Architecture
Orchestration proposes work; authorization bounds it; the verifier checks it; the operator retains the final decision.
Private internal project. Features described here are evidence-limited and are not offered as a public execution service.
Agent and supervisor
The agent interface manages tasks and receives model outputs. The supervisor starts child work under explicit tool restrictions rather than giving every child the root session's host capabilities.
Bounded coding lane
A coding proposal is limited to an approved source snapshot. Visible checks may provide concise failure feedback. Independent withheld checks determine acceptance and are not presented to the proposing worker.
Evidence trail
Execution records include the tested source revision, requested scope, candidate artifact hashes, and signed result. Signatures provide tamper-evident provenance within their trust assumptions; they do not prove code correctness or complete isolation.
Why separation matters
A model's assertion that a patch works is insufficient. The architecture separates proposing, running, grading, and approving so that a single model response cannot authorize a privileged change or declare its own success.
Limits
Measured local demonstrations do not prove distributed throughput improvements, comprehensive container hardening, or production-safe unattended operation.