Part of Temporal Constructs Appendix, which contains details behind the research: dates, quotations, descriptions of instruments, and sources.
General definition
The difficulties that this research has met and the difficulties it anticipates, recorded with the same weight as its discoveries.
Relation to the practice
A difficulty is recorded here when it is met, it is given a date, and it is tied to the pages that state it. In a practice-based doctorate the difficulties establish the method as much as the results do, and a difficulty left in conversation is lost.
The entries are in two parts: what the first year has already met, with the instance and what it cost, and what the work has yet to meet, written down before it arrives. An expectation can later be confirmed, qualified or dropped. Where a difficulty concerns artificial intelligence (AI), it is scoped to the models and arrangements that this practice runs, in keeping with the rule kept on those claims.
Difficulties already met
Serial verification
The method states that an agent's report is not evidence until it has been tested against the machine. Checking is therefore work that only the person can do, and the first year met the cost of it more than once. An architecture reported as running under the name Fabric turned out to be largely invented. A success flag was read as proof that messages had been delivered. A fault then took days to trace that a look at the machine would have placed in minutes. The ceiling of the research is therefore the operator’s capacity to verify, and adding agency does not raise it.
Unrepeatable experiment
A finding made in one month cannot reliably be reproduced in the next, because the conditions move. The audit of August 2026 found the running patches differing from the copies held in version control. The models on offer were withdrawn and replaced at the routing service, and the instruments age by the year. The research therefore cannot offer repetition as its warrant. It offers inspectable conditions instead: the recorded patches, the transcripts, the artefacts and the decision record.
Attribution along a chain
A decision passes through more than one hand, so the same decision can be recorded three times with a different initiator each time. The year supplies the instances. One agent deleted the coding of the system. A second rebuilt it while mixing three generations together.
The sixteen modules were then rebuilt, with each result read off the machine by Gardner. The coding therefore has to tolerate ambiguity. Traced cases, not totals, will carry the argument.
Failure of a dependency
The method rests on things that stop on their own, without warning. The list of them is the following:
- the framework
- the routing service
- the maintained port of HMSL
- the build environment, which needs a current system
- the MIDI interfaces, which can leave production
- the machines, made between 1987 and 1991
Three instances stand for it. A corrupted memory card lost its contents. A file system discarded every write when it restarted. An agent deleted a codebase. Three further instances were met in the instruments. A synthesiser module faulted on a missing buffer flush. A video mixer was locked out of its own memory. A tape deck had a sensing foil that perished. An account written for a human successor would reduce the exposure, and it does not exist yet. Two onboarding documents exist for agents.
Archive lag
The months with the most engineering produced the thinnest accounts. The reason is arithmetic before it is discipline: once the agent was doing the work, the work stopped passing through the conversations that the accounts summarise. The session history is therefore the largest single asset of the project and the least usable. Turning it into an account of who decided what is work that the research still has to do. The difficulty is also evidence for the thesis, because it shows where the decisions went once they were shared.
Comparison with several variables
The premiere and the run at MU differ at once in a number of ways. They differ in the narration, the audience, the venue and the screens, and in the code, the months of practice and the capability of the agents. The strongest evidence of the research is also its most entangled. What the research can offer is a date. Each of those differences is noted, and the decision record shows what was proposed, refused and rebuilt between the two runs, even where the effect cannot be separated.
Anticipated difficulties
Training rights
Fine-tuning a model on KML and HMSL is framed as preservation, and the artistic and academic intentions are clear. A model trained on the corpus of a language raises questions that the research has not opened. They are what its outputs may be used for, and the terms under which any of it can be shared or sold. The partnership with the makers of those systems helps, and it is not a licence. This is to be settled before the corpus is assembled, and not left until the archive is handed on.
Artist’s fluency
If agents write the tools, the question is what happens over four years to the person’s own command of them. The assessment expects reflexive capacity, not confidence. The practice already has the better answer: the system should remain one that its operator understands from end to end. The move to plain Pure Data is that answer in practice. A dated statement of where the artist’s own hand stays irreplaceable would settle the question before it is asked.
Legibility to a committee
The claim has to be assessable by people who do not know what a server for tool access is, so the explanation work of October 2026 was the right instinct. The opposite failure is a defence that demonstrates machinery and never states a finding. The account already offers, in plain language, the sentence any examiner can use: an agent’s report is not evidence until it has been tested against the machine.