How Xamun delivers level five: the Two Minds runtime, the Software Factory, the OS series — and an engagement that starts with your baseline, not our roadmap.
The full argument is on this page. The PDF is the formatted paper, with 6 cited references.
Paper one established where enterprise AI value concentrates: level five of the maturity ladder. Paper two set out the method: inventory the fossil constraints, re-derive the process from the outcome, place every decision, and run it under staged autonomy. This paper answers the question both leave open: execution. Method without a machine dies of its own timeline — re-engineering failed in the 1990s less because the ideas were wrong than because delivery took years. Xamun’s machine has three parts: the Two Minds runtime, a Software Factory that ships a purpose-built application in weeks, and an OS series of vertical starting points — engaged one workflow at a time, against a baseline measured before anything is built.
The constraint that kills level five is not intelligence. It is the time between the design and the cutover.
The re-derivation method in paper two is deliberately public. Any capable team can run the constraint inventory, apply the decision triage and stage autonomy — and we expect many will. What the method cannot supply is speed, and speed is not a convenience here; it is a survival condition. A redesign that takes eighteen months re-creates every failure mode the method exists to avoid: the parallel process runs long enough to win, sponsorship rotates, the pilot never earns a cutover date — the purgatory where over 40% of agentic projects are forecast to die. Hammer’s re-engineering wave stalled on exactly this: right diagnosis, multi-year delivery.
The runtime is the architecture papers one and two argue for, running as software. The deterministic mind executes rules, entitlements, calculations and statutory logic — and because a rule correctly applied needs no probation, deterministic decisions are autonomous by construction from day one. Staged autonomy applies to the judgment and generative classes; determinism is not staged, it is guaranteed. The governed model handles language, extraction, drafting and explanation, bounded and logged, never the last word.
Underneath both sits the graph memory, and this is what makes the event-native design physically possible. A system can only notice that a renewal is due or a threshold is crossed if obligations, dates and precedence exist as first-class objects it can watch, rather than as sentences buried in documents. Trigger inversion is not an interface choice; it is a property of memory.
The obvious objection to paper two is that re-deriving a process and building a purpose-built application for it sounds like the eighteen-month IT programme the paper itself warns against. The Software Factory is the answer. Delivery is specification-first: the re-derived workflow, the decision placements, the rule set and the evidence requirements are captured as a structured specification, and the Factory generates and assembles the application from it — screens, queues, exceptions, approvals, integrations — in weeks.
This is why one-workflow-at-a-time is economically viable for us and ruinous for a traditional integrator: when an application costs quarters to build, only enterprise-wide programmes justify the mobilisation, and enterprise-wide programmes are where redesigns die. When it costs weeks, a single workflow can carry its own business case.
Most engagements do not begin from zero. Xamun maintains a series of vertical operating systems — spanning real-estate portfolio operations, acute-care coordination, port and marine operations, and governed financial workflows — each one a set of already re-derived processes with the constraint inventory done, the decision placements made, and the domain’s compliance obligations built into the rule packs. An engagement in a covered vertical starts from a working system and adapts it to the client’s reality.
The OS products also answer a quieter question: whether the method survives contact with regulation. Rule packs encode the governing frameworks of their jurisdictions, evidence chains are native, and the graph memory’s obligations model is built against the regulator’s definitions — governed activation rather than retrofitted compliance.
The engagement follows the method, because the method is ours too. Week one measures the baseline and runs the constraint inventory on one workflow: the one costing the most time and risk. The following weeks re-derive and build; the Factory ships the purpose-built application; the cutover is scheduled at design time and the parallel process is sunset on schedule. From cutover, judgment and generative decision classes move through staged autonomy at the pace their reversibility allows, with sampled audit throughout — and results are read only one way: cycle time, error rate and throughput against the week-one baseline.
Commercially, the engagement is outcome-aligned: no capital build-out on the client’s side, with fees that move with the workflow’s use. A delivery partner whose revenue depends on the workflow running has already answered the question of whose risk the redesign is.
This model fits an organisation with a P&L owner willing to own the redesign, one workflow with measurable pain, and the intent to decommission the old process. It does not fit an organisation shopping for a chatbot over its knowledge base — that is level two, the rung below the bare model, and no delivery machine improves it. It does not fit a programme that wants a pilot without a cutover date. And it does not fit a buyer who wants the intelligence without the governance: at level five, speed and governance are the same architecture, and we do not ship one without the other.
Bring one workflow and its owner. We bring the baseline in week one, and the application in weeks — not a roadmap.
The complete paper, formatted for reading and circulation, with every figure and the full reference list. We will email it to you.
No. 3 in the Xamun position-paper series. © 2026 Xamun Technologies. Version 1.0, August 2026.