Problem
Completion today is engine-success + satisfied stop hooks; there's no shared concept of evidence-backed completion, which is the gap that lets a green run still be broken.
Proposed solution
A small, optional outcome + evidence vocabulary other orchestrators can target: the deterministic gate result, what was and wasn't checked, and an honest "not verified" disclosure. It also gives a future evals / "test-plugin" surface a completion contract to assert against.
Alternatives considered
Leaving completion meaning entirely to each consumer's stop hooks, which is the status quo and reproduces the green-but-broken failure mode across every orchestrator.
Area
Orchestrator
Additional context
No response
Problem
Completion today is engine-success + satisfied stop hooks; there's no shared concept of evidence-backed completion, which is the gap that lets a green run still be broken.
Proposed solution
A small, optional outcome + evidence vocabulary other orchestrators can target: the deterministic gate result, what was and wasn't checked, and an honest "not verified" disclosure. It also gives a future evals / "test-plugin" surface a completion contract to assert against.
Alternatives considered
Leaving completion meaning entirely to each consumer's stop hooks, which is the status quo and reproduces the green-but-broken failure mode across every orchestrator.
Area
Orchestrator
Additional context
No response