Zivk00 Labs
All systems

Trust-aware AI work routing

ZivLabs

Open source · public V1

What it does

A dependency-free multi-AI workspace: AgentLane routes by capability and trust, ContextForge keeps context grounded, and delivery fails closed without current evidence.

Read the source on GitHub

Proof

A delivery is not a model answer.

The public core stays local by design. It allows delivery only when the conditions below are independently present.

A bounded candidate change

CheckObserved stateDecision
trusted execution surfaceREQUIREDRequired
source-bound handoffCURRENTRequired
non-empty diff receiptPRESENTRequired
declared check receiptPASSRequired
model response aloneUNVERIFIEDNot trusted
stale handoffSTALERefuse

This is a documentation example of the public policy, not a claim that a provider ran or that a change was delivered.

Case study

  • Public V1 modules cover routing, context, execution, verification, delivery, persistence and a desktop shell
  • Deterministic demos and current CI exercise the same orchestration contracts
  • MIT licensed and dependency-free at the public core
Problem
A useful AI worker still needs a reliable answer to a simpler question: may this proposed change cross the delivery boundary?
Approach
Keep routing, handoff freshness, verification, and delivery as small pure functions. A host application supplies any provider adapter and its independently proven isolation boundary.
Architecture
A task signature selects an eligible execution surface. A source-bound handoff and a receipt containing both a diff and a declared passing check are required before delivery is allowed.
What went wrong
A policy library cannot create isolation by itself. Treating a provider name or a successful command as proof of confinement would make the boundary fictional.
Result
The public V1 makes the whole contract inspectable: route by capability and trust, ground continuation against the current revision, and deliver only after verification.
Limits
The repository is a host-neutral foundation, not a claim that providers ran or that a host has proven confinement. Production adapters must establish those boundaries separately.

Contact

Building an AI system or a product people must trust?

Open to engineering roles, product missions, and creative technical collaborations.

vlad@zivk00-labs.com