Start free.
Scale when you need to.

Prova checks whether the argument's claims form a valid dependency graph — no circular reasoning (where a conclusion supports its own premise), no contradictions (where two premises cannot both be true), and no unsupported leaps (where a claim asserts it follows from prior claims that don't actually support it). It does not check whether the facts in the argument are true.

Any chain-of-thought output from any AI model — Claude, GPT-4, Gemini, Llama, and others. Also agent decision logs from orchestration frameworks like LangGraph, CrewAI, and AutoGen. The reasoning can be numbered steps, flowing prose, or conditional branches.

The EU AI Act requires high-risk AI systems (credit, hiring, medical, legal) to provide auditable, explainable reasoning trails. Prova's certificate is a timestamped, formally grounded document that proves a specific AI decision was backed by logically valid reasoning at a specific moment. It fills the gap that logging and monitoring tools leave open.

By default, yes — they're stored alongside the certificate to make the certificate complete and auditable. If you set retain=false in your API request, the reasoning text is processed in memory and never written to disk. Certificate metadata (verdict, confidence, graph structure) is always stored for the certificate to remain valid.

Enterprise plan customers can request that all data (reasoning chains and certificates) be stored in Supabase's EU region, ensuring no data leaves the European Economic Area. Available on Enterprise — contact us to discuss timelines.

Prova uses Čech cohomology to detect structural failures in argument graphs. A logically valid argument corresponds to a graph where H¹(K;ℤ) = 0 — no cohomological obstructions. Circular reasoning, contradictions, and unsupported leaps each correspond to different types of non-trivial H¹. This is formally proved in 2,400+ Lean 4 theorems in the cobound repository.