Python Trust Boundary
Python is responsible for modeling, orchestration, candidate search, and presentation. It is not allowed to declare its own heuristic candidate a proof.
exact Python claim
│
▼
normalization + semantic digest
│
▼
untrusted candidate search ──► candidate bound / cutoff / Krawczyk data
│
▼
negotiated LeanCert checker
│
accepted fixed certificate
│
┌───────────────────┴───────────────────┐
▼ ▼
typed Python outcome exported Lean project
│
▼
independent kernel rebuild
What Verified means
A v1 Verified result means:
- the request was a closed, normalized exact claim;
- the Bridge advertised the operation and schema used;
- the response matched the negotiated backend, certificate schema, and verification route;
- the certificate payload agreed with the original request; and
- the named checked operation accepted it.
It does not mean every Python module, search heuristic, NumPy operation, or compiler optimization has joined the trusted computing base.
Compiled checking and kernel replay are separate events
The bundled Bridge reports a compiled_checker verification route. Exporting
and rebuilding a fixed certificate is a second event: the generated project
kernel-reduces the retained checker input, applies the soundness theorem, and
uses #assert_trust kernel on the resulting theorem.
Do not relabel the original Bridge result as a kernel-replay result. Record both events when an audit requires both.
Provenance
Verified outcomes retain:
- Python claim digest and normalized claim;
- Bridge API and protocol versions;
- Lean and LeanCert versions;
- source revision and source digest;
- build-environment digest and profile;
- resolved Lean toolchain and LeanCert dependency revision; and
- the negotiated capability identity.
Run leancert doctor --json to inspect the installed runtime independently of
a particular proof.