Safety And Portability

The runtime enforces several invariants independently of the mineable ABI. They remain fixed while candidates compete on memory quality and efficiency.

• Canonical event writes remain under runtime control.
• Every derived artifact cites its source events and is schema-validated.
• Retracted, deleted, expired, superseded, or out-of-scope evidence is excluded before serving.
• M5 may rank only material admitted by M4, and M6 may select only material surviving M5.
• Unknown, duplicate, fabricated, or malformed references fail closed.
• The runtime constructs the final context and rechecks its rendered budget.
• Retrieved memory is fenced from automatic reinjection as new source evidence.
• Tenant, user, agent, profile, and optional session scopes are enforced at the adapter boundary.
• Evaluation hooks run with bounded compute and derived-storage allowances inside the networkless evaluator sandbox. The ordinary pip serving worker is process-isolated but does not provide that kernel boundary. Use trusted modules or external confinement on a consumer host; a verified snapshot establishes module identity and admission, not permission to execute hostile code.

Canonical neural replay is qualified for Linux amd64, CPython 3.10, using the release's exact numeric-runtime bundle. Installed native and Python payloads are checked by hash as well as version. ARM CPython 3.10 is serving-qualified; it is not a claim of byte-identical canonical neural replay. The existing amd64 tests do not establish equality on every CPU instruction-set subset.