Passage
A feather, once grown, is chemically finished. The atoms locked into its keratin are the ones the bird took in during the few weeks the feather was forming, and they are not exchanged afterwards, so a feather is a dated sample of a bird's chemistry. The ratio of heavy to light stable isotopes of hydrogen in rainfall changes smoothly with latitude and with elevation, and that ratio passes from rain into plants, from plants into insects, and from insects into the bird. A feather ought therefore to carry a record of the place where it was grown, and researchers have used the correspondence to assign moulting grounds to small migratory songbirds caught and released thousands of kilometres from where they breed.
What a feather records is not rainfall but a diet, and a diet is not local in the way rain is. Insects that emerge from a stream carry the signature of a whole catchment, which may drain ground far higher and colder than the meadow the stream runs through; a bird feeding at the water is sampling a watershed, while a bird feeding in the grass beside it is sampling the sky. Two feathers grown a hundred metres apart can differ by more than two grown a thousand kilometres apart.
Calibration is supposed to absorb this. Reference curves are built from the feathers of birds that do not migrate, collected where those feathers were grown, and the gap between what the feathers show and what the rainfall maps predict is carried forward as an offset. The procedure is reasonable in itself. But the offset is estimated from the same body of material against which later assignments are checked, and a disagreement can always be accommodated by revising it. A method whose corrections are fitted to the discrepancies it produces is hard to catch in an error.
The remedy is not to abandon the assignments. It is to stop reporting them at a precision the signal cannot support, and to look for agreement from a marker that fails differently — a trace element whose distribution is geological rather than meteorological. Two methods that agree are worth more than one method with a narrow interval, because the narrow interval is only as good as the assumption nobody has tested.
