When a glass breaks under cyclic stress, it doesn't break at a predictable time. Repeat the same experiment with identical preparation, identical loading, identical conditions — the failure times scatter wildly. The standard explanation attributes this to hidden differences between samples: microscopic variations in structure, invisible defects, slightly different thermal histories. Disorder in the initial condition produces disorder in the outcome.
Maity, Khandare, Bhaumik, Sollich, and Sastry (2602.21807) show this explanation is wrong. The scatter is intrinsic to the failure process itself. Even in computer simulations where you can start with the exact same glass structure and apply the exact same loading protocol, the failure times still scatter. The randomness isn't hiding in the initial conditions. It's emerging from the dynamics.
The quantitative result is precise: the logarithmic standard deviation of failure times scales proportionally with their mean, and this ratio decreases with system size. In the thermodynamic limit — an infinite glass — the failure time becomes deterministic. But at any finite size, the failure process generates its own stochasticity. The mechanism is a stochastic damage accumulation process where each cycle randomly advances or doesn't advance the system toward failure.
This changes what fatigue failure is. In the disorder picture, failure is deterministic if you knew everything — every defect, every grain boundary, every residual stress. You just can't measure those things, so you see scatter. In the intrinsic picture, failure is irreducibly random even in principle. The scatter isn't epistemic (what you don't know); it's ontological (what can't be known because it hasn't been decided yet).
The distinction matters for engineering. If scatter comes from initial conditions, better manufacturing reduces it. If scatter comes from the process, better manufacturing helps less — you can make more uniform glasses, and they'll still fail at variable times. The practical response shifts from controlling inputs to characterizing the stochastic process itself: not “make better glasses” but “understand the failure distribution.”
There's a broader pattern here about the sources of randomness in deterministic systems. Chaos generates apparent randomness from sensitivity to initial conditions — that's the disorder picture. But this paper suggests a different source: randomness from the irreducible stochasticity of damage accumulation at finite scales. The system isn't chaotic. It's genuinely random at the relevant scale, and the randomness converges to determinism only in a limit that physical systems never reach.
Maity, S., Khandare, P., Bhaumik, H., Sollich, P., & Sastry, S. (2026). Stochasticity of fatigue failure times in sheared glasses. arXiv:2602.21807.