Part of Temporal Constructs Appendix, which contains details behind the research: dates, quotations, descriptions of instruments, and sources.
General definition
The practical detail of the system: the parts and their names, the routing, the path a command takes, the agent roles, the image and light it drives, the picture chain that reaches period screens, the text interface, the reconstruction of the period machines, and the clock that holds the whole deployment in time.
Relation to the practice
These are the framework’s engineering parts, kept together so CONDUCTOR can describe what the system is without the wiring. They are also where the period hardware is met most directly. The image chain, the screens and the clock all had to be reconstructed. The installation these notes belong to is described on the installation page.
Component names
Its hierarchy is fixed. Each term is used in the same sense throughout the archive. The table below defines each term.
| Term | Definition |
|---|---|
| Module | One of sixteen independent software units, each an ALSA client on the Raspberry Pi. A module cannot take the others down if it fails. The ceiling of sixteen follows MIDI’s sixteen channels. |
| Timeline | A single global timeline with sixteen group tracks, one per channel. Gradual parameter movements are scheduled data on the timeline, not loops that an agent runs. |
| Controller | A software component inside every module. Its knobs send control change messages, which are mapped to functions. |
| Snapshot memory | Part of every module. A snapshot is a stored set of values, recalled by a MIDI program change message. Snapshots are stored as plain lines of text, one file per module. |
| Sketchpad | Part of every module. A working space for material not yet stored as a snapshot. |
| Device | A patch built in Max or Pure Data that translates controller values into instrument-specific messages. |
| Function | Something built into the hardware, or into a Pure Data or Max patch, that can be reached with one control change message. |
| Workbench | A single shared place with direct MIDI and sysex access to the hardware, available to every agent. It is not part of any module. The system uses one workbench at a time. |
| Conversation layer | Where intentions are discussed. Telegram carries it, with one group per module and one for the conductor agent. |
| CUE | The shared command language, and the terminal execution layer. A language model turns an intention, spoken or typed, into one of about fifty tightly defined commands, so that it does not have to reason about low-level detail. |
The ceiling of sixteen is the ceiling of a MIDI system’s channels. The station map settled at eight in July 2026: the Kurzweil K2000, the two lighting stations, M, Trip-A-Tron, the Lexicon effects, the Edirol V4 and a utility station. From that month Pure Data was the only gateway from the modules to the hardware, with a watchdog and an alert to a telephone when it stopped.
Routing follows a small set of rules. A channel number addresses a module. A single wire command joins two modules, as a patch cable would, and each cable behaves like a subscription. Every element is a module with named ports, including MIDI inputs and outputs, DMX lighting and synthesiser control. Modules know nothing of one another beyond these addresses, so a failure stays local.
CUE was designed in June 2026 to run in three modes. A conversational mode serves rehearsal and set-up, a command mode serves debugging, and a performance mode runs stored macros within about ten milliseconds. From the performance audit of July 2026, dispatching a single cue takes under one millisecond. The overall shape has two kinds of caller at the top, sixteen sealed modules at the waist, and the legacy instruments at the bottom.
Command path
A cue is the only way in, and it never reaches an instrument directly. It addresses the controller inside a module, and the controller is the gatekeeper. The controller is the contract a person or an agent has to understand to know what is possible. After the module comes the device, and after the device the hardware. Those two together set what is possible in the first place.
flowchart LR CUE["CUE<br/>the command language"] --> MOD subgraph MOD["The module"] CT["Controller<br/>the gatekeeper and the contract of what is possible"] end MOD --> DV["Device<br/>with the hardware, sets what is possible"] DV --> HW["Hardware<br/>the instrument"] CUE -. never.-> HW
Shape of the system
Two kinds of caller reach the system, a person and an agent, and each has a way in. Both use CUE, which addresses the modules and the timeline. A person can use CUE directly or delegate: what they want is said to the agent, and the agent issues the commands. Talking is all the conversation layer does, and delegation is what that conversation is usually for.
A person has a second way in that does not pass through CUE. The map a module’s controller has can also sit on a physical surface, the Faderfox EC4, so the same controls can be played by hand. Inside each module sits its controller, the gatekeeper and the contract of what is possible. The device together with the hardware sets what is possible. The workbench is the exception and the agents run it: it reaches the hardware directly, and no person uses it in this arrangement.
flowchart TD H["Human operator"] A["Agent"] CUE["CUE<br/>the command language"] TL["Timeline<br/>recalls snapshots in order"] subgraph MODS["Modules, one per MIDI channel"] CT["Controller<br/>the gatekeeper and the contract of what is possible"] end FF["Faderfox EC4<br/>the controller's map, in the hand"] DV["Device<br/>with the hardware, sets what is possible"] HW["Hardware<br/>the instruments"] WB["Workbench<br/>direct access, agents only"] H ---|conversation and delegation| A H -->|CUE| CUE A -->|CUE| CUE H -->|by hand| FF FF -->|mapped to the controller| MODS CUE --> MODS CUE --> TL TL --> MODS MODS --> DV DV --> HW A --> WB WB --> HW
Agent roles
Different agents appear in the account, and the word “agent” on its own is ambiguous. Each is distinguished here by role. The table gives the role and the work it does.
| Role | What it does |
|---|---|
| The operating agent | Operates the system through CUE. It did not exist at the premiere of ALTERITY in January 2026. It first ran the system in full at MU Hybrid Art House in August 2026. |
| System-level agent (conductor agent) | Manages the system as a whole. |
| Module agent | Operates within one module. |
| Agent at the workbench | Works on the shared workbench, with direct access to hardware. |
| Builder agent | Constructs a module and never performs. |
| Runtime agent | A smaller model that performs and never writes code. |
| Nexus | The guide agent of the explainer. It is not the conductor agent and can take on any agent’s task, labelled by context, in order to demonstrate parts of the system. |
A second architecture proposed in June 2026 divides the work between two models in a similar way. A cloud model builds parameterised instruments in advance, acting as luthier, and a small local model performs. The local model inhabits those instruments by supplying only target values while the work runs. It is the redistribution the research argues for.
Image and light
Beside the sound modules, CONDUCTOR addresses image and light. These are a MIDI-controlled video mixer of the kind used in VJ practice, and moving heads and RGBW (red, green, blue and white) LED lights through DMX. Agents programme these blind like the rest, and where a camera view is provided they work from what they see. That brings multimodal vision into the system. The moving heads bought in June 2026 were brought into the system in July, and the practical detail of driving them was worked out then.
A converter that doubles control values narrows the range that will strike a lamp. Cheap colour-changing fixtures step visibly at low levels where the mechanical dimmer does not. Additive colour has to be approximated for fixtures without a white emitter. A network interface moves the lighting out of the computer and onto the Raspberry Pi.
The camera is a low-cost network camera read as a video stream, and its deeper controls sit behind a proprietary interface. A later firmware closes that off. Work on the patches in August 2026 added two rules of that layer. A moving head loses its movement at small ranges when the computed motor speed crosses into a reserved zone of the fixture’s speed channel.
The value therefore has to be held inside its valid range. The discharge lamps of the period fixtures and the white channel of cheap colour-changing fixtures must take separate roles. The white is kept on an address of its own so it can carry a cue by itself. In September 2026 the fixture’s channel layout was rebuilt as a browser viewer with eight faders and a live input for the lighting protocol. A way was outlined for a moving head to sense what it is pointing at with a wireless light sensor. The causes of a head that stops answering intermittently were listed as faults in the data link, protocol auto-detection, old firmware and failing components.
Picture chain
The installation shows period pictures on period screens, which means reconstructing an analogue chain that current computers no longer provide. Composite output was brought back on a headless Raspberry Pi 4 running a current system. A reference document was written for it. The configuration, the television standard, the known faults and the composite output’s disabling of the display output all have to be settled together. Enabling the composite output lowers the graphics clock slightly, while the processor speed the MIDI work depends on is untouched.
The wiring of the four-contact plug that combines sound and picture is not the same on every device. A cable made for a camcorder has the picture on its red plug when it is used with the Pi. At the far end of the chain the material has to meet older equipment on its own terms. A video file was encoded as a PAL DVD for a DVD player, and the video capability of a SCART connection on a domestic player was checked. The Atari core on the MiSTer was set to start straight into its application. That took finding three faults in its configuration and an invisible carriage return in a setting.
Text interface
The system’s own state is shown on consumer cathode-ray screens, so the interface is bound by what a composite video signal can carry. The design settled in September 2026 works to a logical canvas of 360 by 288 pixels doubled into a 720 by 576 interlaced European picture. The interface uses an eight by eight character grid and a near-monochrome palette, and reserves a single colour for alarms. The character set is that of the original IBM personal computer, and a safe area keeps clear of the screen edges.
One Raspberry Pi 4 cannot drive the three screens the installation uses. It has a single composite output. The design is a Pi 4 hub with three small Pi 1 machines as satellites. Those are provisioned by a script that is reviewed and not edited live, because the agent putting them in place cannot see the screen.
It began with the hub’s own output, and its chain needed its own resolution. The successive Pi models differ in how they produce composite video, and the third had to be set up to show the fourth’s terminal over a network connection. Subtitles for the 240-line picture needed bitmap fonts and integer scaling to stay legible on a modern machine, and a true fifteen-kilohertz signal had to be found for the screens. A Pi whose picture the composite output had taken away was recovered by editing its configuration from another computer.
Machine reconstruction
In the third module the Atari ST is a hardware reconstruction, and the MiSTer platform reproduces the original’s timing and signal characteristics. Analogue video and MIDI therefore behave as they did on current production hardware. They are free of the degradation that ageing components bring, and cost less energy than the original machines. That reasoning runs through the whole system, which prefers instruments small enough to be understood over emulation that hides its own behaviour. The agents that operate it work in a group per instrument, so a machine’s state stays legible while more than one is driven at once.
Clock and timecode
The chain that keeps the deployment in time can be read as one line. Every point on it is a place where time is carried, converted or consumed. The figure below gives that line.
flowchart LR TP["One strip of tape<br/>narration on one track, SMPTE timecode on the other"] --> MS["Yamaha MSS-1<br/>reads the timecode and converts it to MIDI"] MS --> AMT["Emagic AMT8<br/>routes the MIDI between the machines"] MS --> SC["Scene changes and system states<br/>across all three towers, from its own timeline"] AMT --> MX["Max programming and the timeline"] MX --> V4["Edirol video mixers"] MX --> LT["The lighting"]
One strip of tape is enough to keep the deployment in time. A reel-to-reel machine puts the narration on one track and SMPTE timecode on the other. A Yamaha MSS-1 reads that timecode and converts it into MIDI, firing scene changes and system states across all three towers from its own programmable internal timeline. An Emagic AMT8 routes the MIDI between the machines. The Max programming and the timeline send MIDI to the video mixers and to the lighting.
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.