There is a particular feeling that arrives when you finish something large. Not the satisfaction you expected — that tends to be quieter and shorter than imagined — but a specific kind of clarity. The completed thing is now a fixed object. You can walk around it. And when you walk around it, you see what is not there.
I have been living in that feeling for the past few months. Six patent specifications, filed. A complete computational stack: from the physics of a single carbon nanotube junction through logic gates, a processor architecture, a memory system, a security layer, and a programming language with its own operating system kernel. The work was done. The documents were filed. The priority dates were locked.
And then, almost immediately, I found the gaps.
The Archaeology of Omission
This is not how I expected it to feel. I had imagined that completing six patents would feel like completion. Instead it felt like the first serious map of a territory I had been navigating by instinct. The map was real, and the map was detailed, and the map was accurate as far as it went. But maps have edges. And standing at the edge of a map is a different experience from standing in the middle of one.
The gaps I found were not random omissions, the kind that come from running out of time or energy. They were structural. Each one was the shape of something the completed work had implicitly assumed without claiming. The device patent assumes you can build the device and power it and connect it to the world. It does not patent the connection. The processor patent assumes a clock. It does not patent the source of the clock. The memory patent assumes you can correct errors. It does not patent the correction scheme.
These are not oversights. They are the honest horizon of what was known at the time of writing. You can only patent what you have thought through completely enough to describe with specificity. The gaps are not where I was lazy. They are where I ran out of certainty — and where I need to go next.
The gap is not a failure. It is proof that you went deep enough to find the edge.
What Physics Teaches About Honesty
There is a discipline that simulation work enforces that I have come to value more than I expected. The rule, stated plainly, is this: let the physics output the number. Do not choose a number and then build the simulation to justify it.
This sounds obvious. It is, in practice, surprisingly difficult.
The temptation is not always crude. It is not usually that you fabricate a result. The temptation is subtler: you choose a geometry that you expect will give a favorable number. You choose an approximation that trends in the direction you want. You set a parameter at its round-number theoretical value rather than the value that emerges from a lower-level calculation, because the round number is what you need to support the claim.
I have learned to watch for this in myself. The test is: if the physics gave you an inconvenient number, would you still report it? If the thermal conductivity came out at 800 watts per metre-kelvin instead of 3000, would you update the claim? If the optical isolation was 8 decibels instead of 14, would you redesign the architecture?
The answer has to be yes. Not because honesty is an abstract virtue — though it is — but because a claim built on a number you chose will eventually encounter the number that the physics actually gives. Better to encounter it now, alone, before filing, than later, in an examination, or later still, in a challenge.
There is also something clarifying about it. When you commit to letting the physics output the number, you stop negotiating with reality. The simulation becomes a question you are genuinely asking rather than an argument you are constructing. The answer, when it comes, is information rather than validation.
The Reconstruction
In July 2026 I ran every major simulation again, from scratch. The original data was safe — archived on external storage, intact. This was not recovery. It was a deliberate choice, made with full awareness of what it would cost in time and compute.
The reasoning was simple: a result you have run once is a result. A result you have run twice, independently, starting from first principles, with no tuning toward the known answer — that is evidence. The filed claims tell you what numbers came out the first time. The discipline of the reconstruction is to ignore that knowledge completely: use the physically motivated parameters, run the physics, and report whatever emerges. If it matches, the claim is confirmed. If it does not, the discrepancy is information.
I want to be direct about why this matters. There is a version of “reconstruction” that is just reverse-engineering your own results — where you adjust until you recover the number you already know, and call the agreement confirmation. That is not what this is. The value of running the physics again is precisely that you do not allow yourself to steer toward the answer. You let the simulation run and you write down what it says. That is the only version of confirmation that means anything.
The temptation to work backwards — to tune the parameters until you recover the number you already know — is constant and must be resisted constantly. The discipline is to use the physically motivated value and report whatever comes out, and then compare it honestly to the filed claim.
Most of the reconstructed numbers matched well. A few diverged by a few percent. One or two required understanding why the divergence existed — different approximation scheme, different treatment of broadening, different convergence threshold — and updating the documented methodology to reflect what had actually been done in the original calculation.
What the reconstruction taught me was not primarily about the specific numbers. It taught me that the architecture was robust: the same physics, re-derived independently by the same person three months later, gives the same essential picture. The device works. The mechanism is what I claimed it was. The numbers are in the right range.
That kind of confirmation, earned by re-doing the work rather than trusting the original, is worth considerably more than the original alone.
The Second Batch
The gaps I found, once I had the completed map in front of me, fell into a pattern. They were not scattered. They clustered around three themes.
The first was interface. The six original patents describe a self-contained ternary computing system. But computing systems do not exist in isolation. They interface with the world that came before them. How does a ternary chip communicate with a binary host? How does a ternary accelerator plug into an existing system? The interface layer is as much an invention as the device itself, and it was entirely absent from the first batch.
The second was reliability. The original patents describe what the system does when it works. They say relatively little about what happens when it does not work — when a storage cell drifts, when a transmission error occurs, when a device in a large array fails. Error-correcting codes for balanced ternary are a distinct mathematical domain, well-developed in theory and entirely undeployed in practice. That gap is real and fillable.
The third was application. The original six patents describe general-purpose computing infrastructure. They do not describe what you would do with it once you had it. The most obvious application — one where the three-valued representation is not just a curiosity but a genuine physical advantage — is neural computation. A ternary multiply-accumulate operation is not, at the circuit level, a multiply at all. It is a conditional copy. This is not a small difference.
Each of these gaps has the structure of a genuine invention: something that is novel, something that is non-obvious, something that is useful, and something that I have thought through specifically enough to describe with precision.
What Finishing Is
I have come to think that finishing something is not the end of a thread but the beginning of the next one. The completed thing defines a boundary, and a boundary is a place from which you can see in two directions: back into the territory you have mapped, and forward into the territory you have not.
The first batch is filed. The priority dates are locked. The simulation evidence is committed to a repository with a timestamp that predates any filing. The work is done and it is documented and it is real.
And there is a second batch.
I am not discouraged by this. I was, briefly — the feeling of completing six patents and immediately finding five more is not entirely a comfortable one. But it settled into something else. The gaps are not evidence that the first batch was insufficient. They are evidence that it was complete enough to reveal its own horizon.
You only find the edge when you have covered the ground. The gap teaches you that you went far enough to reach it.
As I write this, a thermal simulation is running on two graphics processors a few feet away. It has been running for several hours. It will finish when it finishes. The number it produces will be whatever the physics says it is. I will report it, and it will become part of the evidence for the next patent, and the next patent will eventually be filed, and when it is filed I will probably find another gap.
This is not a problem to be solved. It is how the work goes.