Bit Ghost Security

Hardened Logic for an Intelligent Era

6 minute read

This is Part 1 of a three-part series about TolmiGround, an evidence-grounded question-answering system I have been building. In this first part I describe what it is and what it does. In Part 2 I tell the story of where the idea came from. In Part 3 I explain why it is only one piece of a larger system I am working toward. TolmiGround is patent pending — U.S. Provisional Patent Application No. 64/181,356, filed October 21, 2026.

The Problem I Kept Coming Back To

Spend enough time around large language models and you notice something unsettling. They sound exactly as confident when they are wrong as when they are right. The sentence is fluent either way. The tone is sure either way. And when you ask where an answer came from, there is usually no honest answer to give, because the model did not retrieve a fact — it predicted a likely string of words.

That is a remarkable engineering achievement and a genuinely poor foundation for trust. If I cannot tell the difference between a confident right answer and a confident wrong one, the confidence is worthless to me. Worse, most systems collapse two very different situations into a single number: “there is no evidence for this yet” and “there is evidence on both sides” come out looking the same. A security engineer learns early that unknown and false are not the same state, and treating them as one is how you get burned.

Crowdsourced knowledge has the opposite problem. Open the question up to a vote and you learn what is popular, which is not the same as what is supported. A coordinated group, a loud majority, or a well-meaning person answering outside their expertise can all move the needle in the wrong direction.

TolmiGround is my attempt to build something that avoids both failure modes. It is not a chatbot. It is an answer engine, and its defining habit is that it would rather say “I don’t know yet” than guess.

What TolmiGround Actually Does

At its core, TolmiGround answers simple, specific yes/no questions about how things relate to one another — the kind of question that has a checkable answer. A few of the formats it handles today:

Question shape Example
Is a… “Is a whale a fish?”
Does… have… “Do birds have feathers?”
Can… “Can penguins fly?”
Part of “Is a wheel part of a car?”
Made of “Is glass made of sand?”
Same meaning “Does ‘canine’ mean dog?”
Opposite of “Is hot the opposite of cold?”
Used for “Is a hammer used for driving nails?”

Each one can be asked in three strengths — “all,” “typically,” or “some.” Can all birds fly? is a different question from Can some birds fly?, and the system treats it that way, because that distinction is exactly where a lot of confident wrong answers come from.

The answers do not come from a model’s intuition about what sounds right. They come from evidence the system can trace back to recognized reference sources — a standard English dictionary and thesaurus, a scientific classification of living things, a commonsense knowledge base, and a curated list of core facts. Automated checks look things up and prove a relationship step by step. A dog is a mammal not because that sounds plausible, but because the taxonomy says so through a chain you can follow and audit.

A language model is still in the loop, but its job is deliberately small: it translates your typed question into a structured one. It never decides the answer. Anything it contributes is set aside before the answer is determined. The model is a translator at the door, not the judge in the room.

Two Scores, Never Blended

Here is the design decision I am most attached to. TolmiGround shows you two separate scores, side by side, and it never mixes them.

  • The evidence score comes only from the reference sources and the automated checks. This is what the answer is actually based on.
  • The people score shows what trusted human participants think.

They sit next to each other. They are never averaged into one tidy number, because averaging them would quietly let opinion contaminate evidence. The evidence side also keeps support, counter-evidence, and still unknown as distinct quantities, so the system can look you in the eye and say “untested” instead of manufacturing a guess to fill the silence.

And when the two scores disagree — when trusted people push back against the evidence — TolmiGround does not hold a vote. Disagreement is treated as a signal that something needs more checking. Maybe a word has two meanings. Maybe someone read “all” where the evidence only supports “some.” The minority is never simply outvoted. The response to conflict is more work, not a show of hands.

Saying “I Don’t Know” Is the Feature, Not the Bug

When the evidence is thin, TolmiGround says so — and then it does something useful with the admission. It breaks the hard question into smaller questions, each with its own status, and keeps working on those. Those sub-questions are saved and reused whenever a related question comes up later, so the work is never thrown away and never repeated.

This is why I describe the whole thing as a system that builds knowledge carefully rather than generating it on demand. The output is a growing, vetted map of questions and answers, each one carrying a status: answered, pending, or contested.

What I Can Honestly Claim

TolmiGround is a working prototype with a private test system — not a public product, and I want to be precise about that. But the prototype has been measured, and these are the results I stand behind:

  • On a benchmark of 500 test questions, it answered 449 and got all 449 right. It declined the remaining ~9% because the evidence was not sufficient. That is the behavior I was after: it says “I don’t know” instead of guessing.
  • Its confidence is well calibrated — when it reports high confidence, it is right about that often. The confidence number actually means something.
  • On a set of newer question formats, it answered 24 of 38 trial questions, all correct.
  • It has pre-created a map of roughly 1.8 million simple questions out of the links already present in its sources, with about 89% answerable from those sources on creation.

I am just as interested in the limits. TolmiGround only handles simple, specific yes/no questions — not open-ended or opinion questions. It is only as good as its sources; in testing, nearly every wrong answer traced back to an error in a reference source rather than in the reasoning. Coverage is uneven — some formats, like “used for,” have thin sources so far. And keeping human participation honest at very large scale is an open, ongoing problem. This is a prototype, not a finished service, and I would rather you hold me to that than to a demo.

Why the Name

TolmiGround is named for Stephen Toulmin, the philosopher who modeled how arguments are actually supported — by grounds and warrants, not by volume — joined with “ground,” as in grounding a claim in evidence. The name is the thesis: an answer is only as good as what it stands on.

What’s Next

That is the what. In Part 2, I want to tell you the why — the idea that started this, which is older and stranger than a product spec, and which began with a thought about whether you could, in principle, write down every possible question and answer. And in Part 3 I will put my cards on the table about what TolmiGround is really a piece of: an attempt to build something that reasons from grounded evidence rather than probability — something that, if it works, would act in a manner closer to general intelligence than a system that only predicts the next likely word.

Join the Mission

TolmiGround is early, and I am building it in the open where I can. If you care about systems that are honest about what they know, I would like to hear from you.

Hardened Logic for an Intelligent Era.

You May Also Enjoy