Disputes between separated parents over where a child will live and how time will be shared are poorly suited to adversarial trial. The parties must go on dealing with each other for years, the court's order governs conduct that no court can supervise, and the legal standard, the best interests of the child, leaves wide room for disagreement. Mnookin and Kornhauser (1979) observed long ago that most such disputes are settled by the parents themselves, bargaining in the shadow of the law. Mediation offers a structured setting for that bargaining, and courts have adopted it widely, often by way of a pilot attached to a single court centre.

Pilots are meant to produce evidence, and frequently they do not. A typical closing report gives the proportion of mediated cases that settled and the proportion of parents who said they were satisfied. Neither figure has a comparison, and both are taken at the moment when goodwill is highest. This article proposes a stronger design. It is addressed to court administrators, mediators and researchers planning a pilot, and it describes what should be decided before the first family is referred. Nothing here reports the results of an actual programme.

The research base for custody mediation is more substantial than for most court reforms. Early longitudinal work in the United States compared mediating and litigating families over time (Pearson and Thoennes, 1984). The best known study assigned families at random to mediation or to the ordinary adversarial route and followed them for twelve years; it found that parents without primary residence who had mediated remained more involved with their children (Emery et al., 2001). Reviews of the field report that mediation settles a majority of cases, that parents tend to rate it more favourably than litigation, and that evidence for lasting effects on children's adjustment is thinner (Kelly, 2004; Emery et al., 2005). More recent trials have tested variants that bring the child's perspective into the sessions (McIntosh et al., 2008; Ballard et al., 2013).

The same literature carries a warning. Grillo (1991) argued that mediation, particularly when mandatory, can disadvantage the party with less power and can pressure a parent who has been abused into negotiating face to face with the abuser. That critique changed practice: screening for intimate partner violence is now treated as a precondition of responsible mediation, and structured instruments exist for the purpose (Holtzworth-Munroe et al., 2010). An evaluation that ignored safety would miss the outcome on which critics and supporters most sharply disagree. There is, finally, the reason the dispute matters at all. Continuing conflict between parents is among the more consistent correlates of poorer adjustment in children after separation (Amato, 2001), so a pilot that produced agreements without reducing conflict would have achieved little.

The design has six components: a statement of programme logic, a comparison strategy, a set of outcomes, a measurement schedule, an analysis plan and a process evaluation. They are described in turn. All of them should be fixed in a written protocol before referrals begin, and the protocol should be lodged somewhere public.

Most of what the design requires is administrative. The lottery can be run by the registry from a sealed list prepared by someone outside the court. The case outcomes are already held in the file and need only a field that marks which group a case belongs to. The questionnaires are short enough to be completed on a telephone, and reminders can be sent by the same staff who send hearing notices, provided the answers go to the evaluators and not to the registry. The work that calls for outside help is the writing of the protocol, the calculation of numbers, and the analysis; a university department or an independent researcher can supply these at modest cost, and an arm's-length analyst also protects the court from the suspicion that it has marked its own work. What the design does demand of the court is patience, since the outcome that matters most cannot be known until eighteen months after the last family has its order.

Every case should be screened for violence, coercion and fear before any offer of mediation is made, using a structured instrument such as the one described by Holtzworth-Munroe et al. (2010) and administered to each parent separately. Cases screened out are excluded from assignment and proceed to the court with the screening information available to it. For cases screened in, safety is monitored as an outcome: any new protective order or reported incident during follow-up is recorded for both groups. Parents must be told that research is taking place, that declining to take part in questionnaires has no effect on their case, and that what they say to researchers will not reach the judge. Independent ethical review should be sought before the pilot starts. The child's right to have his or her views considered in proceedings, recognised in Article 12 of the United Nations Convention on the Rights of the Child, should be reflected in how the mediation model itself is chosen.

This design has not been put into practice or validated in the present article, and several things could defeat it. A single court centre may see too few eligible cases to detect any but a large effect, in which case centres should pool their pilots or the evaluation should be presented as a feasibility exercise. Follow-up questionnaires lose families over time, and those lost are seldom typical; the reliance on court-file outcomes is partly a defence against this. Parent-rated measures of child adjustment are coloured by the parent's own state. Judges and lawyers may resist random assignment, and persuading them requires early consultation. Finally, a result obtained in one court with one group of mediators does not automatically transfer to another, which is why the process record matters.

A custody mediation pilot can be evaluated credibly without a large research budget if the evaluation is designed before the programme opens. Assigning the offer by lot, measuring outcomes that the court file already holds, following families past the final order, screening every case and recording what was delivered are each manageable steps. Together they would let a court say whether mediation reduced repeat litigation and conflict among the families it served, and whether it did so safely.

To be declared by the author(s) before publication.

To be declared by the author(s) before publication.