Evaluation of the methods for Conducting and Analysing 1-shot (Go/No-Go) Experiments

One-shot experiments are crucial in fields where repeated testing is impractical, they trial multiple samples to model the response, improving these methods can lead to better decision-making, enhanced safety, and more efficient resource utilisation. While algorithms have evolved from targeting the threshold, to characterising the whole response curve, model fitting continues to assume symmetry and is calculated by the maximum likelihood of a Generalized Linear Model with a Probit link function. Analysis of large datasets indicates that symmetry cannot be assumed for the underlying response curve.

Using simulated data, we examine the impact of algorithm choice and alternative models at key probabilities, such as 0.1%, the threshold, 99.9% and extending further into the tails. We evaluate the performance of Bruceton, Langlie and Neyer d-optimal algorithms using a probit model, log-probit model and a novel 4-parameter model that can manifest asymmetry. By simulating the underlying response curves we quantify the expected benefits of additional tests, algorithm choice and fitting to an asymmetric model. This study clarifies the accuracy and robustness of approaches to uncertainty assessment for safety and reliability assurance, while also better defining performance and required resource utilisation.

UK Ministry of Defence © Crown Owned Copyright 2025/AWE.