Split at the Seams We Continually Refine

Thoughts from the treadmill ¹

01 Agreeing on the Analysis Is Cheap

Getting close to truth almost always takes multivariate analysis — you hold several variables at once and weigh them against each other, anchored to a few priors you're willing to defend out loud. But that part isn't controversial. No one actually working a non-trivial problem champions one-variable thinking; "don't oversimplify" is a platitude. The agreement is cheap, and it's not where people fail.

Where they fail is the next step. After the analysis, you have to cut. You have to draw a seam — break the big problem into smaller problems you can actually reason about, commit to a module, decide what's in and what's out. Most people won't. They run the multivariate analysis and then leave it open: every perspective accommodated, every edge case nodded at, the claim sprawling across every dimension. They call that rigor. It isn't. It's dilution — a coverage problem dressed up as care — and it's a problem nobody can actually act on. Multivariate analysis is the price of entry, not the achievement. The achievement is progress — and the cut is how you start making it.

02 Engineers Cut — Then Keep Cutting

Engineers are biased the other way — biased to draw the seam even when it's uncomfortable, and, just as important, biased to keep moving it. You bound the module, ship against it, watch it fail somewhere, redraw the boundary. The cut isn't sacred; it's version one. Draw it now, refine it continuously — that instinct is the whole game, and it's exactly what the analysis-only crowd never reaches for. They'd rather hold the problem whole and pristine than cut it wrong, so they never cut it at all.

03 What Holds and What Slips

Here's how I split it once I've cut. There's a major slice and a minor slice (the 90 and the 10, if you want numbers), and you hold the line between them loosely, because we often can't actually know where it falls. The major slice is the empirical part — the stuff the evidence actually backs, the part that survives a regression. The minor slice is everything that won't sit still under a microscope: what hurts people, what crosses a line, what a culture quietly refuses. And the way smart people fail isn't by ignoring that minor slice — it's the opposite. Philosophers and pure scientists keep stretching the model to formally account for it, over-fitting to the tail, and somewhere in that stretch they bury the major slice that was solid. The signal was right there. They buried it under noise by trying to be precise about variables that don't take precision.

Because nobody can cleanly weigh that minor slice. There's no clean coefficient for how much something offends, or how much it matters to the people living inside it. But you can't make a real call without it — drop the coefficient term and you ship something technically correct and culturally dead. So you don't compute it, you sample it — there are sensors who can. The people in the trenches and at the fringes: not the expert with the cleanest credential, but the genuinely multidisciplinary ones who've touched enough different systems to feel where the lines are. They're sensors, not vetoes. And this isn't the inclusion I warned about — like any sensor, they're placed, weighted by how often they've been right, not by how loud they are. They tell you where the minor slice sits; they don't get to re-litigate the major. Depth across a few domains beats noise across many.

04 The Split Was Never Out There

Now the part that takes experience to see. That split between major and minor — ninety-ten, or whatever it really is — is domain-specific, and at the scale of the big human problems we often can't even establish it. Generalized self-driving, fixing inequality, the case for scaling rocket propulsion past its ludicrous-sounding surface — different domains, same move: drawing a seam the world never drew for you. The split only resolves once you've bounded the problem down to something tractable. Which means the split isn't a property of the world you discover — it's something you create the moment you decide where to cut. Splits are made, not found.

That's the builder's move, and neither the scientist nor the philosopher reaches for it by reflex: take a macro problem and decompose it into bounded modules small enough that the trade-offs inside them can start to make sense. Decomposition isn't the warm-up before the work. On the hard problems, decomposition is the work.

05 When the Seams Aren't Real

We know the seams aren't real — a seam is something you draw, not something you find. For complicated systems that barely matters; the cut is closely estimable, the pieces genuinely separable, and the seam basically holds. For complex ones it breaks wide open. Rittel and Webber called these wicked problems: the formulation and the solution are the same act, there's no stopping rule, and every intervention rewrites the problem underneath you. In systems terms, you can only modularize cleanly when the coupling between modules is weak. Much of the trouble with human-scale problems is that the coupling is strong — feedback loops run straight through every boundary you try to draw. Bound a wicked problem badly and you don't eliminate the hard trade-offs, you just push them into the interfaces and lose sight of them. You optimize each module locally, integrate, and the system behaves worse than any component promised. Local optima don't sum to a global optimum. The split looked clean because you drew a seam that was never there.

06 So You Cut Again

This is also why the cut can't be a one-time thing. On a wicked problem ² the seam moves — it shifts as the system exposes more of its structure — so re-drawing module boundaries isn't a sign you got it wrong the first time, it's the job. Seam-finding is a continuous operation, an agility problem, not a planning problem. You re-cut on feedback. And that changes who you actually need in the loop. Three roles, one loop. The scientist optimizes one variable inside a module someone already bounded — deep, narrow, useless until handed the box. The philosopher audits the framing: is the perception clean, is the objective function honest, or are we optimizing a convenient proxy. And the builder makes the cut, commits, and ships the consequence.

That third role is the one that usually gets forgotten, and it's load-bearing — because on a wicked problem, shipping is what generates the information no amount of analysis can. It carries its own failure mode, though: from the inside, decisive and correct are indistinguishable. A confident builder who cuts wrong still hands you a clean, executable, internally coherent decomposition, and because it runs, it reads as progress. That's how execution bias launders "we committed to a cut" into "we found the right cut." And out here there's no rollback — every intervention is irreversible and mutates the terrain you were measuring.

Which is why the builder's job was never to win the room or outvote the other two. It's to close the loop — to ship the one thing a wicked problem respects, which is real-world feedback that updates the formulation. The edge was never decisiveness. It was being the one most willing to be wrong in public, on a clock. And often that's what it takes to move a complex system at all.

And that's the trade I'll take every time. Progress is the achievement — not perfection, not a straight line. Some stretches run two steps back for one step forward, and for a short while that's fine; you're still moving, still generating feedback. What it saves you from is the cesspool — the pristine, uncut problem left to sit in its own fog while everything inside it quietly decays and dies. A seam can always be redrawn. A problem no one will touch just rots.

07 Plant Your Feet, Leave It Open

So when I write to persuade — I plant my feet on the major slice and leave the minor one standing open. Named, visible, handed to you to push on. No slice is softened into mush or dissolved into everyone's exceptions. Open isn't dropped. The honest, pragmatic parts survive precisely because I stopped trying to include my way to the truth, drew the seam instead — and kept refining it as it moved.

¹ Refined by Claude

² On wicked problems: Rittel, H. W. J., & Webber, M. M. (1973). "Dilemmas in a General Theory of Planning." Policy Sciences, 4(2), 155–169. https://doi.org/10.1007/BF01405730