A sensor-linked modelling tool for live prediction of uncontrolled compartment fires, K-CRISP, has been developed in order to facilitate emergency response via novel systems such as FireGrid. The modelling strategy is an extension of the Monte-Carlo fire model, CRISP, linking simulations to sensor inputs which controls evolution of the parametric space in which new scenarios are generated, thereby representing real-time "learning" about the fire. CRISP itself is based on a zone model representation of the fire, with linked capabilities for egress modelling and failure prediction for structural members, thus providing a major advantage over more detailed approaches in terms of flexibility and practicality, though with the conventional limitations of zone models. Large numbers of scenarios are required, but computational demands are mitigated to some extent by various procedures to limit the parameters which need to be varied. HPC (high performance computing) resources are exploited in "urgent computing" mode. The approach adopted for steering is shown to be effective in directing the evolution of the fire parameters, thereby driving the fire predictions towards the measurements. Moreover, the availability of probabilistic information in the output assists in providing potential end users with an indication of the likelihood of various hazard scenarios. The best forecasts are those for the immediate future, or for relatively simple fires, with progressively less confidence at longer lead times and in more complex scenarios. Given the uncertainties in real fire development the benefits of more detailed model representations may be marginal and the system developed thus far is considered to be an appropriate engineering approach to the problem, providing information of potential benefit in emergency response.