Confidence: low. Evidence: taxonomies plentiful; validated assessments rare. Last substantive change: 2026-07.
Calling a factory mature or autonomous is a claim, not a branding choice. The claim needs evidence, a way to check it, and conditions that revoke it.
The conclusion
A high autonomy claim should be an earned, domain-specific assurance case, not a self-description. A factory needs measurable gates for quality, rollback, containment, audit reconstruction, human attention, and controlled adaptation. Whether the lights may safely go off is a conclusion you demonstrate for one domain at a time, not a level you assign yourself.
How the thinking got here
Dan Shapiro's ladder from spicy autocomplete to the dark factory usefully named the destination. Later work showed that agency, orchestration, risk, and verification vary independently, so a one-dimensional rung cannot describe a real factory. The field moved toward two-axis frameworks, capability profiles, and assurance cases, but validated maturity assessments have not yet arrived.
Credible alternatives, and when each is right
| Approach | Right when |
|---|---|
| One-dimensional maturity ladder | communicating direction to newcomers |
| Two-axis agency and orchestration matrix | placing multi-agent systems |
| Capability profile | describing uneven strengths honestly |
| Assurance case | making a defensible autonomy claim |
| Risk-class scorecard | gating autonomy by blast radius |
Where it fails and what we still don't know
Taxonomies are plentiful and validated assessments are not, so "Level 5" is easy to claim and hard to substantiate. Open questions include a standard assurance case, minimum evidence profiles by risk class, recurring certification, third-party audit, and falsifiable criteria for losing a maturity claim.
What would change our mind
A widely accepted, falsifiable assurance standard that distinguishes a governed dark factory from an impressive demo would turn maturity from a marketing term into a measurable state.