A long-exposure photograph of a highway at night shows something that is not there.
The streaks are not cars. They are the record of cars moving faster than the shutter could resolve. Everything about the flow is visible — the direction, the density, where the lanes merge. Nothing about any individual vehicle is. No driver, no destination, no reason for being in that lane rather than another.
Move to the outside lane and travel alongside it, and the streak resolves. It becomes a car with someone in it going somewhere specific. The road did not change. The traffic did not change. What changed was the relationship between your speed and theirs.
That is the clearest statement I can make of what Strand Commonality™ describes, and why it took me forty years to be able to work with more than two or three strands at a time.
What a strand is
A strand is a line of causation running through an operation. Most of them are invisible from inside the function that owns them, because each one lives in a different frame.
In 1998 I was brought into a defence maintenance operation running at 51 percent next-day delivery against a 90 percent target. Every measure on the board was accurate. The determining question was what time of day the technicians submitted their orders — because a next-day requirement is a duration, and a duration has a start point nobody had established.
Four o’clock in the afternoon. Procurement cannot act on an order it has not received.
That was a correct answer. It was not the end of the trace. The next question was why four o’clock, and the answer was in the service department, where staff were rated on calls answered and held parts orders to the end of the day. Then back into procurement, where roughly 80 percent of orders had been pushed to US small and medium suppliers with no cross-border customs experience. Then out again at the border, where every supplier used its own courier and no consolidated clearance existed.
And then out of the operation entirely, into finance. Purchase orders captured part cost only. Courier charges arrived separately on the invoices. The landed cost of a part existed in no single place, and the contractual markup was calculated on part cost alone — so the shipping the purchase order never captured was eating the margin. On transactions the program was booking as revenue, it was losing money.
The reason nobody had caught it is the finding. Finance had done nothing wrong. Part cost and shipping sat in two line items, each managed, each reconciling cleanly on its own. The loss existed only in the relationship between the two lines, and nobody was looking at both at once.
Competent accounting had built the blindness.
Why the limit was two or three
I first noticed this in the early 1980s, programming in dBase II, where adding a better feature in one line of code would break the performance of another. After I made the fix, I turned around and asked what else those two lines touched, and what the repair itself would reach into.
I stopped at two or three. Not from lack of interest. One head could hold that many at once.
The reason the limit is that low is structural rather than a matter of intelligence. Each additional strand requires leaving the frame that made the previous one visible. Nobody working inside a delivery-performance frame asks about time of day, because time of day is not a delivery-performance variable. It becomes one only after you have left. And nobody inside procurement asks about the relationship between two finance line items, because that relationship is not a procurement variable either.
Every strand you reach costs you the vantage point that found the last one.
What changed
For most of forty years the instrument was one person, and the instrument set the ceiling.
The structured multimodel environment I now work in — several independent models running the same question in parallel, with a separate sealed panel whose only task is to attack the collective result — did not make me better at this. It made me faster.
Several frames run at once instead of in sequence. A strand I could not match speed with is now travelling alongside something that can. It resolves, and I can read attributes on it that were previously part of the blur.
Eight to ten strands, roughly as fast as I used to reach two or three.
⚠ The sealed panel is doing something different and it matters. Moving fast enough to resolve a pattern is also moving fast enough to resolve one that is not there. Its job is to attack whatever the main panel produced. On a recent assessment it moved a score of mine downward by nearly two points on two specific defects, and I accepted the correction. Velocity without an adversarial check produces confident artifacts.
The strands do not stop
Over four days this month I ran three of these — one on an anomaly in this site’s own traffic data, one on a set of AI governance graphics, one on a newsletter about where procurement savings disappear before reaching the P&L.
Two of them connected, and nobody built the bridge.
The savings piece identified baseline integrity as the item that decides all the others: a saving is a difference from a prior price, and if the prior price was never established, the difference is not a measurement. That is the same condition as the 1998 finance crossing — a measurement whose reference point was never constituted, with every individual number accurate. And the same condition as a very large software company in 1990, which recognized licence revenue at contract signing rather than at delivery, because nobody had settled which point a sale is measured from, and restated three years of results.
Three domains. Defence logistics, savings reporting, revenue recognition. One condition running through all of them.
That connection surfaced because a lab ran, not because anyone was looking for it. The other one had been parked the night before.
Parked, not finished
I used to describe these as closing. That was wrong, and a better description came out of the same conversation.
You do not stop. You take an off-ramp, because at that moment continuing would not change what you do next. The exit is a judgment about where you are, not a claim that the road ended — and the on-ramp is still there.
The traffic analytics lab was parked at a confidence rating of three out of ten, on the grounds that the instrument could not support a decision. Two days later a mechanism arrived from a board meeting recording in an entirely different domain. Nothing was reopened. Something new pulled up at the intersection.
What this has to do with AI implementations
Every framework in circulation for deploying AI agents is drawn as a static structure. Blocks, stacks, control layers, authority matrices. All of them assume the problem, the boundary and the intervention have already been named correctly.
They are photographs of a road, taken from a bridge, at one shutter speed.
The conditions that determine whether an implementation works are moving at their own speeds, in frames the deployment team does not occupy. Some of them are in finance. Some are in a function that has no stake in the project. Some, as in 1998, are in what time of day something happens — a variable that appears on no scorecard anywhere.
You cannot govern a strand you cannot see. And you cannot see a strand you cannot match speed with.
⚠ This does not guarantee anything, and I want to be exact about that. Reaching more strands means finding more conditions. It does not mean the conditions are favourable, and it does not tell you which of them is determining rather than merely contributing. That selection is still a judgment, and it is still mine to make.
The instrument supplies the velocity. Somebody still has to decide which exit to take.
Truth Is Believing. Accuracy Is Knowing. Outcome Is Proof.™
-30-
Related
Is Strand Commonality™ the Missing Key in Harnessing AI’s Power and Promise?
Posted on September 10, 2026
0
A long-exposure photograph of a highway at night shows something that is not there.
The streaks are not cars. They are the record of cars moving faster than the shutter could resolve. Everything about the flow is visible — the direction, the density, where the lanes merge. Nothing about any individual vehicle is. No driver, no destination, no reason for being in that lane rather than another.
Move to the outside lane and travel alongside it, and the streak resolves. It becomes a car with someone in it going somewhere specific. The road did not change. The traffic did not change. What changed was the relationship between your speed and theirs.
That is the clearest statement I can make of what Strand Commonality™ describes, and why it took me forty years to be able to work with more than two or three strands at a time.
What a strand is
A strand is a line of causation running through an operation. Most of them are invisible from inside the function that owns them, because each one lives in a different frame.
In 1998 I was brought into a defence maintenance operation running at 51 percent next-day delivery against a 90 percent target. Every measure on the board was accurate. The determining question was what time of day the technicians submitted their orders — because a next-day requirement is a duration, and a duration has a start point nobody had established.
Four o’clock in the afternoon. Procurement cannot act on an order it has not received.
That was a correct answer. It was not the end of the trace. The next question was why four o’clock, and the answer was in the service department, where staff were rated on calls answered and held parts orders to the end of the day. Then back into procurement, where roughly 80 percent of orders had been pushed to US small and medium suppliers with no cross-border customs experience. Then out again at the border, where every supplier used its own courier and no consolidated clearance existed.
And then out of the operation entirely, into finance. Purchase orders captured part cost only. Courier charges arrived separately on the invoices. The landed cost of a part existed in no single place, and the contractual markup was calculated on part cost alone — so the shipping the purchase order never captured was eating the margin. On transactions the program was booking as revenue, it was losing money.
The reason nobody had caught it is the finding. Finance had done nothing wrong. Part cost and shipping sat in two line items, each managed, each reconciling cleanly on its own. The loss existed only in the relationship between the two lines, and nobody was looking at both at once.
Competent accounting had built the blindness.
Why the limit was two or three
I first noticed this in the early 1980s, programming in dBase II, where adding a better feature in one line of code would break the performance of another. After I made the fix, I turned around and asked what else those two lines touched, and what the repair itself would reach into.
I stopped at two or three. Not from lack of interest. One head could hold that many at once.
The reason the limit is that low is structural rather than a matter of intelligence. Each additional strand requires leaving the frame that made the previous one visible. Nobody working inside a delivery-performance frame asks about time of day, because time of day is not a delivery-performance variable. It becomes one only after you have left. And nobody inside procurement asks about the relationship between two finance line items, because that relationship is not a procurement variable either.
Every strand you reach costs you the vantage point that found the last one.
What changed
For most of forty years the instrument was one person, and the instrument set the ceiling.
The structured multimodel environment I now work in — several independent models running the same question in parallel, with a separate sealed panel whose only task is to attack the collective result — did not make me better at this. It made me faster.
Several frames run at once instead of in sequence. A strand I could not match speed with is now travelling alongside something that can. It resolves, and I can read attributes on it that were previously part of the blur.
Eight to ten strands, roughly as fast as I used to reach two or three.
⚠ The sealed panel is doing something different and it matters. Moving fast enough to resolve a pattern is also moving fast enough to resolve one that is not there. Its job is to attack whatever the main panel produced. On a recent assessment it moved a score of mine downward by nearly two points on two specific defects, and I accepted the correction. Velocity without an adversarial check produces confident artifacts.
The strands do not stop
Over four days this month I ran three of these — one on an anomaly in this site’s own traffic data, one on a set of AI governance graphics, one on a newsletter about where procurement savings disappear before reaching the P&L.
Two of them connected, and nobody built the bridge.
The savings piece identified baseline integrity as the item that decides all the others: a saving is a difference from a prior price, and if the prior price was never established, the difference is not a measurement. That is the same condition as the 1998 finance crossing — a measurement whose reference point was never constituted, with every individual number accurate. And the same condition as a very large software company in 1990, which recognized licence revenue at contract signing rather than at delivery, because nobody had settled which point a sale is measured from, and restated three years of results.
Three domains. Defence logistics, savings reporting, revenue recognition. One condition running through all of them.
That connection surfaced because a lab ran, not because anyone was looking for it. The other one had been parked the night before.
Parked, not finished
I used to describe these as closing. That was wrong, and a better description came out of the same conversation.
You do not stop. You take an off-ramp, because at that moment continuing would not change what you do next. The exit is a judgment about where you are, not a claim that the road ended — and the on-ramp is still there.
The traffic analytics lab was parked at a confidence rating of three out of ten, on the grounds that the instrument could not support a decision. Two days later a mechanism arrived from a board meeting recording in an entirely different domain. Nothing was reopened. Something new pulled up at the intersection.
What this has to do with AI implementations
Every framework in circulation for deploying AI agents is drawn as a static structure. Blocks, stacks, control layers, authority matrices. All of them assume the problem, the boundary and the intervention have already been named correctly.
They are photographs of a road, taken from a bridge, at one shutter speed.
The conditions that determine whether an implementation works are moving at their own speeds, in frames the deployment team does not occupy. Some of them are in finance. Some are in a function that has no stake in the project. Some, as in 1998, are in what time of day something happens — a variable that appears on no scorecard anywhere.
You cannot govern a strand you cannot see. And you cannot see a strand you cannot match speed with.
⚠ This does not guarantee anything, and I want to be exact about that. Reaching more strands means finding more conditions. It does not mean the conditions are favourable, and it does not tell you which of them is determining rather than merely contributing. That selection is still a judgment, and it is still mine to make.
The instrument supplies the velocity. Somebody still has to decide which exit to take.
Truth Is Believing. Accuracy Is Knowing. Outcome Is Proof.™
-30-
Share this:
Like this:
Related