It makes the work and checks itself as it goes, against the spec and what was actually asked. That's the honest first pass, and it can't see its own blind spots.
the pilot
It re-checks every step, and no LLM does the checking. It passes nothing it can't tie back to the contract, and fixes only what it can prove is safe.
the watcher
It answers to the contract, not to the pilot. It checks the result from scratch and holds back anything that can't show it conformed.
the point
Whatever makes the work shares its blind spots. The checks that catch them are the ones that owe it nothing. Three catches, not one.
the pass
Every output earns the same pass.
Intent goes in as a contract. Work advances one gate at a time, each step checked
and written down. What comes out is locked, audited, and traceable back to the ask. Nothing changes in silence.
Intent enters
The handover becomes the contract: the spec, the rules, and what "right" means for this task. Everything downstream answers to it.
Gated, step by step
Work advances through checkpoints. Each one has to pass before the next opens.
Checked against the contract
The pilot verifies conformance; the independent watcher verifies the pilot.
Locked & audited verified
What ships is sealed, with a record of every decision behind it. Nothing leaves unproven.
the breadth
One discipline, whatever AI produces.
The output can be anything a model emits. The standard doesn't bend to fit it.
The same proof lands on a document, a code change, or an analysis the same way.
DocumentsA brief, a report, a spec. Checked to say what the ask required, and to claim nothing it can't support.
CodeA change, a review, a build. Checked to do what the spec says, and to break nothing that already worked.
AnalysisAn audit, a finding, a dataset. Checked so the conclusion actually follows from the evidence under it.
Anything elseWhatever else a model emits. The standard doesn't bend to the format; the same proof lands on all of it.
runs anywhere
Change the model. The proof holds.
Open or frontier, cloud or your own hardware, Praxis doesn't care which model
made the work. The substrate is swappable; the verification stays constant. Sovereign and on-prem are just one
end of the same dial.
claude
substrateany model,
swappable
wherecloud or your own
metal
posturethe verification,
constant
the category
Models commoditise. Proof is what lasts.
Generation gets cheaper and more interchangeable by the month. The durable
layer isn't which model you used. It's the independent proof that what shipped was right. That's the part that
holds its value when everything underneath it converges.