FAQ

FAQ

Is this a framework I install and use in production?

No. It is a reference blueprint: docs + framework-agnostic specs + a small runnable reference. Steal the contracts, implement them in your stack, and keep the reference as a conformance oracle. It favours legibility over throughput and ships no warranty (see LICENSE).

Which model / LLM does it use?

None, by default. The Capability Provider Model makes the model a swappable provider, and the reference ships a deterministic offline EchoProvider so everything runs with no key and no network. Plug in a real model by writing one provider class — see How To Extend.

How is this different from LangChain / AutoGPT / CrewAI / a typical agent framework?

Those optimise for capability (more tools, more autonomy). This blueprint is about the other half — accountability: a fail-closed policy surface on every action, and outcomes graded against evidence rather than the model's self-report. It is structural and model-agnostic, not a prompt or persona layer. The two are complementary: you could run such a framework as a provider underneath.

What does "capability must never outrun accountability" actually mean in code?

Concretely: nothing acts without passing the Policy Hook Surface (fail-closed, audited), and no mission is credited with success unless the Reality Grading Loop confirms it against evidence. As you plug in a stronger model, those guards matter more, because a stronger provider can do more damage per unpoliced action.

Does it need a GPU or internet?

No. Python 3.9+ and the standard library. CI proves it on 3.9 / 3.11 / 3.12.

Why "fail-closed"? Isn't that annoying?

An error in a safety check that lets the action through anyway is security theatre. If a policy can't decide, the safe answer is "no." It is deliberately not disable-able.

Can a human be part of the loop?

Yes — a human is modelled as the highest-authority Capability Provider (typically for governance or high-stakes actions), and only the human principal may ratify a change to the rules.

What's the license? Can I use it commercially?

Apache-2.0. Yes, including commercially — keep the license and NOTICE, and note any changed files.

How do I contribute?

See Contributing and the repo CONTRIBUTING.md. The one rule: a change must increase either reliability or evidence.

Where do I report a problem or ask a question?

GitHub Issues.