The Accountable Author
AI-Generated Systems for Accountable Human Sign-Off
James Wolstencroft · June 2026
The Tekmerium · Paper VI of VI
Abstract
The earlier papers treat artificial intelligence as a component built into a system, bounded, monitored, and externally accountable. This paper concerns the inversion the established framework names but does not develop: AI as the author doing the building. When a machine composes systems faster than any human can review, the architect’s task shifts from drawing to directing, and the load-bearing act becomes not the design but the signature beneath it. Drawing on generative architecting and the “Accountability Horizon”, it sets out how an AI author is kept inside the accountability apparatus: judged by the Digital Backbone of ‘Engineering the Decision Trail’, fed from the validated Vault of ‘Governing the Model Vault’, and answerable through a human who signs. It then develops the Omnium: the enterprise model of many products and many authors reasoning through one shared Continuum, a move that runs ahead of the established framework’s built content and is flagged throughout as an extension.
1. When the tool becomes the author
The question is not whether a machine can design, but how a machine’s authorship remains within the apparatus that makes engineering accountable.
Every prior paper treats AI as something within a system, assuming a human still does the designing; this paper takes up the moment when that assumption breaks down. The change is not cosmetic: an authoring machine can propose a hundred coherent configurations before lunch, none reasoned through by a person, and the accountability the series built does not automatically survive that acceleration. Until now, the Continuum’s trail was enough: because, even with AI assisting, a human still authored each decision, a traceable record carried a responsible author with it. When the machine authors the record, the author falls out of it: traceability without accountability. A design accepted in seconds is one whose reasoning was examined in seconds, or not at all. The capability is coming regardless; the open question is authority: how machine authorship stays accountable, and who answers when it does not.
2. From governed component to governing author
Governing the author is not a stronger version of governing the component; it is a different problem, one level up.
The established framework already draws the line: everything it says about bounding and monitoring AI assumes an architect integrating AI as a component, and the next frontier reverses that, AI as a co-designer in the architecting process. That reversal is the boundary between ‘Engineering the Decision Trail’, which governs the AI within a system, and this paper, which governs the AI that produces systems. The roles fail differently: a misbehaving component is contained by the guardrails around it, but a misbehaving author propagates its error into the design itself, into the thing the guardrails are meant to protect.
3. Generative architecting
Generative authorship is already tethered at two points: the Vault bounds its inputs, the Digital Backbone judges its outputs.
The developed account is a division of labour: the architect sets goals and constraints, the AI explores a far larger design space than a human could enumerate, and the Digital Backbone acts as the fitness function that scores the options against encoded objectives (The Continuum, Ch. 22; worked examples are hypotheticals). Two features are load-bearing. The AI does not mark its own work: every generated option is validated against the project’s criteria before a human weighs it; and it may build only from a managed, validated store, not models conjured to fit the prompt. That store is the Vault. This paper’s contribution is to treat those two engines as the anchors that keep an author accountable, and to ask what a third anchor, the human, must still supply.
4. The Accountability Horizon
A named human still signs. Machine authorship does not remove the signature; it stretches the distance between the signature and the work.
In every regulated industry, a named human places their authority behind the claim that something is safe; that signature is the load-bearing element of the entire edifice of trust. Follow the trend and you arrive at a role that barely exists yet: the engineer accountable atop a body of decisions a machine made and no human fully checked. If the signatory cannot see which validated models built a design, what evidence supports each decision, and to what level it was proven, they can only sign in good faith and hope, which is how a person ends up carrying liability for a decision they let happen but never controlled. The signature is honest only if the apparatus makes the work reviewable at the speed it is produced; accountability is not a posture but an infrastructure requirement.
5. The human and AI operating model
Humans set the envelope, the apparatus enforces it, and humans adjudicate the exceptions: neither reviewing everything nor reviewing nothing.
The established framework provides the endpoints (the Digital Backbone judges, the human signs) but not the protocol in between; this section proposes one, assembled from parts the framework already provides. Authority is delegated to the AI author within an explicit envelope: a brief the Digital Backbone scores against, and a permitted Vault asset base to compose from. Inside it, the machine ranges freely, and only options that meet the criteria surface; authority is reclaimed by the human at defined gates: where a choice exceeds the envelope, trades off unranked objectives, or reaches a reserved threshold of consequence. Every machine-made decision re-enters the same audit trail through the capture gate, doing for the author what it does for the self-modifying component: converting an autonomous act into a traceable decision before it is trusted.
6. The Omnium: one Continuum and many authors
The Omnium is the enterprise composed of many products and many authors, human and machine, reasoning through one shared Continuum (an extension, not an extraction, and flagged as such).
The framework names this state in a one-paragraph synopsis but does not build it; what follows is a proposal, not an extraction. Its defining property is plurality on a single fabric: not one system with one author but a portfolio, each with its own Continuum, worked by human architects, generative engines and the reviewers who sign, all reasoning through one substrate of connected, evidenced decisions. Three things must hold for it to be more than a slogan:
One accountability fabric, so every author produces decisions in the same traceable form.
Authority legible across the whole, so who may delegate to which author over which assets is itself governed and visible.
A signature that scales without thinning, so the chain of authors and evidence behind any signed artefact is available on demand.
The Omnium is less a new technology than the disciplined generalisation of the whole series, and the most speculative section of this paper. The picture specified here is the one the prologue asserts as the series’ destination: the prologue states it, this section names the conditions under which it stands.
7. Limitations and open problems
The strongest claims here run ahead of the established framework, and the accountability case assumes the apparatus exists and is honest.
The Omnium operating model (§6) elaborates a concept the framework only names; it is grounded in that synopsis and the earlier apparatus, but not extracted from developed content, and should not be cited as if it were. The delegation protocol (§5) is proposed, not validated: built from existing parts, untested against a real authoring system. The generative figures are illustrative hypotheticals, not measured results. The whole accountability argument assumes a complete trail and truthful assurance stamps; where either is incomplete, the human is back to signing in good faith, the very failure the Accountability Horizon warns against. And liability is described, not resolved: the law is likely to tighten around who answers for a machine’s decision, but how it is apportioned between organisation, signatory and tool provider is flagged, not answered.
Boundary notes
What this paper borrows, and what it leaves to its companions.
It uses the Digital Backbone and capture gate of ‘Engineering the Decision Trail’ as the engine that judges and records machine-made decisions and draws only from the validated Vault of ‘Governing the Model Vault’ as its bounded input, re-deriving neither. It keeps strictly to the author role, handing the component role (bounding, drift, monitoring) back to ‘Engineering the Decision Trail’, and defers family-scale reuse and the cross-product mechanics of the Hyper-Continuum to the forthcoming companion volume. Where accountability touches regulation, the governance argument belongs to ‘Breaking the Reconciliation Trap’.
References
Regulation (EU) 2024/1689 of the European Parliament, 12 July 2024.
Tabassi, E. (2023). AI Risk Management Framework doi:10.6028/NIST.AI.100-1.
National Institute of Standards and Technology (2024). AI Risk NIST AI 600-1.
ISO/IEC 42001:2023, Information technology — AI — Management system.