Pygmy hippopotamus optimization algorithm (PHOA): A bio-inspired metaheuristic under ablation and benchmark controls with a large language model phase controller

A new metaheuristic is credited with the gain it reports, and a language model attached to one with the gain that follows. Neither can be examined without a control. This article builds four and applies them to one algorithm and its language model phase controller. The controls are an origin translation that moves the optimum away from the coordinate origin, a budget-matched ablation of every claimed operator from shared initial populations, a prompt ablation that removes the reference rule from the controller prompt, and an audit of every benchmark against its declared minimizer. The algorithm is the Pygmy Hippopotamus Optimization Algorithm, compared against fourteen methods including a covariance-based optimizer and an adaptive differential evolution, on 62 problems with 30 paired repetitions under one budget. It places third of fifteen at a Friedman mean rank of 4.073, and Holm places two methods above it. Of seven ablated variants only two separate from the full method after correction, and one is a null control, so the evidence credits one operator of the seven. The language model reproduces the reference rule written into its own prompt, with the progress ratio alone accounting for a coefficient of determination of 1.0000. Removing that one sentence from the prompt leaves 0.0004, and the two prompts do not separate on search performance. The audit found incorrect benchmark definitions, including one function of the official CEC2017 suite whose published shift vector is not its minimizer. Each control cost one run, and each changed what the study could claim.