How far up the ladder of life can environments carry a bottom-up chemistry? An ordered search over 13 environmental variables

Which environments can turn non-living chemistry into something that inherits, encloses itself and evolves?
We built a fully bottom-up artificial chemistry with nothing living put in. It runs from feedstock through
activated monomers, chains, catalysts and template copying to amphiphiles and vesicles. Its only inputs are 13
environmental variables, encoded as 19 dimensions with ranges taken from the literature. Rate constants that no
measurement fixes are kept as explicit unknown dimensions, up to 54 dimensions in all. We searched this space
layer by layer with COMPASS, an ordered, non-repeating active-learning engine, and detected seven levels
automatically against paired controls.
The model climbs the ladder up to coupled protocells, and each step narrows the environments sharply:
activated feedstock appears in about 45% of the space;
persistent chains in about 8.5–16%;
self-sustaining catalytic networks only when short chains are highly catalytic (0.3% of a focused
region);
heredity by template copying in a narrow class of environments (5 confirmed);
heredity inside stable vesicles in 4 environments found by the search;
coupled protocells wherever heredity and vesicles coexist (298 of 435 coexistence runs).
All environments at the top are cool to mild (0–25 °C), slowly flushed, enclosed water bodies with temperature
cycles of days to weeks across the melting point of short duplexes, and near-neutral to slightly acidic water. The
ladder shows no conflict between consecutive levels: one stable environment of this kind could carry the
chemistry through all levels reached. The hardest step is having heredity and stable membranes in the same
place. The ladder stops before Darwinian evolution. Copying families sweep once, after which copying erases
all heritable variation in vesicle content, because sequences in the model have no phenotype. We therefore
predict that sustained evolution in such systems needs a sequence-dependent trait that copying does not erase.
For every result we state whether it follows from the model's assumptions or is derived.