Field note 07 / 10
The Witness-to-Claim Spine
The project began to move again when it narrowed the machine: witness behavior, derive a claim, share a vector, replay it in ports, preserve lineage, and leave every step inspectable.
The active path became clearer when the project stopped asking archives to do living work.
That sentence hides a lot of pain.
The earlier dream wanted breadth: more packages, more generated artifacts, more languages, more proof-looking output. The working path demanded sequence before scale. If the project could not explain exactly how one observation became one claim and one vector, then more languages would only multiply confusion.
The new spine was narrower:
Observe a witness.
Preserve the observation.
Derive a candidate claim.
Curate it.
Materialize a shared vector.
Make a port satisfy the vector.
Verify the lineage.
The sequence looks obvious after it exists. In the middle, it was not obvious at all. The tempting path was always to let a rich source archive, a generated projection, or a clever parser smuggle more meaning than it had earned.
The spine forced the project to slow down in the right place.
What exactly was observed? What exactly was derived? What did tooling parse from witnessed bytes? What was excluded? Which language consumed the vector? What does the evidence prove? What does it not prove?
Those questions became more important than any single implementation.
The public story has to stay deliberately modest. A witness can provide bounded behavior evidence. Tooling can interpret observed bytes only when the artifact admits that interpretation is not the source of truth. A port can satisfy a shared vector. Other languages can replay the same expectation or stress it from an unfamiliar angle.
None of those roles is an oracle.
That is the point.
The port is pressure. The vector is contract. The claim is narrow. The witness is evidence. Review remains human.
This is the machine the private repo is being prepared to show.
It is also where the working method started to show its bones. Every step had a card. Every card had a boundary. Every boundary had to survive a cold reader, not just the person who remembered why it existed.
But a machine that only works in one language is still too comfortable. The next test was whether other languages would expose the assumptions hiding in the first success.