What I see as the biggest issue facing wealth-oriented, cost-cutting capitalism is what happens when a business can book a saving today and leave someone else to discover the cost years later.
Cancel the afternoon an engineer was going to spend tracing a circuit and updating the drawing. The saving appears in this quarter’s accounts. Where do we record the fact that the next engineer will have to open the ceiling to find out what that circuit serves?
Look at what happened during the FAA ground stop. Several sites were trying to connect, and older equipment became unstable.
Now suppose I give an AI the job of planning a network change. I give it the target design and the new equipment specifications. It does an excellent job with everything I put in front of it. But did I give it the AS-IS?
That is the solve for Mush. Before asking anyone to design the change, document and verify the AS-IS.
Buildings were where I first learned the challenge of as-builts. A plan tells you where the conduit was supposed to go. When you return to change the building, you need to know where it actually went.
At National Data Corporation (NDC), I encountered the same problem with networks. Before planning a change, I needed to know how the network was actually connected. I was 24, maybe 25, when I learned the seven Ps at NDC: piss poor planning provides piss poor performance. Planning starts with the system we have.
Give the designer the as-built diagram, the specifications and running configuration of each machine, and a record of when somebody last checked them. Follow the new traffic across that drawing. Which old box must answer when several sites reconnect at once? Leave that box off the drawing and even a careful analysis can miss the very limit the change needs to respect. That is the question the outage raises for me. Its cause is a matter for the incident record.
Take a rack installed before its equipment could report identifying tags to software. A designer sees two network connections on a screen. Do their cables run through the same conduit? Does one breaker feed both power supplies?
An AI can record what a technician traces, compare it with the drawings and flag a discrepancy. Reading the software inventory alone cannot tell it where somebody actually ran the cable. If nobody has looked since the last change, the honest entry on the drawing is “unknown.”
In my piece on the FAA ground stop, I said to start with the router manuals. How much traffic can this model carry? How many new connections can it handle when several sites reconnect at once? Those are different limits.
Give me the as-built network diagram, with each machine tied to its model, software version, settings and physical connections. Then I can follow the traffic to the next machine and ask what it can absorb. Where the specification runs out, measure.
The same goes for servers and data stores. How much space can this database actually use, and how much is free when the overnight job starts? How many CPU cores can that job use while the shop is still taking payments?
The server inventory gives me the installed capacity. I want the allocation and the measurements from a busy night. Otherwise an AI can design a job that fits the machine on paper and still runs out of space halfway through.
My concern for the future is what happens when one AI fills a gap with a plausible assumption and the next designer inherits it as a fact. A guess about a router setting can travel through three generations of drawings without anyone opening the configuration. That is how we pass the mush on.
Consider the change freeze before Thanksgiving. PowerReviews scheduled its 2026 feature freeze to begin at close of business on November 20. Azure engineers have described a December pause in feature rollouts.
A freeze buys the team time. What do we do with it? I would take the drawing into the equipment room and check it. Otherwise, when changes resume, the AI still has the same drawing with the same missing router. Give it mush for the AS-IS and it can design better-looking mush. PowerReviews holiday schedule · Azure engineers’ outage review
Software, software, software. Give an AI the instruction “authorize every sale online” and it can draw the terminal, the bank and the arrow between them. What does it know about that arrow? Has anyone told it whether there is a telephone in the shop?
I remember an INTUG discussion in Budapest about Russia and Eastern Europe. I had a mobile phone with me. The figure that stayed with me was a hundred people to one fixed line.
The regional figures give that memory some perspective. In 1989 Russia had about eight people per line. Albania had about a hundred. Even at eight, the shopkeeper with a customer at the counter had a practical question. Where could I make the call?
The bank had already decided whom it would trust with a line of credit. The merchant had a floor limit. Within that limit the shop could accept an eligible sale under the card rules. Above it someone had to obtain authorization. The floor limit was a business decision about how much risk to carry when asking the bank took time and money.
When did asking on every sale become ordinary? Visa dates its BASE I electronic authorization system to 1973. By November 2008, Visa’s US supermarket incentive programme specified a zero floor limit. Those dates tell us when a network could carry an authorization and when one US programme expected one on every purchase.
The change at any particular counter depended on what the merchant could connect to and at what cost. The internet, 3G and the small cells that extended wireless coverage belong to that later story. Telecom engineers built the connection that a designer today can so easily draw as an arrow.
What happens when the answer never comes back?
Suppose the bank approves the sale just as the connection drops. The cashier sees a spinning wheel. The customer is standing there. Should the cashier try again? Before I let an AI design that interaction, I want it to account for four things.
- Networks. Can the request get out and the reply get back during the time a customer will wait? What have we measured in that shop on a busy afternoon?
- Servers. Who made the decision? Can the terminal ask that server for the outcome of the first request?
- Clients. What can the terminal remember through a power cut? What does it tell the cashier when the outcome is still unknown?
- Data stores. Where is the durable record of the approval? Can a second request be matched to the first before the customer is charged twice?
I would start with the people who run the network, maintain the servers, service the terminals and reconcile the books. Ask the cashier what happened the last time the screen froze. Put those answers alongside the measurements. Now the AI has something to design from.
That instinct turns up in our story of Alexandria. Mary Magdalene looks at the goods and asks about the people who brought them there. At the counter, I want to hear from the cashier for much the same reason. She knows something the drawing cannot tell me. That is where Mush begins for me. Ask her what happened. Check it against the record. Then leave what we learned where the next designer can find it. That last step is what I mean by Design the Present as the Past of the Future.
I have a lasting reverence for Fred Brooks and The Mythical Man-Month. Brooks counted time that a project schedule can easily hide. Someone who knows the work has to bring each newcomer up to speed. I think of that when an AI produces a design in seconds. Who spent the afternoon in the equipment room finding out where those two cables actually go?
That is the wisdom I take from Mush and DPPF. Before we redesign the system, trace the cables. If they share a conduit, put that fact on the drawing. Date it. Leave the next engineer a record worth inheriting. Whether that shared conduit figured in a particular outage is a question for the incident record.
Finding the truth is not political.
If You Re-engineer Mush · Design the Present as the Past of the Future · EMV
Sources: EBRD, Transition Report 1999, fixed-line penetration · VisaNet history · Visa USA operating regulations, November 2008.
Philip Andreae, in dialogue with ChatGPT. November 22, ’26.
