The Drawing Says 600 CFM. The Data Says 400.

A commissioning director at a university told this story to a smart buildings newsletter this week, and it has been rattling around my head since.

A project team is planning work on a space. They pull the spec for the VAV box that serves it. The spec is fifteen years old and says the box is rated at 600 CFM, so that's what they design against.

The building has six months of trend data. In six months, that box has never once produced more than 400 CFM.

The drawing wins. It usually does.

I read that and thought: that's it. That's the whole argument for monitoring anything, and it has nothing to do with catching failures.

The Record Dies at Turnover

The newsletter was Nexus Labs, and the piece was called Substantial Completion is Fake. It was written by a former controls contractor named Brad, and the reason it's good is that it's a confession rather than a thesis.

He was turning over a hospital wing, months late. His team was fighting a control valve for a pharmacy that wouldn't hit the required airflow inside the required startup time. Nights piled up. They changed who knows how many settings across the air handler and the terminal unit until it finally worked. The commissioning agent, who was supposed to be off the job months earlier, gave a thumbs-up remotely while watching numbers on a graphic. Staff were already moving in.

Then came the last deliverable: as-builts. Nobody had time or budget to review what had actually been changed, so the PDFs went to the GC without the story. The GC handed them to the owner. Boxes checked. Everyone went home.

"I'm sure a hospital technician serviced that valve within a couple of years. If he even found our as-builts, they were useless."

A general contractor gave him the line that ought to be printed on a poster somewhere:

Construction busts its ass for the first 90% of the job, then dumps the turnover over the fence and walks.

I don't work on hospital wings. I monitor a sump pump in a basement and forty devices in a small building. But I have walked into both of those situations, and I recognized the aftermath immediately, because the aftermath is what my customers hand me on day one. There is no record. There has never been a record. Whatever record once existed described a building that no longer exists.

It Gets Worse After Turnover

Two more stories from that same piece, and they're the ones that actually changed my mind about what I'm selling.

The first: an energy manager at one of the giant tech companies said that as her operators tune the building for energy savings, the building slowly morphs, and nobody updates the record. Walk the site and it looks nothing like the drawings.

Notice that the changes were good. This is undocumented improvement. The building got better and the paperwork got wronger, at the same time, for the same reason.

The second story is what happens next. A healthcare operator ran an upgrade project. During the upgrade, the control logic was copied and pasted back to its original design intent.

Two years of tuning, gone. Erased by a restore. All of that knowledge had been living in one controller's memory and a couple of people's heads, and a well-meaning upgrade wiped it, because there was no record anywhere that said this deviation from the drawing is deliberate and it took us two years to find.

The Signal Was Already in the Wire

Here's the other thing that landed this week, from a completely different corner of the industry.

Renesas published work on detecting belt damage in a motor system. Tears in a drive belt, caught before the belt lets go. The interesting part isn't the detection. It's what they detected it with.

No sensor. Nothing added to the machine at all. The model reads the voltage and current feedback that the inverter is already producing to run the motor.

• 9 KB of RAM. 4.8 KB of ROM.
• 98% cross-validation accuracy.
• Stable from cold start through thermal equilibrium.

Sit with the size of that for a second. The evidence that a belt is tearing has been traveling down a wire that exists in every one of these systems, for as long as those systems have existed, and the amount of computation required to hear it is nine kilobytes.

The signal was never the problem. Nobody was writing it down.

I have a personal stake in that finding, because it describes my own system almost exactly. There is no sensor on my sump pump. There's a fifteen-dollar smart plug between the pump and the wall, and everything I know comes from how much current the motor pulls and for how long. Running, not running, running too long, cycling too often, all of it inferred from draw. I did it that way to be cheap and to avoid drilling into somebody else's equipment. It turns out to be the same idea Renesas is publishing: the machine is already telling you, in the electricity.

What I've Actually Been Building

Put the two together and something uncomfortable falls out.

I have spent a year and a half describing my work as monitoring. Monitoring means alerts. The pump ran too long. The compressor is short-cycling at 3 AM. Here's a text message.

Alerts are a commodity. Every prospect I talk to has already been pitched alerts by somebody with a nicer logo and a bigger sales team, and they've mostly learned to ignore them, because most alerting systems produce alarm fatigue faster than they produce insight.

But go back to the VAV box. Six months of trend data beat a fifteen-year-old drawing. Not because the data caught a failure. Nothing failed. The data simply knew what the equipment does, and the drawing only knew what somebody once expected it to do.

That reframing has teeth, and I'm annoyed it took me eighteen months and somebody else's newsletter to see it.

My sump pump has a nameplate. The nameplate is a fact about a category of pumps. What I have instead is how this pump behaves on a dry week in February versus a wet week in April, how long a normal cycle runs, how many cycles an ordinary storm costs, and how all of that has drifted across eighteen months. That is a different kind of fact, and it's the kind you can make decisions with.

A vision vendor published a line this week that says the same thing from the other direction: drift is not a failure, drift is a normal consequence of production change. Your model doesn't decay. Your building changes and the record doesn't. Same disease as the morphing building. Same cure.

Four Things a Record Does That an Alert Can't

It settles arguments. A contractor bids against design capacity. Your data knows actual capacity. When those disagree by 200 CFM, one of them is a document and one of them is a measurement, and only one of them has been in your building recently.

It outlives the equipment. My monitoring isn't bolted to the pump. When that pump gets replaced, eighteen months of history for the old one becomes the baseline for judging the new one. You will know whether the replacement is actually better or just newer, which is a question almost nobody can answer about their own equipment.

It's affordable evidence. Sophisticated owners are now writing quantitative acceptance criteria into their contracts with general contractors, redefining what "substantially complete" means so that it requires proof. Yale and Altura are doing exactly that. A small building owner cannot afford that program. But six months of continuous data is the same evidence in the only form a small owner can buy.

It's worth money to someone other than the maintenance guy. Kilroy Realty runs roughly four thousand water sensors across eighty buildings, and they've turned that alert history into a record they use in insurance negotiations. Same data. Different reader. I had never once thought about who else in a building owner's life might want to see this history.

The Small Team Thing

One more signal, because it's the one I keep thinking about.

A senior controls and commissioning engineer at UCSF Health is about to present on how an owner-built AI agent and one capable BAS engineer beat a multi-year, seven-figure vendor engagement.

A name-brand health system is going public with the claim that a small competent team, working with its own data, outperformed the enterprise contract.

I am not going to pretend that's a story about me. I'm one person with two sites. But it is a story about the direction the buy side is moving, and the direction is toward owning your own record instead of renting access to somebody else's. That is a market where being small and being permanent are the same advantage.

What I'm Changing

Starting with the two sites I already run, I'm producing a written report at ninety days for every site I monitor. Not a dashboard. A document.

What the equipment actually did, month by month. Measured range against whatever the nameplate or the spec claims. What normal looks like for this specific unit in this specific building. What has changed since install, and whether the change was good.

And it belongs to the customer, whether or not they keep paying me. That last part costs me nothing and answers the objection I get most, which is a reasonable fear of being locked into a small vendor.

Because here is the thing that took me too long to understand. The buildings I work in have no institutional memory. The tuning knowledge lives in one person's head, and that person retires or takes another job, and a well-intentioned upgrade copies the settings back to a design intent that was wrong in 2011.

The drawing says 600 CFM. The data says 400.

Somebody has to be writing it down.