Someone asked the right question. Waving it off was rational.

Part 1 ended in a review room. Someone asked "wait — what do we mean by active?" and the room had answers ready: that's over-complicating it, we'll handle it in code. The signal existed, was received, and was thrown away. This part answers why — and why the answer is not carelessness.

Start by dissolving the tempting version of the question. It's tempting to ask why the error happens — why organizations full of capable people end up holding wrong or undecided concepts. That question needs no mechanism. A complex domain outruns anybody's first model of it; being wrong is the default state of contact with the world, and no hiring bar or process changes that. The question that pays is different: why isn't the error caught? Wrong models should die young — colliding with a colleague's different model, or with reality. In these failures they don't. They live for a year under green dashboards, accruing changes built on top of them. Something protects them. Two things, it turns out: one keeps the wrongness invisible, and one discards it when it becomes visible anyway. Neither is a character flaw. Both are structural, and both are rational for every person inside them — which is why exhortation, be more rigorous, ask more questions, has never fixed either.

The first mechanism: the concept fragments

Melvin Conway observed in 1968 that an organization designing a system is constrained to produce a design that copies its own communication structure. The law is usually cited about architecture — your services end up mirroring your org chart. Point it at concepts instead, and it explains the first protection.

A concept like active customer spans marketing, product, finance, and data. Each team talks constantly inside itself and thinly across its boundaries, so each builds and maintains its own shard of the concept — and each shard is locally sound. Marketing's active is engaged-recently: defensible. Finance's is billable: defensible. Product's is logged-in: defensible. The contradiction between them doesn't live in any team. It lives in the seams — precisely where, by Conway, the communication is thinnest and no role stands. Nobody is negligent. Everybody's piece is fine. The wrongness exists only in the whole, and the whole is nobody's view.

This is why the ownership symptoms look the way they do: the question that gets forwarded around the org and lands nowhere, the "not really our team's thing" said in turn by every team it touches. Each team is telling the truth. And it's why the talent Part 1 called defining is structurally rare, not just statistically rare. Naming the case that breaks requires holding the whole concept at once, and the org chart guarantees that no position comes with that view. Whoever assembles it does so against the grain of the structure.

The second mechanism: the held signal loses

Sometimes someone assembles it anyway. They cross the seams, hold enough of the whole, and raise it: these definitions don't agree, and here is where that will break. Now the second protection takes over. The room has to price the question, and the pricing is rigged — not by anyone, by what can be measured.

Take everything a room can score about a piece of work today: speed — did it ship; cost — what it took; the qualities you can test on the build in front of you — does it stay up, does it compute what the spec says, can it be broken into. Now subtract those. The remainder is the thing the raised question is about: whether the concept underneath is sound, whether the work will still make sense when reality presents its next case. That remainder can't be scored by the room today, because it's about builds and cases that don't exist yet. And a value you can't score until next year does not win an argument against two you can score this afternoon. It doesn't lose because anyone is short-sighted. It loses on the arithmetic of the meeting.

Goodhart's law is the famous relative — put pressure on a measure and it stops measuring — but the room's failure is plainer and older: substitution, the countable proxy standing in for the uncountable value. Every function in the building runs a local version. Engineering's version is the purest. An engineer is rewarded for building exactly what was asked, reliably and quickly. Pushing back on the ask almost never buys more time or budget; it only adds ways to miss the agreed scope. So the rational move is to become a clean interface with no slot for why: tell me precisely what to build, and I'll make it solid. Each engineer becomes their own microservice — a strict contract, the reasoning sealed inside. The people best positioned to ask what a term means are the ones whose incentives most reliably train them not to.

And there is a third layer, the one that stings, because it implicates the person trying hardest to get this right. Even a leader who wants the truth is rational to discount the raiser. At the moment it's raised, this concept is unsound is an unverifiable claim: cheap to assert, expensive to check, and indistinguishable on its face from over-engineering, from taste, from stalling. You have been burned by all three. Lacking any instrument to tell the rare true claim from the common false ones, discounting is not a bias — it's the correct move. That's over-complicating it is authority's honest defense against a talent it has no way to test. Which is the diagnosis in one line: the room isn't short on rigor or goodwill. It's short on an instrument.

Note that the same pricing protects the other origin from Part 1 even better. Questioning a senior's confident model of the domain carries every cost above, plus one. So the misconceived concept — the one with no disagreement to expose it — gets the strongest protection of all.

What the two mechanisms make together

Fragmentation makes the whole-concept view rare, so the signal is scarce. Measurement pressure prices the scarce signal below the afternoon's deadline, so when it appears, it loses. An organization built this way — which is to say, an organization — produces smoke constantly and is structured to ignore it. Nobody decided this. It's what Conway's law and the arithmetic of the meeting do when nothing counters them, and it's why the misattribution from Part 1 is so stable: a failure nobody could see whole, waved off by a room that was pricing correctly, gets filed a year later under whatever label is nearest.

So the fix is not more seers. The defining talent is rare and you can't manufacture it — no pipeline mints it on demand, and a training program that claims to is selling you the sensing half. The fix is to stop wasting the seeing you already get. That has two halves: hold the signal, which takes no talent, and then get the rare act performed and verified, which does. The first half you can start this week.

Holding the signal: convergence, not diagnosis

We built an instrument for the first half, and its whole logic fits in a paragraph. Conceptual debt can't be proven from the hallway; proving it is the rare act. But its presence throws off a wide, recognizable family of cheap symptoms: words that won't hold still, problems that return after being fixed, questions no team will own, official systems everyone quietly routes around, confidence that doesn't fit the difficulty, numbers reality keeps contradicting. Any single one has an innocent explanation — a shadow spreadsheet can just mean the tool is clunky. The instrument is the convergence: count the independent kinds of symptom pointing at the same concept — kinds, not instances, because three copies of the same symptom are one symptom in three coats. One kind is noise, and two is worth a conversation; three or more converging on one concept is where we draw the line. The line is a judgment call, not a statistic, but the logic behind it is hard to argue around: one innocent story rarely stretches across three different kinds of trouble on the same idea. That much smoke means a conceptual risk that no current role will catch. Its output, deliberately, is not you have conceptual debt. It's this concept needs an owner.

Notice what that does to the second mechanism. The waved-off question was unverifiable; a convergence count is not. A non-expert can check it — walk the symptoms, count the kinds — which means, for the first time, the room can price the signal on something other than the raiser's say-so. The written-up version is called the Conceptual-Debt Screen, and it is a heuristic, not a validated test; the symptom families are representative, not complete; a clean screen is reassurance, not proof. But it converts smoke from a feeling into a count, and a count can survive a meeting.

What it cannot do is the rare act. The screen ends with a named person standing over an orphaned concept — an owner, not a definition. Someone still has to name the case that breaks, and that someone is a talent no title reliably marks. So Part 3: how to recognize the definer by observable behavior rather than role, and how to reward a naming you had no way to verify in advance — the artifact that turns a seer's claim into a credential a non-expert can read. That's next.