Bit Ghost Security

Hardened Logic for an Intelligent Era

7 minute read

This is the final part of a three-part series about TolmiGround. Part 1 described the system; Part 2 told the story of where it came from. Here I explain why TolmiGround is only one component of a larger system I am building — one meant to reason in a manner closer to general intelligence. TolmiGround is patent pending — U.S. Provisional Patent Application No. 64/181,356, filed October 21, 2026.

Saying the Quiet Part Out Loud

I have described TolmiGround carefully and modestly for two articles, because the engineering deserves to be judged on its own terms. But I built it in service of a larger ambition, and it would be dishonest to leave that unsaid.

I want to build a system that behaves, in its reasoning, in a manner closer to what people mean by general intelligence — and I think the dominant approach is driving down the wrong road to get there.

Let me be careful about words, because this is exactly the kind of claim that gets oversold. I am not claiming TolmiGround is AGI. I am not claiming the larger system is AGI, or that it will be. What I am claiming is that there is a different road — one built on grounded reasoning rather than probability — and that TolmiGround is the first working stretch of pavement on it.

The Road Most Traveled

The intelligence we are all watching emerge is, at its foundation, prediction at enormous scale. Show a model enough text and it gets astonishingly good at producing the most likely continuation. That capability is real and it is not going away. But a system optimized to sound right is not the same as a system that is right, and the gap between the two is precisely where the failures live — the confident fabrications, the answers that cannot cite their grounds, the inability to distinguish “I have no evidence” from “the evidence is mixed.”

You can patch around these problems. You can bolt retrieval onto the front and guardrails onto the back. But the core of the system is still a probability engine, and you are always negotiating with its instinct to produce something plausible. The grounding is an accessory. The prediction is the engine.

I wanted to try building it the other way around: grounding as the engine, with everything else in service to it.

What the Other Road Looks Like

Think about how a careful person actually reasons their way to something they are willing to defend. They do not generate the most fluent-sounding answer and stop. They start from things they are confident are true. They test a supposition against those. When a step is shaky, they do not paper over it — they break it into smaller steps and test those, and if the smaller steps are shaky too, they go further down. They hold “I don’t know yet” as a legitimate, stable state rather than a gap to be filled. And they keep the results, so the next hard question starts from more solid ground than the last one did.

Read that paragraph again and you are reading the design brief for TolmiGround. Ground the claim in recognized sources. Prove relationships step by step. Decompose what you cannot yet answer into smaller questions and keep them. Treat uncertainty honestly and keep evidence separate from opinion. Reuse what has been proven so work is never repeated.

That is not a chatbot feature set. It is the beginning of a cognitive discipline, implemented in software. TolmiGround practices that discipline on a narrow class of simple, checkable questions — because that is where you prove a method works before you trust it with anything harder.

Why Start Somewhere So Small

People sometimes hear “is a whale a fish” and wonder how that could possibly be on the road to anything like general intelligence. The smallness is deliberate, and it is the most important design decision I made.

If you want to build a system whose defining virtue is that it is grounded — that it knows the difference between knowing and guessing — then you have to start where “grounded” has an unambiguous meaning and where you can measure whether the system got it right. Simple relationships between well-defined words are exactly that place. The answers are checkable. The reasoning is auditable. The calibration is measurable. You can prove, with numbers, that the system declines to answer when it should rather than bluffing — that is the whole point of a benchmark where it answered 449 of 500 questions, got all 449 right, and walked away from the rest.

A method that cannot be trusted on the simple cases has no business being trusted on the hard ones. So TolmiGround earns the method on the simple cases first. Breadth comes later, and only on top of a foundation you can verify.

TolmiGround as a Component

Here is how the pieces fit. TolmiGround is a grounded reasoning engine for a specific, verifiable class of questions. In the larger system I am working toward, I think of it as one organ, not the whole organism:

  • It is a place where claims get grounded and proven, rather than merely generated.
  • It is a growing, vetted memory — a map of questions and answers that accumulates rather than resets, each entry carrying its status and its evidence.
  • It is a discipline for handling uncertainty: a working example of a system that decomposes what it cannot answer and holds “unknown” as a first-class state.
  • It is a model for safe human participation, where influence is earned per topic, continuously re-tested, auditable, and never allowed to overwrite evidence.

A language model, in this picture, is not the mind. It is one faculty among several — the part that handles fluent language at the door. Grounding, memory, decomposition, and honest uncertainty are not accessories bolted onto a predictor; they are meant to be the load-bearing structure, with prediction serving them rather than the reverse. The long-term work is connecting these faculties into something that reasons across far more than the simple relationships TolmiGround handles today.

I am not going to pretend I have that whole system built. I do not. What I have is a working prototype of the hardest-to-fake part — the part that insists on grounding — with measured results behind it and a patent-pending design. That is the piece you build first, because every other faculty depends on being able to trust what this one proves.

Staying Honest About It

The fastest way to discredit an idea this big is to oversell it, so let me hold myself to the same standard I hold the system to.

TolmiGround answers simple, specific yes/no questions today — not open-ended ones, not opinions, not the open-world reasoning that “general intelligence” really demands. It is only as good as its sources. Its coverage is uneven. Keeping human participation honest at massive scale is unsolved. It is a prototype with a private test system, not a public service. The larger system is a direction I am building toward, not a product I am announcing. Every one of those sentences is true at the same time as everything I wrote above, and I would rather you hold all of them in your head at once.

The patent, for what it is worth, is mainly about getting the invention recognized and on the record. My intent is to keep licensing inexpensive rather than to fence anyone out, with an open, shared base of vetted questions and answers as a long-term goal. I would rather this road get built than own the only lane on it.

The Bet

So this is the bet I am making. The road to systems that reason well does not run only through bigger predictors. There is a second road — grounding first, probability second; evidence before fluency; “I don’t know yet” treated as an answer rather than a failure — and I think it leads somewhere the first road has trouble reaching. TolmiGround is the first working stretch of it.

I would be glad for company on the way.

Join the Mission

If you have read all three parts, you know what TolmiGround is, where it came from, and what I think it is a piece of. If any of that resonates — if you build, research, or just think hard about how machines ought to reason — I would like to hear from you.

Hardened Logic for an Intelligent Era.

You May Also Enjoy