Language models predict what comes next.
The Ground Model knows what things stand on.
An instrument that sits outside every model. Point it at a body of work and it reports what that work rests on, how many layers sit between it and checked ground, and how much it cannot account for. When the ground cannot carry the work, it declines and says why.
Evaluators are now inside the labs. What they publish can still be redacted.
An evaluator reads the work, forms a view and writes it up. If the finding rests on material the lab marks confidential, the redaction takes the ground with it. Readers are left with a conclusion to trust.
The finding carries a record of what it stands on, taken by an instrument outside the lab and outside the model. A record like that can travel without the confidential material. The lab keeps its secrets, the evaluator keeps the finding, and the reader gets something to check.
Three steps, all outside the model.
Point it at the work
A repository or a package. The same method has also read a library of machine-checked math. It reads the work itself, never the model that made it.
It traces each piece down
Every part is followed to what it depends on, layer by layer, until it reaches a declared floor of sources that have already been checked.
You get a verdict and a record
One of three verdicts, the measures behind it, and the exact revision and floor it was read against.
The unit of value is the record, not the token.
Every read ends in the same record, so anyone you hand it to can see what it was checked against. A signed certificate an evaluator can put their name to is what we are building next.
Anyone asked to trust work they did not make.
Independent evaluators
Teams checking AI from the outside, and the labs that seat them. A finding with its ground attached, that holds through redaction.
Teams shipping AI-written code
See what a repository stands on before you depend on it. Public repositories will be free to read.
Publishers, funders and research
Check the ground under a claim before it travels further under your name.
Is this a language model?
No. It does not write text. It reads what work stands on and records the answer.
Does it tell you whether the work is correct?
No. It tells you what the work stands on, not whether it is right or secure. That is a different question, and we keep them apart.
Does it need access to a lab's model or training data?
No. It reads the work itself. That is what lets the certificate hold without the confidential material.
What happens when it cannot tell?
It declines and says why. A decline is a result, and it is recorded like any other.
What does a read cost?
Public repositories will be free to read. Private and large ones will carry a flat fee per read while we build. Evaluation teams, ask us.
Who is behind it?
MirrorMind, Inc. Built by Rollie Stanich and Cory Smith. Backed by Evan Sharp. True Read is the product built on the Ground Model.
Bring us the work you are asked to trust.
Then check the certificate yourself.