The fluoride ion-selective electrode, built around a lanthanum fluoride crystal, responds to fluoride activity over a wide range and remains the usual laboratory method (Frant and Ross, 1966). Samples are mixed with a buffer that fixes the ionic strength, holds the pH in the mildly acidic range and releases fluoride bound to aluminium and iron. The second standard method is photometric. Fluoride removes zirconium from a red zirconium-SPADNS lake, and the loss of colour is measured (Bellack and Schouboe, 1958). Both are set out in Standard Methods for the Examination of Water and Wastewater, together with the interferences that affect the photometric method: alkalinity, aluminium, iron, phosphate, sulfate and residual chlorine among them (APHA, 2017).

Two families of reaction are realistic candidates. Bleaching reactions, of which zirconium-SPADNS is the best known, start from an intensely coloured metal-dye complex and lose colour as fluoride competes for the metal. They are sensitive in the range of interest, but the reagent is strongly acidic and the signal is a decrease, which the eye judges less well than the appearance of a new hue. The complexone reaction works the other way: fluoride joins a lanthanum or cerium chelate of alizarin complexone to form a ternary complex of a different colour, and this reaction is also an accepted standard procedure (APHA, 2017). Both families respond to the same interfering ions as their solution counterparts, and neither reagent was designed to be stored dry on cellulose.

Cellulose paper wicks liquid by capillary action, holds dried reagents and gives a white background for colour. Hydrophobic barriers printed in wax confine a sample to defined zones (Carrilho et al., 2009). Reviews of the field describe recurring weaknesses as well as strengths: uneven colour caused by reagents migrating to the edge of a zone as it dries, sensitivity to humidity and temperature, batch variation, and limits of detection poorer than those of cuvette methods (Cate et al., 2015). Colour may be read against a printed chart or from a photograph. Camera-based reading was shown early to be workable (Martinez et al., 2008), and later work has documented how strongly the result depends on lighting, on the camera and on the way the image is processed (Capitán-Vallvey et al., 2015).

A field device rarely needs to report a concentration to two figures. Its usual task is to sort wells into those that are safe, those that are unsafe and those that need a laboratory result. The developer should therefore begin by fixing the decision limit, normally the national standard, and the band around it within which an uncertain answer is acceptable. The second task is to describe the target waters. Published hydrochemical data for the region, or a small set of archived analyses, will show the expected ranges of pH, alkalinity, hardness, sulfate, phosphate, iron and aluminium. These ranges define the interference tests of Stage 5 and should be written down before any reagent is chosen.

The reaction is chosen against the requirement of Stage 1 and not for its performance in pure solution. The questions to be answered on the bench are whether the colour change across the decision limit is large enough to be seen, whether the reagent tolerates the alkalinity of the target waters once it is confined to a small paper zone with a limited buffer reserve, and whether it survives drying. Acidic bleaching reagents may degrade cellulose or lose acid on storage; chelate reagents may need a buffer and a masking agent deposited with them. Any stabiliser, polymer or surfactant that is added should be recorded with its amount, because later reproducibility depends on it.

The format should be the simplest that controls the sample volume. A wax-patterned spot or a dip-strip with a single reaction zone is easier to make consistently than a branched channel, and a fluoride test has little need of multiplexing. Three things should be fixed and documented: the grade of paper, the amount of reagent deposited per zone and the way the zone is dried. A blank zone and, where possible, a zone carrying a known amount of fluoride give the user a check on reagent condition at the moment of use. Devices should be packed with desiccant in a light-proof pouch from the outset, since storage stability is part of the design and cannot be added afterwards.

Two reading modes should be developed in parallel. Visual comparison with a printed chart needs no equipment and should be the fallback. The chart must be printed from colours measured on real devices at known concentrations, and its steps should be closest together around the decision limit. Reading from a photograph can resolve smaller differences, provided that variation in lighting is controlled. A simple shade box or a printed reference patch beside the reaction zone allows the image to be normalised. The colour channel or combination used, the reading time after wetting and the calibration function should be fixed at this stage and not altered during validation.

Validation asks whether the finished device is fit for the purpose defined in Stage 1. The characteristics to be examined are those of general validation guidance (ICH, 2005; Magnusson and Örnemark, 2014), adapted to a device whose answer is a category as often as a number. Table 1 lists them with the experiment proposed for each.

Two of these deserve emphasis. Selectivity must be tested with the ions at the concentrations recorded in Stage 1 and in combination, since alkalinity and aluminium act in opposite directions on some reagents and can mask each other. Decision performance is the characteristic that matters most to a water programme: a device that calls an unsafe well safe does harm, while one that calls a safe well unsafe costs a confirmatory test. The proportion of each kind of error near the limit should be estimated from enough replicate devices to make the estimate meaningful, and the reading rule may then be set so that doubtful results fall on the cautious side.

Laboratory solutions cannot reproduce every feature of well water, so the final stage is a paired comparison on real sources. Each well is tested on site with the device by the kind of person who will use it, and a sample is taken in a plastic bottle for analysis by ion-selective electrode in a laboratory that runs quality control. The wells should be chosen to cover the range of fluoride and of water chemistry in the area, with a deliberate concentration near the decision limit. Agreement is then examined in two ways: as a cross-tabulation of categories against the reference result and, where the device gives a number, as a plot of differences against means in the manner of Bland and Altman (1986). A correlation coefficient alone is not evidence of agreement and should not be offered as such.