The Scaling Asymmetry: Generating Outpaces Reconstructing

1. Two Curves That Do Not Match

There are two things an AI system does, and they scale on completely different curves. The first is generating outputs — answers, decisions, actions. This is cheap and near-instant, and it scales with compute almost without friction. The second is reconstructing the basis for any one of those outputs later — showing why this answer, on this evidence, under which rules. That does not scale the same way at all.

The clearest statement of the gap came from Tim Zlomke, a managing partner working on AI systems, in a public comment in May 2026:

One sentence, but it names a structural problem the rest of the governance conversation keeps circling. We have called it the scaling asymmetry, and it has consequences that compound.

2. The Debt Compounds with Usage, Not Time

The usual mental model for technical debt is that it grows with time — the longer you leave it, the worse it gets. Reconstruction debt is different. It grows with usage. Every output produced without its basis captured is one more decision that cannot be reconstructed later. The backlog tracks the number of decisions made, not the number of months elapsed.

This inverts the intuition about which systems are most exposed. It is not the old, neglected system that carries the most risk. It is the busy one — the high-volume, business-critical model making thousands of decisions a day. That system accumulates reconstruction debt thousands of times faster than a quiet one, even if both were switched on the same morning. And it is precisely the busy, consequential system that a regulator, a customer, or a court is most likely to ask about.

3. Why the Architecture Choice Is Hard to Reverse

Here is the part that turns an interesting observation into an architecture decision. If a system was built to capture the reasoning path as it runs, the record exists for every decision from the first one. If it was not, the records for every past decision are simply gone — and they cannot be regenerated faithfully after the fact. You cannot reconstruct the framework version that governed a decision, the evidence state at the time, or the path the reasoning actually walked, if none of those were captured when it happened.

So the two directions are not symmetric. Choosing the capturing architecture up front is cheap: it is a design decision, paid once. Switching to it later leaves a permanent gap covering everything decided before the switch — and, because of the usage curve above, that gap is largest exactly where it matters most. The decision compounds in one direction and cannot be back-filled in the other. That is the precise sense in which it is far more expensive to reverse than to make.

4. The Same Asymmetry the EU AI Act Will Test

The scaling asymmetry is usually felt first as a cost problem — the bill for reconstructing an answer after the fact, in engineering time and re-derivation, dwarfs the cost of having produced the answer. But it expires as a legal problem on the same curve. An EU AI Act inspection in 2027 can ask why a specific past decision was made. If reconstruction debt has been accumulating, the honest answer for older or higher-volume decisions may be that the basis no longer exists.

That is the same gap the rest of this series is about, viewed through the economics. A system that can replay the path a decision walked, under the rules in force at the time, has paid the architecture cost once and converted the asymmetry into a solved property. A system that cannot is carrying a liability that grows every time it is used.

5. The Fix Is Architectural, Not a Logging Bolt-On

The tempting remedy — “we will add logging” — only partly works, and the part it misses is the expensive part. Adding logging captures decisions from that point forward, but it cannot recover the basis for decisions already made, and bolted-on logging tends to record outputs rather than the reasoning path and framework version that produced them.

The durable fix is to make the path the reasoning walked be the record itself, so the basis is captured structurally for every decision rather than reconstructed afterwards — the argument made in detail in proof precedes permission. An honest limit applies here too: the substrate records the basis; it does not make the decision correct, and it does not decide whether the resulting action was permitted. What it does is ensure the asymmetry never opens in the first place.

Continue Reading

• Proof Precedes Permission: The Substrate Seam
• Recall vs Verifiability in EU AI Act Audit Trails
• The Hidden Cost of AI Adoption Friction
• The EU AI Act Audit-Trail Stack (pillar)

Frequently Asked Questions