Tim Palmer - On seamless prediction and the reliability of climate forecasts It is well understood that forecasts of weather and climate, whether 1 day ahead or 1 decade ahead, are necessarily probabilistic - neither the initial state, nor the models used to integrate the equations of motion are perfectly known. This itself of itself does not make decision making impossible or even, necessarily, difficult. However, it is crucially important that if decision makers are not to be misled, forecast probabilities must be reliable. A summary of the reliability of ECMWF System 4 seasonal forecasts is described. For decadal timescales, an assessment of reliability poses a severe challenge because the number of independent initial dates available to make decadal forecasts is very limited and on these timescales models continue to show substantial biases against observations. Here it is proposed that the reliability of coupled model seasonal forecasts can be a useful guide to longer timescale probabilistic reliability. This is tested by studying idealised seasonal ensemble forecasts made at T95 atmospheric resolution, and initialised and verified using mutlidecadal T959 integrations run in AMIP and in climatechange timeslice mode. Time permitting, some remarks will be made on some new proposals for using inexact computing hardware in next generation cloud-resolved global climate models.