Task
Purpose, scope, constraints, owner, and acceptance criteria.
active development / public preview
APCP Conductor coordinates people, independent AI collaborators, and a bounded execution agent — through structured packets, exact-action approval, and a record that doesn't move once it's written.
canonical artifacts
Shape identifies what an artifact is. Color identifies its condition. Neither depends on the other alone.
Purpose, scope, constraints, owner, and acceptance criteria.
What was produced, observed, or learned in response to a Task.
A choice, approval, rejection, or disposition with the reasoning kept intact.
Expected state, actual state, evidence, warnings, and outcome.
Canonical material that informs work without duplicating it.
An explicit hold that preserves state, reason, and restart condition.
four roles, deliberately unequal
Something can be able to act without being allowed to decide. That split is the whole architecture.
Grants scope, sets cost ceilings, approves anything irreversible. No packet reaches an authority state without this.
Holds the packet lifecycle, the audit trail, and every decision's evidence. The only place state actually lives.
Separate AI systems analyze, critique, and propose — without collapsing into one unverifiable answer.
Sensory and motor access to a real environment. No standing authority, no memory of its own between actions.
where this actually is
The six packet types, review queue, audit log, validation snapshots, bundle import/export.
The execution interface, observation channel, and a thin approval client for the one thing that still needs a human's actual attention.
Secure remote execution, cross-model handoffs reviewed the same way local ones are, real deployment discipline.
apcp conductor
There's no polished demo yet — there's a working protocol, a real audit trail, and a lot of decisions that were argued out, not assumed. If that's the kind of thing you want to watch or contribute to, say so.