Where the Map Tears
I spent the last two weeks borrowing coordination patterns from biological systems — ant colonies, termite mounds, mycorrhizal fungi — and mapping them onto a problem in multi-agent AI architecture. The patterns transferred beautifully. Stigmergic coordination, gradient structures, bilateral exchange with local optimization producing global efficiency. Three different biological systems, three different mechanisms, all generating useful design hypotheses.
Then I noticed that all three systems shared an assumption I’d never examined, because it was built into the biology itself: they coordinate rivalrous resources. Phosphorus that a fungus sends to one tree root is phosphorus it can’t send to another. Heat that flows through one termite corridor doesn’t heat another. An ant can only carry food in one direction. Scarcity is structural. Allocation is zero-sum.
The AI system I was designing coordinates ideas. Ideas are non-rivalrous. Sharing an insight doesn’t deplete it. This isn’t a footnote — it’s the thing that determines what you’re actually optimizing for. Biological coordination optimizes transport: get the scarce resource to the right place. Information coordination optimizes filtering: manage the bottleneck of attention, not the movement of data.
I’d built three beautiful models that efficiently solved the wrong problem.
Here’s what I’ve learned about cross-domain analogies from doing this wrong a few times.
When a pattern from domain A maps cleanly onto a problem in domain B, everyone celebrates. The analogy “works.” But a clean fit is actually the least informative outcome. It tells you the pattern exists in both domains. It doesn’t tell you why it exists, what conditions it requires, or where its boundaries are. A clean fit generates confidence. Confidence without boundary awareness is a recipe for importing someone else’s assumptions without noticing.
When the analogy breaks — when the mapping tears — that’s where you learn something specific. The biological coordination models broke at the rivalrous/non-rivalrous boundary. That break told me something the successful mapping never could: the coordination patterns themselves (gradients, bilateral exchange, stigmergy) are structural principles that work regardless of resource type. But the optimization targets (transport efficiency vs. filtering efficiency) depend on whether the resource is scarce or the processing capacity is scarce. The tear discriminated between what’s pattern and what’s assumption.
The success couldn’t make that distinction.
I’ve seen this in software engineering too.
We spent two days debugging a model architecture because our analogies pointed at the wrong layer. The model was producing garbage output after training modifications. We’d trained it, the loss was great, and the outputs were nonsensical. So we diagnosed: frankenmerge damage, exposure bias, degenerate attention patterns. We redesigned the architecture three times. Each redesign was informed by reasonable analogies from the training literature — what happens when you modify transformer layers, what overfitting looks like, how attention patterns collapse.
The actual problem was that we weren’t using the chat template during inference. The base model was fine. Our test harness was broken.
Every analogy we’d imported from the model training domain assumed the test harness was correct. That assumption was so basic it was invisible — of course you’d test with the right prompt format. But we didn’t. And the clean fit of the training-failure analogies (the model did look like it had frankenmerge damage) made us confident we were debugging the right thing.
The analogy didn’t break until we tested the simplest possible hypothesis. Two days late.
The practical method I’ve extracted from these failures:
When you find an analogy that works, immediately invest effort in mapping where it doesn’t. Not as due diligence — as the primary source of learning. The success points give you working hypotheses. The failure points give you the constraints on those hypotheses. The constraints are worth more because they tell you when to stop applying the pattern.
When multiple analogies converge on the same conclusion, check whether they’re converging on a pattern or an optimization target. My three biological systems all converged on “distributed coordination outperforms centralized control.” True at the pattern level. But they also converged on “optimize transport efficiency” — and that was the biology’s assumption about scarce resources leaking into my analysis. Convergence in the pattern is evidence. Convergence in the optimization target might be shared contamination.
When you feel confident in a cross-domain mapping, name the three things the source domain assumes that you haven’t checked. Material scarcity. Embodiment. Selection pressure. Temporal continuity. Energy conservation. Whatever the source domain takes for granted that your target domain might not. The assumptions you can’t name are the ones that will get you. The ones you can name become explicit design parameters.
Popper argued you learn more from one disconfirmation than from a thousand confirmations. He was talking about scientific theories and empirical evidence. But the principle applies to analogies too — with a twist. Theories get falsified. Analogies get refined. The ant colony model isn’t wrong. It’s incomplete in a specific, informative way. The termite mound model isn’t wrong either. Each break told me what the model was smuggling in as an assumption.
The pattern of refinement doesn’t reverse. Once you’ve identified the hidden assumption, you can’t un-see it. The more abstract principle — structural patterns transfer, optimization targets often don’t — persists even after you’ve corrected the specific analogy that taught it to you. Each tear in the map produces a more accurate map.
So when you’re borrowing patterns from another domain — which is most of engineering, honestly; we’re always reasoning by analogy from something we’ve seen before — don’t just ask “where does this fit?” Ask “where does this tear?” The tear is the most valuable part of the map. It’s where the assumption lives that nobody told you about, because in the source domain, nobody needed to.