Two Bounds, Two Papers
A governance function can fail for two reasons that have nothing to do with how mature it is: the system it governs cannot be seen into, or the authority it exercises was never properly granted. Two preprints, published in November 2026, take one reason each.
Most of what is written about AI governance asks how good the governance is: which committees exist, which policies apply, how mature the processes are. The two papers I published on Zenodo in November 2026 start one step earlier. They ask what limits governance before its quality comes into play at all. The answer they give is: two things, and they are independent of each other.
The first bound: can the system be governed at all?
Governability as a Computed Property treats governability as a property of the system being governed, not of the governance function. How far can the exercise of an organisational capability be observed, represented, attributed, constrained, reversed and adapted? Where it cannot, a mature governance function produces documents, not control. Governability sets a ceiling.
The paper proposes to compute that ceiling from a semantic model of the enterprise and the runtime evidence of its agents, rather than to assert it in a self-assessment. It is a research agenda, not a results paper, and it says how its central claim could fail: if high governance effectiveness turns up where governability is low, the ceiling does not exist. And if the index predicts nothing that the maturity of formal governance does not already predict, it has no use of its own. It would be measuring ritual, not capacity.
The second bound: is the authority legitimate?
A capability can be fully observable and constrainable while the person who governs it has no entitlement to do so. Under Which Mandate? represents where agent authority comes from: which body granted which decision right, under which mandate, within which limits, and who remains answerable.
That chain is built into a checked metamodel and consulted by an agent's tool gate before every call that changes something. In a demonstration with two agents and seventeen scenarios, a baseline of identity, permission and some human approval executed actions that the institutional check stopped: the approver held the wrong right, or the right one not yet. The demonstration also showed two things I did not design for. One was an agent that, once refused, tried another tool. The other was a defect in my own artefact: an approval that was not bound to the action it approved.
Why two papers
Effectiveness and legitimacy are separate properties, and an arrangement can have one without the other. Keeping them in separate papers keeps them separable in measurement too: the two constructs must share no indicator, or neither can be shown to bound governance on its own.
Both are preprints, open under CC BY 4.0. Comments and objections are welcome, especially the ones that would make a claim fail.
New posts by email: subscribe to the newsletter. One email per post, no tracking.