Systems and subsystems: the decision every other number depends on.

1. Two ways of cutting the same plant

Construction works geographically: this deck, that module, the piping crew, the electrical crew. It is the right structure for building, because it matches how labour and materials move.

Commissioning works functionally. Nothing starts up 'by deck'. What starts up is the instrument air system, the seawater cooling system, the crude export system — each a functional whole crossing several areas and every discipline.

So the scope is re-cut. The plant is divided into systems, each system into subsystems, and every physical tag is assigned to exactly one subsystem.

2. Why the subsystem is the unit that matters

The subsystem is what gets mechanically completed, what gets a walkdown, what gets an RFSU certificate and what gets handed over. It is the unit of completion in every sense that counts.

Which means progress reporting is only meaningful at that level. '82% complete' is a useful sentence when it refers to a subsystem someone could plausibly start up. It is a meaningless one when it refers to a discipline across an area.

3. Setting good subsystem boundaries

• Follow function, not geography. The boundary should sit where the system can be isolated and tested independently.
• Cut at isolatable points. A boundary that cannot be physically isolated cannot be tested or certified independently.
• Size for start-up sequence. Subsystems should be small enough to be handed over incrementally and large enough to be functionally coherent.
• Avoid splitting loops. A control loop divided across two subsystems means neither can be fully function-tested alone.
• Mind utilities. Common utilities feed many subsystems and typically need to be commissioned first; model that dependency explicitly.

4. Tag allocation is where the effort is

Every valve, instrument, vessel, motor and cable has to be extracted from P&IDs, line lists and equipment schedules, then assigned to the subsystem it belongs to. On a mid-size facility that is tens of thousands of decisions.

It is also error-prone, because the source documents disagree. The P&ID shows a tag the line list does not carry; the equipment schedule uses a slightly different numbering convention; a revision changes a boundary after allocation has started.

This is precisely the work Deskely's AI drawing parser and AI population compress: extracting tags, surfacing conflicts between documents, and drafting the allocation for engineers to review rather than create from scratch.

5. What goes wrong when it is wrong

1. Phantom completion. A subsystem reports complete because the tags that would have failed it were assigned elsewhere.
2. Untestable subsystems. The boundary cuts a loop, so function testing cannot be performed without borrowing scope from a neighbour.
3. Certificates that cannot be issued. Prerequisites reference tags in the wrong place, so the gate never clears cleanly.
4. Late rework. The error surfaces at walkdown, and re-allocation at that stage means re-running progress reporting across the affected subsystems.

Because everything references it, the tag-to-subsystem map has to be controlled data with change history — not a spreadsheet column. That is the role of the master data register.