AIOps for AI-native development
DevOps made software releasable. AIOps makes minds releasable.
A build pipeline turned releasing software from an event into a habit: every change kept, tested, gated and shipped often. The MDLC is meant to do the same for expertise. What moves through it is not a binary but a mind, and what it must never ship is a confident guess.
“This ie the evolution of all I have learned about DevOps and Platform Engineering so that a multi-deminsional synaptic pipeline may be established to create the rapid releases of fit for purpose sovereign Ai.”
John David Marx, 28 September 2026 — the charter, verbatim
Practice by practice
What replaces what, and what runs it.
Each row is a practice DevOps made ordinary, the practice that replaces it when the thing being shipped is a mind, and the machinery that performs it in the MDLC today. The figures are live from the ledger.
| DevOps | AIOps in the MDLC | Today |
|---|---|---|
| Source control every change kept and attributable | The evidence vault. Sources are kept as the bytes that arrived, with who fetched them, from where, when, and with which tool at which version. Evidence cannot be edited or deleted; the database refuses. | sources |
| Code review a second person approves | Admission. Machines may gather and grade. Only a named authority may admit a claim, and every admission records whose authority it was. | admitted refused |
| Unit tests small checks that run alone | Atomic processes with controls. Each process declares cases whose answers are known before it runs. If a control fails, the harness refuses to report a score at all. A check that passes on nothing is not a check. | every process, every run |
| Build one command makes the artefact | Mint. A Mind is a contract, a frozen set of admitted claims and a base model, versioned. Its claims are frozen when it is minted, so it answers for what it said even after the corpus moves on. | versions minted |
| Release gate nothing ships red | The promotion gate. No evaluation, no promotion. The control must hold, the mean lift must be above zero, the run must be on the model the Mind was minted with, and a single invented citation retires the run. | Minds serving |
| Rollback the last good version returns | One version serves. Promotion is exclusive. What it displaces is retired, not deleted, and can be measured again at any time. | retired |
| Canary release a change meets a few first | The canary arm. The Hebbian ranking term runs for one arm of questions, assigned by the question itself so the same question always meets the same arm, and the arms are compared by whether the work that followed was accepted. | Hebbian edges |
| Monitoring production is watched | The Learn stage. Every question is logged with what came back and whether the work was accepted. That record is what the Hebbian edges strengthen on and what promotion is judged by. | logged judged |
| Scheduled operations the pipeline runs unattended | The night runner. Deterministic work under the operating system's own scheduler: the harness, the quality suites, the evaluations. Every task has a deadline, the queue re-seeds each cycle so a service that dies at one in the morning is found at one in the morning, and work that needs a model is parked with its reason rather than failed. | every night |
| Dependency policy know what you run on | No rented inference at runtime. Models run on the Foundation's own hardware, and the runtime is measured before it is chosen. | measured below |
AI-native development
The software is designed from evidence too.
The same loop is turned on the Inference Desktop itself. Before a line of its runtime was written, corpora on orchestration and on desktop design were gathered and admitted, and they changed the design. Three decisions came straight out of them:
- A bus, not a pipe. A pipe has exactly one reader, and a room of Minds has many listeners. So broadcast and queueing are separate types, chosen on purpose.
- A lease, not a spawn. A job is leased, and returns to the queue if its worker does not come back.
- A named foreground, not a boolean. Exactly one holder at a time, named, and everything else waits on an event.
That is what AI-native means here: the design decisions of the software cite the evidence they came from, and the evidence was admitted by a named authority before it could decide anything.
| Machine | Runtime | Decode, 3B 4-bit |
|---|---|---|
| Kepler · M4 Max | Ollama (llama.cpp) | 161.4 tok/s |
| Kepler · M4 Max | MLX | 183.7 tok/s |
| Spindle · M5 Max | MLX | 211.0 tok/s |
Same prompt, greedy decoding, measured on the Foundation's own machines, 29 September 2026. From inference-desktop/docs/can-we-train-a-mind.md.
The person in the loop
Knowing when to step in, and when to stay out of the way.
“The ultimate Co-Pilot is one that accelerates the work by knowing when a human-in-the-loop is required, and when we are just in the way!”
John David Marx, 28 September 2026
AIOps here is not the removal of people. It puts a person exactly where judgement is irreplaceable, which is deciding what a mind may learn from, and it automates everything around that decision. In the week of work that preceded the MDLC, three faults in one night were caught by a person looking at a picture. No automated measure caught any of them, and several reported that everything was fine. That is the argument for the gate. AI augments the people who use it; it does not replace them.
One cycle, end to end
A narrated film of one Mind going once round the loop: gathered, admitted, minted, evaluated, promoted, and asked. In production.