Every architecture diagram is a claim. This one shows its working.

Six months on, half the diagram has quietly stopped being true.

A Maximo architecture drawing is a set of claims — which products they run, what talks to what, what IBM still supports. Every one of them was true the day somebody drew it.

Then a support line ends. A product stops being available standalone and becomes an add-on inside another one. A page is re-published with different wording. None of that changes the picture on the wall, and none of it announces itself: the diagram looks exactly as correct as it did the day it was approved.

The tools that draw these pictures cannot help, because a box on a canvas has no idea where it came from. Ask any of them how do you know that and there is no answer in the file.

That sentence invalidated an approved design decision. Atlas found it because it re-reads the page, not the drawing.

Four states. There is no fifth, and keeping it that way took discipline.

Most tools have one state: on the diagram. Atlas has four, and the difference between them is the entire product.

Across all 92 evidence-bearing records — elements, relationships and constraints

	State	What it entitles a reader to	Now
●	Documented	IBM states this, and the sentence where they state it travels with the record — along with the page it came from and the date it was read.	75
◇	Customer-declared	Something the customer says they run. Never an IBM claim and never dressed as one. It stays in the browser and never enters the catalogue.	10
○	Inferred	Our architectural judgement. The sources were searched, and what was searched is recorded on the record.	7
?	Evidence gap	We do not know. Drawn rather than hidden, because a silent gap reads to every reader as a fact.	0

The states never blur into one another, and there is no comfortable middle one to promote a weak claim into. That is deliberate: the moment a diagram can say sort of documented, every figure on it becomes worth arguing about, and the person holding the drawing loses the one thing they came for.

Three readings, and only one of them needs a plan.

1. 01

   What they already have

   Record what a customer runs and Atlas answers without being told anything else: what is expiring and when, what has been deprecated or renamed underneath them, what cannot be evidenced, and which constraints bite. No destination required — this is the resident architect's whole job.

2. 02

   Where they are going

   Add a target release and the move appears, scoped to the elements they actually run rather than the whole product line: what carries, what survives but changed underneath them, what leaves, what arrives. The first reading does not go away — a thing that is expiring is still expiring on the way to 9.2.

3. 03

   What moved since you last looked

   The catalogue keeps an append-only record of what every claim's evidence looked like at each build. Cross that with the date the estate was recorded and Atlas can say: IBM changed this on the 14th, you recorded it on the 2nd, here is the excerpt as it read on both dates. That is the reading nobody can shortcut, because what it accumulates is time.

How the record is kept current — the method in full →

Every one of these readings states the window it looked at. Atlas will tell you it has not been watching long enough to answer rather than report a quiet period it never observed — because on this particular question, a confident nothing is the most expensive answer a tool can give.

Estate	Release	Elements	Needs attention	Nearest dated concern
Northern Water	8.11	19	6	2026-04-30 · 108 days past
Metro Rail	9.0	24	3	2026-09-12 · in 27 days
Coastal Ports	9.1	16	2	2027-02-01 · in 169 days
Regional Power	9.2	21	1	none dated

What Atlas deliberately will not do.

A validator that knows the edge of its own evidence is worth more than one that always has an answer. Each of these is stated inside the product, not buried on a page like this one.

• ○
  It refuses verdicts on the classic line.Maximo 7.6.1 is carried as documented history, but no compatibility, prerequisite or lifecycle verdict is offered for it, because no captured source backs one. A 7.6.1 → 9.2 read produces the deltas and then declines the question.
• ○
  Fix packs resolve on the MAS line only.959 builds across 18 applications, 8.4 through 9.2, taken from IBM's own releases index. The classic line publishes elsewhere, so Atlas reports we did not look rather than reporting those builds absent.
• ○
  It does not read your system.No agent, no CMDB connector, no scraping a production Maximo. An estate is recorded by a person, and it stays marked as that person's word for as long as it exists.
• ○
  It will not write the recommendation.The proposal draft assembles the evidence and marks every judgement for a human. What to advise depends on appetite, timeline and budget, and none of those are in the catalogue.

• The resident architect

  Holds the same estate for years, and is the person everyone asks when something breaks. The problem is never the work — it is that IBM keeps changing things under a customer nobody has looked at since autumn.

  What changes: Monday morning starts with a list instead of a hunch. Atlas re-reads the sources, crosses them against the date the estate was recorded, and hands over what actually moved — so the first conversation of the week is about a decision, not a search.

• The bid team

  Has an afternoon to turn a discovery call into something a client will read, and no time to chase down where each claim came from. So the claims go in unsourced, and the review meeting finds that out.

  What changes: The draft comes out of the record rather than out of memory, with every fact already carrying the page it came from. Atlas marks the judgement calls for a human and refuses to write the recommendation — which is the part you were being paid for anyway.

• The design authority

  Signs off other people's architecture and carries the risk when it is wrong. Asks one question all day long, and mostly receives a reassuring tone of voice in reply.

  What changes: That question gets an answer instead of an assurance: the sentence IBM wrote, the page it is on, and the date it was read. Where nobody can answer it, the drawing says so on its face rather than leaving the reviewer to find out later.