Does a green verification authorize anything?
No. Every result is structural evidence. Identity, tenancy, live
policy, effect ownership, execution, billing, and evaluation truth are
host-owned surfaces; the Evidence record names them in not_verified.
Why canonical-first decoding? Digests are only meaningful over exact bytes. Non-canonical bytes deny before any semantic read, so two spellings of the same value cannot both be "the artifact".
What happens to extensions I don't know? Optional extensions quarantine: retained byte-exactly, typed as unscanned, never executed. Critical ones deny — an unknown critical surface is a surface you cannot enforce.
Can I widen a bound for a special case? No. Effective bounds never widen host policy. Narrow, clamp with evidence, or deny — the algebra is property-tested.
How do I upgrade an exact-pinned release without breaking
historical replays?
Bind the replay-identity fields (manifest_version,
verification_semantics_version, protocol_revision,
registry_digest, the artifact digests, effective_bounds) and record
the provenance fields (package_version, package checksum,
corpus_digest, spec_digest) as metadata. Upgrades between releases
that share a semantics version and an unchanged registry census replay
green; registry movement conflicts by design (it is the fail-closed
vocabulary detector); a true semantics change conflicts loudly on the
integer. The full partition and its reasoning:
upgrading guide.
Why is the registry compiled in? Registry content is a code release: drift between shipped code and a shipped registry file would be unrepresentable. The governance-canonical json is bound to both compiled twins by the registry-equality gate.
How do I run the corpus in my own language? The runner protocol ships as verifier/RUNNER.md: corpus layout, load-time integrity, the report format, and the byte-agreement contract. The standalone kit (verifier + corpus) is attached to every release tag.