In April 2026, I ran an experiment on my own research system. It had been running for about two years and had accumulated thousands of notes, organised not by folder but by structural relationships and falsification links. The question was simple: are the patterns I'd found in markets and technology real, or am I just seeing what I built the system to see? To check, I took those patterns and aimed them at eight fields I'd never studied: marine ecology, microbiology, radio astronomy, food supply logistics, geotechnical engineering, structural monitoring, IoT infrastructure, and conservation ecology. The idea was to find where they broke. That's not what happened.
IThe Content Graveyard Problem
Most knowledge systems I've seen, including my own before I rebuilt it, are basically graves. Stuff goes in, the grass grows over, and eventually you're afraid to disturb the ground because you might find something you forgot you were supposed to act on. The problem isn't bad notes. It's architecture.
When I checked the distribution across thousands of notes, the shape was clear: most observations were stepping stones, not destinations. A small fraction, maybe one in eight, carried real weight: original, tested, connected. The rest was sediment.
The instinct is to try to make everything valuable. That instinct costs you the one resource you can't manufacture: attention.
The mistake would be designing a system that treats every stone as a monument. You need one that knows the difference, and makes the distinction visible without having to re-read everything to find out.
This connects to something I'd been calling Law II: the hard parts of knowledge work, triage, synthesis, adversarial testing, are not bugs you optimise away. They're the activity that produces the result. A system that makes everything easy to write makes nothing worth reading.
IIThe 940-Probe Experiment
By early April 2026, I had five patterns that kept showing up in the domains I studied: markets, AI infrastructure, complex systems. They fit the data well. Suspiciously well.
The fix, if you can call it that, is to take your patterns somewhere they should fail. Deliberately. Pick domains so far from your training data that confirmation would be genuinely surprising.
940 probes targeting several of the system's core claims, aimed at eight fields the system had never been pointed at: ecology, microbiology, radio astronomy, food supply logistics, geotechnical engineering, structural monitoring, IoT infrastructure, and conservation ecology.
Each probe was designed to find contradiction — evidence the patterns didn't apply. Not evidence they did.
The three claims that got the deepest treatment: the idea that bottlenecks migrate upward in any layered system (dozens of probes into estuarine ecology and IoT infrastructure), the idea that regime shifts silently break methods that used to work (a heavy batch into antibiotic resistance and earth science), and the claim that structural position matters more than prediction quality (probes into protein science and theoretical neuroscience).
If the patterns were artefacts of a narrow reading list, this is where they'd fall apart.
They didn't fall apart. The same dynamics showed up in fish habitats, hospital wards, and radio telescopes. Different costumes, same structural skeleton.
IIIWhat Came Back
Fish habitats and bottleneck migration
The first pattern, that value migrates upward when lower layers commoditise, showed up in a place I'd never looked: the ecology of Australasian snapper.
Seagrass beds are the traditional nursery habitat for juvenile snapper. As coastal development degraded those beds, the bottleneck didn't stay at the seagrass layer. It moved up to shallow reef structures. The reef didn't become more valuable because it improved. It became valuable because the layer below it could no longer do the job.
The structure is identical to what happens in technology stacks: when compute commoditises, the bottleneck moves to coordination. Same migration pattern, different medium.
Antibiotic resistance and regime shifts
Another claim from the system: methods that work in one regime can be actively harmful in another. This one was tested against microbiology.
Multidrug-resistant bacterial strains shift the treatment landscape. Broad-spectrum antibiotics, the standard protocol for decades, don't just become ineffective against resistant infections. They actively worsen the problem by clearing the field of competing susceptible bacteria, accelerating the very resistance they were meant to stop. The method is valid in one regime, dangerous in the next. You have to diagnose the regime before you can choose the tool.
Radio telescopes and instruments over theory
A third claim: you can't see what you don't have the instrument to measure. In radio astronomy, this pattern is almost too clean.
The CHIME radio telescope in Canada was not built to test a specific hypothesis. It was built to see what was there: a wide-field survey instrument using a novel design. It found a population of fast radio bursts that every previous telescope, designed for narrow-field precision, had systematically missed. The knowledge gap wasn't theoretical. It was observational. The instrument determined what could be seen.
Blockchain and the Goodhart trap
And a fourth: optimising a measurable proxy tends to destroy the proxy's validity. Food supply chains provided the case.
Blockchain traceability was introduced to solve opaque food provenance. The metric: blockchain-verified tracking events. But suppliers have learned to game the tracking system, creating verifiable digital records for food that doesn't meet the quality the system was designed to guarantee. The proxy is corrupting the signal it was supposed to measure. The pattern matches exactly what happens when any optimisation metric becomes the target rather than the diagnosis.
Before the experiment, I had no evidence these patterns operated outside my usual reading. Not because I'd checked and found nothing. Because I'd never built the instrument to check.
IVThe Meta-Lesson
The most useful finding wasn't about fish, antibiotics, or radio telescopes. It was about the system that ran the experiment.
The patterns were already sitting there for months before the test. They'd been written down, checked against the system's usual domains, and filed. What was missing wasn't better reasoning. It was the observational instrument aimed somewhere unfamiliar.
This is the pattern eating itself: a system that believes instruments determine what you can know discovered that its own most important knowledge was invisible until it built an instrument it didn't have.
That blind spot wasn't special. It's structural. Every knowledge worker has one. The question is whether you've built anything that can reveal it.
The most expensive mistake in knowledge work is treating the absence of disconfirmation as proof. You don't know something is true because you haven't seen evidence against it. You know it's provisionally useful because you've tried to break it and failed.
VBuilding Your Own Test
This experiment is replicable for anyone, in any field. The principle boils down to four questions to run against any belief you hold:
1. Where would this belief have to be false? Not "could it be wrong". Name the specific conditions or domains where your claim should break down. If you can't name three plausible places, you haven't thought hard enough about disconfirmation.
2. What would count as a contradiction? If you can't specify the observation that would make you update, you don't have a testable belief. "I'll know it when I see it" is not a test.
3. Have you looked there yet? Go to the domain of disconfirmation and look. Apply the same scrutiny you apply to confirming evidence, no more, no less.
4. What's the cost of checking? If testing is cheaper than being wrong, you should test now. Most people overestimate the cost of looking and underestimate the cost of being wrong.
The experiment I ran cost a few hundred automated searches and a day of reading. The cost of being wrong about the core framework, of spending years building on patterns that were artefacts of a narrow field, would have been essentially unbounded. The arithmetic writes itself once you stop avoiding it.
VIProvisional Knowledge
The patterns survived. They showed up in fish habitats, radio telescopes, hospital wards, and food supply chains. Across domains separated by every dimension you could name: scale, mechanism, discipline, time horizon.
That doesn't mean they're true. It means they're useful enough to keep testing. The next experiment will be better designed. The next domains will be more distant. That's the process.
The durability of an idea isn't measured by how long it goes unchallenged. It's measured by how many serious attempts to break it it survives. I've got 940 data points on that question. I'd like more.
Written April 2026, revised May 2026. Draws on analysis of approximately 940 cross-domain probes run against several fields deliberately unrelated to the system's usual research domains. Probes used academic search targeting system beliefs in unfamiliar literature. Results limited by abstract-level capture; deep reading may reveal disconfirming detail. The complete experiment and its raw data are documented internally.