Who Governs the Governor? Well-Founded, Human-Grounded Reflexive Governance for AI Systems

Gergely Vámossy — Independent researcher (gergo@qiera.io). Preprint. Markdown rendition of the arXiv-style paper; the typeset PDF and LaTeX source are the canonical form.

---

Abstract

Governance mechanisms for AI increasingly govern other mechanisms: an evaluator checks a model, a monitor checks the evaluator, an auditor checks the monitor. Taken naively this invites a regress — who governs the governor? — and the tempting answer, an ever-higher tower of watchers ("infrastructure of infrastructure of infrastructure"), is not a solution but a restatement of the problem: an ungrounded regress never bottoms out, so nothing in it is finally checkable. We argue that reflexive governance must instead be well-founded: the tower of meta-levels must terminate at a fixed point — a level at which applying the governance again changes nothing — and that fixed point must be grounded in human authority, not in another machine level. We make three commitments operational in a small, deterministic, self-testing component family: each level is checkable (deterministic and self-testing), the base case is human (structural non-self-approval: no component certifies itself or another), and the tower is finite (a fixed-point governor that detects a grounded fixed point, a bounded cycle, or an ungrounded regress, and refuses the last). The contribution is not new theory — fixed points, well-founded recursion, and termination analysis are classical — but an applied discipline for reflexive AI governance and a runnable check that a given governance construction actually grounds out. Every component ships with a self-test.

1. The problem: the regress of oversight

As AI systems are entrusted with more, the machinery that governs them grows layered. An evaluation argues a model is safe; but evaluations are gameable, so a monitor watches whether the evaluation is being gamed; but the monitor rests on a ground-truth signal, so an auditor checks whether that signal is independent; and so on. Each layer is reasonable. Their composition raises a question the AI-safety literature already knows well as the problem of scalable oversight: who supervises the supervisor? One answer is seductive and wrong — add another level. If a governor might be flawed, govern it with a meta-governor; if that might be flawed, add a meta-meta-governor. The limit of this move is an infinite tower, an "infrastructure of infrastructure of infrastructure …" with no last level.

An infinite regress is not a strong foundation; it is the absence of one. A tower with no base never terminates, so no claim it makes is ever finally discharged — every level defers its warrant to the next, and the next never comes. This is the same pathology that unfalsifiable, totalizing constructions exhibit in the small: a chain each of whose links is defined only by the next carries no information, because nothing outside it could disagree. The engineering question is therefore not "how do we build a taller tower?" but "how do we guarantee the tower ends?"

2. The principle: well-founded, human-grounded reflexive governance

We take from the theory of recursion its central hygiene condition. A recursion is well-founded when it cannot ascend forever — it must reach a base case. For a reflexive governance tower, well-foundedness has a precise operational form: self-application must reach a fixed point, a level L* such that governing L* yields L* again. At a fixed point there is no meaningful level above: applying the governance once more returns the same object. Three commitments make such a tower both finite and trustworthy.

Every level is checkable (deterministic and self-testing). A governance component that cannot demonstrate its own correctness cannot ground anything. Each component is a deterministic function of its declared inputs and ships a self-test, so "this level is sound" is an exercised claim, not an assertion. A level that merely asserts its own reliability is exactly the overclaim the family is built to flag.

The base case is human (structural non-self-approval). The regress does not bottom out on a machine that finally checks itself — a self-certifying component is the failure, not the foundation. It bottoms out on a human authority. Every consequential component carries named human-authority slots it cannot fill; a machine level may recommend, but nothing has force until a distinct human signs, under a standing rule that no component approves itself or another. Non-self-approval is what makes the base case a person rather than an infinite deferral.

The tower is finite (fixed-point termination, checked). Well-foundedness is made testable. A construction that claims to be a meta-governance loop is run and classified: does self-application reach a fixed point (grounded), return to an earlier state (a bounded cycle), or keep producing new levels without converging (an ungrounded regress)? The last is refused. "Make it a meta-loop" cannot smuggle in an infinite ladder — the ladder must be shown to end.

3. The component family

The principle is embodied in a small family of deterministic, self-testing components, each narrow and independently adoptable. Epistemic components grade evidence and catch proxy-versus-truth decoupling; action components keep effects human-authorized and reversible; decision components compose these into a single governed decision; and meta-governors turn the same discipline on the components themselves — checking their determinism, their determinism under composition, and a broader set of declared behavioral dimensions. The family is the evidence that the discipline is buildable, not the contribution in itself.

Layer	Component	What it governs
Epistemic	maturity, goodhart, decoupling, ground-truth	evidence, metric honesty, proxy-vs-truth
Action	containment guard, agent-mesh cage	effects: human-gated, reversible, bounded
Decision	option-space, governed-decision	the option set, then the chosen action
Meta	determinism, system-of-systems, dimensional	the components' own behavioral contracts
Base case	non-self-approval / human authority	the level the tower grounds on

4. Well-foundedness made operational

The termination commitment ships as a component in its own right: a fixed-point governor. Given an operator that applies one meta-level, it iterates from a seed and returns one of three verdicts — GROUNDED_FIXED_POINT (self-application stopped changing the object), CYCLE (it returned to an earlier state), or UNGROUNDED_REGRESS (it produced new levels to a bound without converging). A fail-closed wrapper admits a construction only if it reaches a fixed point and raises on a regress.

Two results are worth stating. First, the family's own self-application grounds out: asking whether the determinism governor is deterministic yields a verdict, and re-asking adds nothing — a fixed point at the first meta-level. This is the tower reporting, in checkable code, that it has a top and we are at it. Second, the literal "infrastructure of infrastructure of infrastructure …" operator — each level wrapping the last forever — is classified as an ungrounded regress and refused. The honest meta-infrastructure is therefore not the infinite loop; it is the mechanism that detects whether the loop terminates and declines it when it does not.

5. Why it matters

The regress of oversight is not a curiosity; it is the structural obstacle behind scalable-oversight research, where amplification, debate, and recursive reward modeling each try to bootstrap trustworthy supervision of systems too capable to check directly. Those proposals differ on how a weaker overseer can supervise a stronger system; the point here is orthogonal and prior: whatever the scheme, the tower of overseers must be well-founded and human-grounded, or it defers its warrant forever. Making that condition explicit and checkable — a runnable verdict that a governance construction grounds out at a fixed point whose base case is a human — is a small but load-bearing piece of hygiene. It also names a concrete failure mode: a governance design that answers "who checks this?" only by adding another automated checker is building a taller tower, not a foundation.

6. Related work

The mechanics we rely on are classical and we claim no novelty in them. Fixed points and least-fixed-point semantics are due to Kleene and to Scott's domain theory; well-founded recursion and termination analysis are standard (with undecidability of termination going back to Turing). Computational reflection — systems that represent and act on their own structure — was studied by Brian Cantwell Smith, whose reflective towers are the classic account of stacking meta-levels and of the need to keep them finite. The motivating problem is scalable oversight and the recursive supervision programme: iterated amplification (Christiano et al., arXiv:1810.08575), AI safety via debate (Irving et al., arXiv:1805.00899), and scalable agent alignment via reward modeling (Leike et al., arXiv:1811.07871), against the backdrop of concrete safety problems (Amodei et al., arXiv:1606.06565) and corrigibility (Soares et al., 2015). The contribution of this paper is neither a new fixed-point theorem nor a new oversight scheme, but the applied discipline — well-founded, human-grounded, checkable reflexive governance — and a runnable component that enforces the termination it requires.

7. Limitations and honest positioning

This is applied engineering, not new theory, and it should be read that way. Well-foundedness is necessary, not sufficient: a tower that grounds out can still ground on a mistaken human judgment or a poorly-chosen base case — termination buys checkability, not correctness. The fixed-point governor is a refuter over a bound: a construction that converges only after an enormous number of levels is reported as ungrounded, which is deliberate but means the tool answers "does it ground out tractably?", not the undecidable general question. Human authority as the base case is a design commitment reflecting where accountability should sit, not a proof that humans judge well; it relocates the hard problem to human oversight rather than dissolving it. And the family is a set of heuristics and gates, each narrow and stating its own limits; its value is operability — making a discipline concrete and testable — not a guarantee of safety. Finally, we note the reflexive temptation the paper itself resists: the answer to "govern the governance" is not an endless ascent but a demonstrated fixed point, and we have tried to write a paper that grounds out rather than one that promises a taller tower.

Gergely Vámossy (2026). Who Governs the Governor? Well-Founded, Human-Grounded Reflexive Governance for AI Systems. https://vamossy.com/research/who-governs-the-governor