Part of Temporal Constructs Appendix, which contains details behind the research: dates, quotations, descriptions of instruments, and sources.
General definition
The rules the practice holds itself to when it uses artificial intelligence, and the arrangements it has in place so that the work stays its own.
Relation to the practice
Gardner works with artificial intelligence on terms he sets, and these guidelines state those terms as rules. They cover what the practice runs on, what it retains, which models it chooses and how much computing it needs. They stand beside the programme's ethics requirements.
No closed infrastructure
The practice does not build on closed artificial intelligence infrastructure. It runs the Hermes agent framework, which is open source, and routes model calls through OpenRouter. The framework and the routing can each be replaced or held independently of any single supplier. A four-year research project cannot rely on an arrangement a provider can change or withdraw.
Open source software
The practice works mainly with open source software, which is a question of security as much as of principle. Running Linux is part of it, and so is preferring the tools the open source community maintains. That community is largely Linux-based, so the tools an agent is given are more easily made to work there. Microsoft infrastructure goes against the guidelines.
European alignment
Linux originated in Europe, and the practice is based in Europe. The arrangement is therefore kept European where it can be and aligned with European rules. What those rules require of artificial intelligence in research is set out on the frameworks the research works to.
Europe is moving in the same direction, at institutional scale. Public bodies have left Microsoft for Linux for reasons of digital sovereignty. Among them are the Danish ministry that switched in the summer of 2025 and the German state of Schleswig-Holstein. France announced a plan in April 2026 that covers the computers of its civil service. The European Commission has also set out a technology sovereignty package and an open source strategy. The practice’s preference for open source software and European suppliers sits within that movement.
Zero retention and open weights
Model calls go to inference providers that retain no data, and the models used are open source or open weight. A model that originates in China is run on infrastructure outside that country, generally in the United States. The legal position was checked in 2026 and is noted on the agent‘s page. Neither the European Union nor the United States prohibits the private use of such models through such a service. In Europe the binding constraint is data protection, and export control does not bind.
Owning the data
Session history and everything the models generate are kept locally, so the material can be read and inspected by hand. Holding the material makes a session record and a decision log possible. An account of who decided what can be written only from material the practice has in hand.
Language models only
The practice uses language models and translates what they produce through legacy hardware and its own systems. It uses no generative music model. A generative music model is trained on recorded music, much of it without the permission of the artists who made it. Putting one to musical tasks would take from contemporary creatives, endanger their livelihoods and blur their authorship. A language model writes into a period instrument and works through the practice’s own arrangement.
Small and local models
The practice runs small models for cost reasons, as the direction of travel is towards local models, and a local model runs on the artist’s own hardware. Work that does not need an agent, such as text editing, is already done with fully local models. The decision is both economic and technical.
This system’s command vocabulary is bounded, so small models can operate it. capability gradient exists to keep the system usable across model sizes. When a task is complex enough to need a larger model, the practice calls Mistral through the Mistral API, a European provider. Mistral is also cheaper than reaching a large model through OpenRouter.
Hardware as an asset
This practice owns no expensive artificial intelligence hardware. That is a limitation. It is treated as an asset. Working without such hardware teaches the use of smaller models from practice. It also keeps the tools built this way usable by others who do not own a graphics card.
Efficiency
When a task has to be repeated and does not need artificial intelligence, the practice has an agent build a tool it can run locally. The task can then be repeated without a model, and the agents can work through local toolkits. The reason is practical and economic. A large model on a simple task spends more power than the task needs. Moving repeated work to local tools is the more economical use of energy.
Retained context
The agent’s efficiency comes from what it keeps. The agent retains concepts, and the archive contains context the agent can read for whatever a task needs. Nothing is explained twice. The archive is maintained by the artist and the agents together, and its structure is kept friendly to both. A record written for a person alone would cost an agent more work to read the context out of before it could begin.
Transcription
Recorded audio that has to be transcribed, such as interview material for oral history, is transcribed with a local speech to text model. Transcription needs no expensive hardware, and the recordings are part of the archive. Whatever the transcription gets wrong is put right by hand. A local model can also check the quality of the transcription. The material is therefore not sent anywhere to be corrected.
Where it stands
The practice has not reached fully local operation. Small models run locally for part of the work. The agent runs on a rented server away from the artist’s computers and is reached from a telephone. Running the agentic models locally was costed in October 2026 at a workstation. The terms on which a model trained on the practice’s own languages may be used, shared or sold remain open.
How it started
The arrangement did not begin as it is set out above. It began by working through Claude. The concerns the guidelines name made the migration to the Hermes agent framework preferable. Those concerns are reliance on closed infrastructure and how data is held, above all. The migration is not complete. The practice is not yet entirely free of Claude’s infrastructure. Full independence is expected within the next year of the doctorate.
Weak points of the guidelines
These guidelines are not the same strength throughout. The gaps are stated here. Owning the data applies locally and not in transit. Model calls pass through OpenRouter and Mistral. Whether anything is retained rests on their policy. The practice’s architecture does not control that. That makes the zero retention rule the weakest of the set.
This supply chain is American even where the practice is European. The framework, the routing service and the code hosting are all based in the United States. That position is defensible on licence and on control of the code. It is weaker on jurisdiction. A permissive licence answers none of the questions the guidelines turn on. Disclosure and accountability rest with the artist whatever the licence says.
Compromise and intention
The arrangement is not presented as settled. When an alternative offers the same capability, the practice would move to it. The present setup is a compromise accepted because no European alternative exists. Working through the Hermes agent framework and OpenRouter allows the research to be conducted ethically in the ways set out above. There is nothing of the kind to move to.
What the practice would prefer is an arrangement in which the weak points are answered by European or Dutch infrastructure, which as far as it knows does not exist. It follows the question so that it can move when such infrastructure appears. Beneath that sits the aim of a fully local inference system, which the practice is working towards. The hardware costs a great deal, and experiencing the present arrangement first is the better order. The decisions about what to buy can only be made well after that experience.
What the agents contribute
The practice values one use of artificial intelligence above the others, which is the reading back of the record. An agent working through the archive can see what a practice does and name it, and that is how improvisation became a subject in this research. The agent supplied the occasion and the artist made the recognition, which is the case recorded on Method articulation. This is the contribution the disclosure should carry, because it makes the practitioner’s own knowledge visible.
Contribution to share
The guidelines are a contribution as well as a private arrangement. The text sets out how a practice can work with artificial intelligence on terms it sets, with the weak points named. That knowledge travels to the institutions the practice works with. It belongs with the transferability the doctorate is assessed on.
It also bears on the practice’s own conduct. It is what the practice brings to a body that is setting policy on the subject. One case is the meeting on the ethical guidelines of artificial intelligence use in the professional doctorate on 5 November 2026.