Five experiments on whether complexity produces its own fragility
In the systems tested, and only where complexity is built from elements that compete for something finite. Where arrival costs the incumbents nothing, none of this happens.
The question behind this is cosmological and cannot be tested. The claim it rests on can be. If complexity does not generically produce its own fragility, then the argument from complexification fails whatever else is true, so that mechanism is what these experiments go after. Below is the yardstick everything is measured against, live, and then what five experiments did to the thesis that motivated them.
At a fixed link budget, compartmentalizing a community does not move the stability threshold. Grilli, Rogers and Allesina proved this analytically in 2016, so the experiment reproduces a known theorem rather than finding one. What compartments do instead is cap the reach of a failure, and that cap is graded: it weakens as the system nears criticality.
A community that grows its own complexity, one arriving species at a time, drives its leading eigenvalue from 0.84 of the bulk spectrum down to 0.03. Every interaction strength tested converges on the same endpoint while reaching very different complexity. Nothing tunes it there. No maintenance cost is present. This was established by Biroli, Bunin and Cammarota in 2018 and is reproduced here, not discovered here.
Predicted before it was run, then tested in one model. Compartments do not slow the approach to criticality at all, and by the time a community arrives there, containment has already failed for every partial separation. Only complete decoupling holds. The unpredicted part: a uniform background coupling of 0.00125 per pair breaks containment even with all of the interaction variance sealed inside compartments. Walls have to stop everything, not just the variable part.
Associative memories share no mathematics with interacting populations, having discrete states, an energy function and no interaction matrix whose eigenvalues decide anything. They converge to their own critical point anyway. Random Boolean networks do not, and their endpoint tracks how they were built. What separates them is interference: whether a new element degrades the ones already present. Species and memories both charge the incumbents. A Boolean node is free. The two that charge converge, the one that does not fails to.
Superlinear maintenance cost alone never collapses anything, it just shrinks the system. Collapse needs the cost to fall on whoever is left, so that losing members raises the burden on the rest. And the standard early warnings do not work here: across 61 collapsing and 88 surviving histories, variance separates them by 0.04 and autocorrelation by minus 0.14, rising less often before collapse than without it. The reason follows from experiment 02. A system permanently at marginal stability recovers slowly all the time, so critical slowing down has nothing left to signal with.
Complexification drives systems to marginality wherever complexity is built from elements that compete for something finite. That is a cross-domain statement formal enough to fail, and failing it is specific: find a substrate with interference that does not converge, or one without interference that does.
Universality. Accretion alone does nothing, so "structurally predisposed" holds only under a condition the thesis does not state. The Maintenance Thesis is also over-specified: it stipulates three ingredients and the phenomenon appears with none of them. And whether the universe is a system of this kind is not a question any of this can answer.