COMPASS: mapping where life-like behaviour appears in stochastic models, with an ordered, non-repeating active search

Models of the origin of life are stochastic, slow and depend on many conditions. The useful question is oftennot what one run does, but where in condition space a level of life-like behaviour appears at all. That questionasks for a boundary. We present COMPASS (Condition-space Ordered Mapping by Prioritised Active Samplingof Stochastic systems). It tests conditions in an ordered, never-repeating sequence, simulates each conditiononce, fits a shared neural map of the success probability, and spends most of its budget where a bootstrapensemble of maps disagrees. None of the parts is new. Our contribution is the combination for this question, abenchmark protocol in which all methods share one map-maker and are scored against brute-force ground truth,and a record of what did not work.On six models, from a toy protocell to a fully bottom-up chemistry with 29 dimensions, COMPASS needed1.6–8× fewer simulations than Sobol or random sampling for the same map accuracy. The saving held at 20–30dimensions (2.8–4.0×), and the engine's own cost did not grow with dimension. On autocatalytic sets, thethreshold read off a COMPASS map lay within 7% of direct simulation. Ablations show that the saving comesfrom the loop as a whole: a standard boundary rule (straddle) does as well as our scoring. On a 54-dimensionalbottom-up model, COMPASS found and confirmed environments for levels that a ground-truth sample neverhit. But it does not raise hits per simulation uniformly, and most of its apparent advantage in raw hit countscomes from the design of the comparison.