https://mateopetel.substack.com/p/provability-fabric-the-saf...

Most “AI safety” today is vibes. Prompts, heuristics, and dashboards don’t scale when agents can call tools, stream outputs, or mutate state.

Provability Fabric introduces a different contract:

Proof-carrying bundles (PAB-1.0): specs, Lean proofs, SBOM, and provenance in one signed package

ActionDSL → Lean 4: policies compile into proof obligations and runtime monitors

Complete mediation: every effect passes through a Rust sidecar enforcing the same semantics the proofs assume

Egress certificates: every emission leaves with a signed verdict — pass, fail, or inapplicable

The principle is simple:

Safety = correspondence between proofs and runtime enforcement.

Would be interested in HN’s view:

Should deployment of AI agents require proof-carrying bundles?

Where should falsifiability end — policy layer, runtime, or both?

What would make you trust an agent in production?