The Three Gates
Every idea, before it gets a dollar, a day, or your belief. Three questions, asked in that order. They came out of a year spent chasing a mathematical pattern that turned out not to be there — and the discipline was worth more than the pattern would have been.
Is it real?
Does it survive a test designed to kill it — not one designed to confirm it? Feed it something that cannot contain the structure you're claiming. If the structure still appears, you may be measuring your method instead of reality.
Is it new?
Or are you following someone else's path in your own notation? Search the prior work before you convince yourself you've found new ground. A new name, notation, or presentation does not make an old idea new.
Does it have a foundation?
Is there a mechanism, or only a resemblance? A beautiful analogy between two domains is not a bridge between them. If you cannot explain what connects them — mathematically, physically, or causally — you may have a metaphor, not a mechanism.
Gate One — Is it real?
The instinct, when a result looks good, is to run more tests that could confirm it. This is precisely backwards, and it is the single most expensive habit in independent research.
The operative move is the negative control: feed your pipeline something that cannot possibly contain the structure you are claiming, and see whether your result shows up anyway. I once had a constant that pinned to −0.333333, stable to seven digits across four orders of magnitude. It was forced by a mean-subtraction step in my own normalization. A list of random numbers reproduced it exactly.
Two supporting habits. Pre-register your predictions — write down what you expect, and how many things you are testing, before you look. And plot it. A summary statistic is blind to whole classes of error that a human eye catches instantly. I once had a metric report "no signal" when the signal was present but offset; the numbers said 0.20, and a person looking at the figure noticed the peaks were landing systematically between the prediction lines rather than on them. The corrected value was 2.91.
Gate Two — Is it new?
This gate is unglamorous and it saves the most time. Search the prior work before you commit — and search it against the claim you intend to make, not against the tool you intend to use. Old methods produce new results constantly; that is most of how any field advances. What advances nothing is rediscovering a result that already has someone's name on it, and that is where a year disappears.
It also cuts the other way, and this half is harder. When a result is significant, go looking for the boring reason. I once had a correlation at p ≈ 10⁻⁹⁰ that turned out to be partly a mathematical identity — the two quantities shared a term. Overwhelming significance and no content. If a result is strong, your first job is to find out whether it is strong for a trivial reason, because a reviewer will.
The stance that makes this bearable: I am not trying to ignore prior mathematics. I am trying to find a simpler structure underneath it.
Gate Three — Does it have a foundation?
Many attractive ideas die here, because resemblance is easy and mechanism is hard. The question is whether there is a mechanism connecting your cause to your effect, or only a resemblance between two things that both happen to be interesting.
I spent a long time on a framework whose stated ambition was to predict how a strand of DNA or RNA folds, starting from the prime factorisation of integers. The images were genuinely productive; the ambition was not. There is no known and no plausible mechanism by which the factorisation of an integer determines the folding of a molecule. The two belong to different worlds, and a resemblance between them is an analogy, not a bridge.
The same gate kills engineering ideas faster than any prototype does. A tailpipe attachment that captures CO₂ fails on mass balance — the sorbent that would be needed weighs more than the car. That is not an engineering shortfall you can iterate out of. It is arithmetic, and it is available in an afternoon, before you build anything.
The gate I applied late, and expensively
Run these three questions on the plan, not only on the invention. This is the lesson that cost me the most, and I did not learn it until long after I should have.
For thirty years I tested every formulation and every result rigorously, and never once tested the business model wrapped around them: file it, then enforce it. That plan does not survive Gate Three.
A patent is not a right. It is a license to sue — and the suing is priced beyond nearly everyone it was written to protect. The grant costs a few hundred dollars. Enforcement costs more than a house. There is no mechanism by which an individual outspends a corporation in federal court; the outcome there is set by capital, not by claims. Nobody explains this at the front of the process, and an entire industry depends on you not knowing it.
I found that out the way you would expect.
Ask of every beautiful pattern not only is it real? but is it new, and does it have a foundation? That discipline is, in the end, the most durable result of the work. Purple Mathematics, concluding section
What a properly killed idea leaves behind
Killing an idea well is not the same as abandoning it. Three things should survive every execution:
A number. Close the branch with a measurement, not a mood. "It didn't feel promising" will let you wander back in two years from now. "Variation: 0.0000 mm, assist 0.3 W against 6 W draw" will not.
The corpse, labeled. Never delete a failure. In patent work, documented dead ends can help show that the successful route was less straightforward than hindsight suggests — they do not prove non-obviousness by themselves, but they can become useful evidence. In research they may save the next person the year you just spent.
The salvage. Almost every killed idea has a smaller, true version inside it. The CO₂ capture device is dead; the honest particulate scrubber is real. The satellite-magnetosphere claim is wrong; the ionospheric aluminium question underneath it is wide open and unmonitored. Killing the flattering version is usually what lets you see the real one.