FAQ

Is Corvid a Python wrapper?

No. Corvid is its own language with its own compiler (Cranelift backend), runtime, type system, and stdlib. You can call into Python via the Phase 30 FFI when you need numpy or sklearn, but the language core is independent.

What’s the runtime overhead vs. Python LangChain?

On orchestration-only benches (no model call), Corvid is roughly 25–36× faster than Python and 1.7–2.6× slower than TypeScript on the canonical reference apps. The numbers were published in Phase 17 deliberately as “honest slower” — Corvid is not optimized for raw throughput; it is optimized for the AI-correctness surface. See the benchmarks page for the full numbers.

Does the compiler need network access?

No. Compilation is local. Network access is only for runtime LLM calls (when you actually run an agent), and even then corvid replay reproduces a recorded run with no network.

Which LLM providers are supported?

The model substrate (Phase 31) ships adapters for OpenAI, Anthropic, Google, AWS Bedrock, Cohere, and local models via Ollama. Provider selection happens at deploy time, not in source code. New providers are pluggable via the typed adapter trait.

Can I use Corvid for non-AI code?

You can, but it’s not the point. Corvid’s value is the AI-orchestration surface. For tight numeric loops, systems code, or general-purpose backend code, use the language that fits best — and call Corvid through the cdylib FFI when you cross into the AI surface.

What happens if my LLM provider changes its API?

The model substrate is the boundary. Adapter changes go in the substrate; your source code does not change. The eval surface (corvid eval --swap-model) verifies whether the change altered behavior on your saved traces.

Is approve enforceable across module boundaries?

Yes. The compiler resolves call graphs across modules. An imported tool whose effect row carries trust: supervisor_required requires an approve at every call site, regardless of where the tool is defined.

Can I bypass approve with metaprogramming?

The source-fuzz corpus (crates/corvid-types/tests/source_bypass_corpus.rs) exercises four classes of attempted bypass. Each fails to compile with the matching guarantee_id. New bypass attempts are welcome; file them as adversarial test cases.

What about formal verification?

Corvid v1.0 ships engineering-grade compile-time guarantees: a registry of properties, adversarial tests for each, a separate-binary ABI descriptor verifier, and a CI gate. Formal mechanized proof of the type system is a post-v1.0 research agenda. The security model is explicit about what’s verified by tests vs. what’s verified by proof.

How does this relate to DSPy / LangChain / Pydantic-AI?

Those are libraries. Corvid is a language. They patch over the problems Corvid solves at the type-system level. If you are already hitting the limits of Python’s type system on your AI code, you are the audience.

Is the runtime open source?

Yes. MIT/Apache-2.0 dual-licensed. Micrurus-Ai/Corvid-lang on GitHub.

What’s the road from here to v2?

v1.0 is the launch. The post-v1.0 research agenda includes formal proof of the type system, true second-implementation TCB shrinkage, multi-process worker pools for the rendered backend server, custom middleware injection from Corvid source, and additional connector families. The roadmap lives in ROADMAP.md in the repo.