Part of Temporal Constructs Appendix, which contains details behind the research: dates, quotations, descriptions of instruments, and sources.
General definition
The working method of RETRIEVAL: the loop in which an agent writes code, reads the picture it produces and tries again, and the conditions that make the collaboration observable.
Relation to the practice
The method of RETRIEVAL was worked out in 2026 and is recorded here. The account also gives the reasoning behind using a language model and the constraint that makes a model’s contributions visible. What the work is meant to produce, and the material it compares against, are on RETRIEVAL, the proposal.
Automated loop
Work in May 2026 built the loop that RETRIEVAL rests on. The agent writes KML, sends it to Trip-A-Tron, reads the resulting picture and tries again, without a person in between. Code, visual output and the reasoning between them accumulate as a body of examples. Deterministic pixel analysis reads the picture, with a small vision model kept as a fallback. The editor’s own text is read by matching its six by six pixel font. By the end of the month the agent could reach all 128 events of the editor unaided, and a standalone tool for the pixel work had been written for it.
The stage runs under emulation, which makes continuous operation practical. The working arrangement was reframed in July 2026. The file that contains the KML code of a composition is the place the work is authored, and the emulator is only where it is written down. The agent is the file loader the original system lacks, and it types the code keystroke by keystroke. Before any input was tested, a protocol was agreed for checking the agent’s reading of the screen against a known reference. The earlier work had been done without vision, and its findings were still hypotheses.
Rationale for the model
The aim is to locate the boundary between human and machine fluency in creative programming. A model reaches a fluency in the syntax and logic of these languages that would take a person years. The human remains necessary for creative intention, validation and direction. The question is one of division, which tasks belong to the machine and which remain with the human. These languages were written for humans, and a practitioner can spend years developing fluency in them. A model reached the same fluency after a single day of preparation.
Method of the loop
The work uses artificial intelligence in its making and never in its execution. A session can begin with the human describing a desired visual behaviour in plain language, and the model proposing KML code to achieve it. The human then tests it on the actual hardware. The two refine it together until it performs what was intended, or until it performs something the two did not intend. Sometimes the model produces finished code, and sometimes the human reshapes a draft, so the boundary shifts. The research tracks that boundary.
Once the work is complete, the instruments run entirely on legacy hardware. An Atari ST runs Trip-A-Tron, and a Macintosh runs HMSL. No artificial intelligence is involved in that execution, a choice grounded as much in sustainability as in artistic intention.
Constraint as research condition
To test what a model can do creatively, the space must be contained enough for its contributions to be visible and measurable. The strict form of a haiku makes a machine’s feel for poetic form observable, showing where it succeeds and where it fails. KML and HMSL have a defined and finite possibility space. This makes it possible to observe where creative decisions are made, where limitations are met, and how they are overcome.
A body of work by human practitioners alone already exists, so collaborative output can be placed directly alongside that historical record. That comparison turns on what a model changes in a repertoire that humans built. Max and Pure Data survived the period because they lowered the barrier to entry. KML and HMSL ceased to be used, although they could express more. The same difficulty that kept them from being adopted also kept their possibilities from being explored. RETRIEVAL turns on whether a model can make a retroactive difference: whether these languages can be activated as living tools, and whether what was latent in them can be reached.
Unforeseen
Here the unforeseen situations, sometimes difficult, sometimes unexpectedly interesting, will be described as the project runs. Many of these unforeseen situations have already occurred and have been recorded. They will be added in the time to come.