; advanced lines from national breeding programmes; and landraces from other hill districts held in genebanks. The core subset of the finger millet collection described by Upadhyaya et al. (2006) offers a structured sample of the last. Between six and ten entries, including the most popular local landrace as a check, is a manageable number. Seed must be obtained under the proper agreements and checked for germination and for seed-borne disease before it is distributed.

A mother trial containing all entries in two or three replicates is placed on each main land type, on land lent by a farmer who then manages it in the usual local way. Around each mother trial, a larger number of households grow baby trials. Each baby consists of one or two new entries next to the household's own variety, on plots large enough to yield grain for a family meal and straw for a fair comparison. Entries are allotted so that every one of them appears on a similar number of farms. The design is preferred here to tricot for two reasons. Finger millet programmes in the hills usually need measured grain and straw yields to support a release proposal, which mothers supply. And terraces are small, so that a baby with one new entry is easier to host than three unfamiliar ones. Where a programme already has release data and seeks only wide testing, tricot would be the better choice.

Evaluation takes place at three moments. Near maturity, farmers visit the mother trial in a farm walk. Entries are identified by number, not by name, and each visitor casts votes privately for the entries liked most and least before the group discusses reasons; women and men vote with tokens of different colours so that the tallies can be separated. At harvest, grain and straw are weighed from the mothers, and blast and lodging are scored. After harvest, households that grew babies are asked to rank the new entry against their own for each trait in the first-stage list, and a tasting session is held in which flour from each entry is cooked into the local staple, such as roti or the thick porridge called dhindo, and scored without the entries being identified.

Selection is pointless if seed of the chosen varieties cannot be had. In the second season, entries that farmers rejected are dropped and the remainder are repeated, while the team arranges seed multiplication of the leading entries with a farmers' group or a community seed bank. Finger millet is self-pollinating and its seed rate is low, so a variety can spread quickly from farmer to farmer once a modest quantity is in the village. In parallel, the mother trial data and the preference records are compiled in the format the national release committee requires, since release procedures in the region, notably in Nepal, have been adapted to admit participatory data as supporting evidence.