Prompt optimisation

A hosted model can only be changed through its input. The thesis's loop, derived from GEPA, plays the current prompt on three training floors, lets a teacher model read the traces and write a diagnosis and a revised prompt, and keeps the revision if it scores at least as well. A fixed validator rejects any candidate that contains coordinates, monster names, routes, lists of allowed fights or numbers of 100 or more. The question (RQ2) was whether evolved prompts beat a hand-written prompt or one that states only the rules, and how that depends on the observation.

The study stated its hypotheses in advance and measured the noise floor directly: the variation between two evaluations of the same prompt, paired on thirteen held-out generated floors.

Evolution rediscovers the rules that follow from the combat formula and the floor structure, but rarely the limit on HP spent per fight or the scarcity of keys (Table 6.7). With the annotations present, its gains stay within the variation between repeated evaluations; across the experiments, the observation constrained success more than the prompt did.

Every run, candidate by candidate, with the prompt diff, the teacher's diagnosis, the validator's verdict and the training episodes, is on the prompt-optimisation pages; the one rewrite that changed behaviour is here.