Straight answers

Is this a chatbot?

No. There's a conversational surface, but it's the least interesting part. The product is the 11-layer reasoning, the deterministic tools, and the closed learning loop. The model formats; the tool decides.

What does 'substrate-agnostic' mean?

The same agent works whether your fabric is EFA, InfiniBand, RoCE, plain TCP, NVLink inside one box, or a single workstation. Every lesson is tagged with where it applies, so knowledge learned on one substrate never leaks into a place it doesn't belong. That is what substrate-agnostic means: knowledge that knows where it holds.

How do you avoid hallucinated numbers?

The language model never invents a formula or a benchmark. It gathers inputs, calls a deterministic tool, and formats the result. Any number that reaches you carries a provenance stamp: commit, image digest, command, timestamp. A value without provenance fails the output gate.

What happens at an irreversible or costly step?

The agent stops. Anything touching spend, capacity, destructive operations, credentials, or evicting someone else's work is human-gated: it prepares the exact action and hands you one command to run.

Which fabrics and clusters are supported?

EFA, InfiniBand, RoCE, TCP, NVLink, and single-host, on managed Kubernetes (e.g. EKS) or any conformant cluster. These are examples of its reach, not the definition of the product.

What are the 11 layers?

L0 compute kernel, L1 launch config, L2 framework, L3 communication library, L4 transport plugin, L5 CI / build environment, L6 network fabric, L7 OS / infra / driver, L8 benchmark harness, L9 autonomous loop, and L10 knowledge base. Eleven layers, kernel to knowledge base; the interactive stack walks each one.

Can I use it today?

The read-only recall plugin is the entry point; the plan and orchestrate lanes are early access, invite-only while the coverage ratchet climbs. See the roadmap for an honest view of what's proven versus still being earned.

Still have a question? Read the docs or request access.