// the method
One method.
Several instruments.
I build systems that cannot assert past their evidence. That is one discipline, not one product. It was built for language models, where the failure is obvious and expensive, but nothing in it is specific to them. This page states the method, the criterion for where it applies, and what has actually been built on it so far.
// § 01 · the four commitments
01
Bars before runs
What counts as a pass is fixed before anything is scored, and entered in a public append-only log. A threshold chosen after seeing the result is not a threshold, and now the instrument can tell the difference.
02
Claims gated by what can be witnessed
Nothing is asserted that is not backed by something checkable. The gate is the same whether the claim comes from a model or from a build system.
03
Records the recipient can check
Every result ships bound to the exact bytes it describes, and names the analyser that produced it. Both are checkable from your copy. Checking me must not require trusting me, including about my own tools.
04
Unknown stays unknown
Where the evidence runs out, the system says so and stops, rather than rounding a plausible answer up to a confident one. Most of the engineering is here.
// § 02 · where it applies
The method fits any domain with two properties. First, a claim is being made about a constrained system. Second, those constraints are declared somewhere that can be read. When both hold, the claim can be checked against the declaration exactly, rather than sampled, estimated or asserted.
That is a wide criterion, and deliberately stated as a criterion rather than a roadmap. Listing verticals I have not built would be a claim with nothing behind it. That is the habit this whole method exists to break.
// § 03 · instruments
now
Arcifact Gate
Continuous integration. Computed from the declared workflow: the ways a required check can report success without the validation behind its claim succeeding. The narrowest and most concrete application, and the easiest to check in an afternoon. The current commercial instrument.
origin
Model Evidence
Language models. Fabrication measured on frozen banks and attack-hardened instruments, with an evidence gate that holds generation to what can be supported. Where the method was built, and the deeper of the two.
Each instrument is a contract: a declared source, a claim type, an explicit envelope, and a verifier the recipient runs themselves. Gate reads a workflow and rules on merge closure. Model Evidence reads a bank and rules on fabrication. The envelope and the verifier are the same shape in both.
// § 04 · why two unrelated domains
Because generality is itself a claim, and claims need evidence. A method that only worked where it was invented would be a technique. Measuring what a model fabricates and computing what a merge gate fails to cover have nothing in common at the surface, and the same four commitments produced both. That is the beginning of evidence for generality. It is not proof, and it is not described here as proof.
Research, and where it is going.
Model Evidence · the same method
applied to model evaluation: frozen instruments, digest-bound banks,
an executable scorer
Oracle · a laboratory rather than a
product, and labelled as one
Evidence · how a claim is bound to what
produced it
About · the company, the founder, the
filings
Gate is the commercial product. These are the programme it came from, kept separate deliberately: a buyer evaluating a CI tool should not have to read a research agenda first, and a reader interested in the research should not be sold to.