what is guardrails-core fr?

guardrails-core is a small Python and JavaScript library that decides whether a money-moving action should proceed. It targets two related problems: AI agents that transact on your behalf, and scammers who try to trick you into sending money yourself. In both cases something proposes an action and this engine decides if it should happen. The core package has zero external dependencies. The core idea is that a language model proposes, but deterministic code decides. A model emits a structured Intent describing the proposed action. The engine evaluates that Intent against rules and produces a Verdict. The model never signs or executes anything directly. Rules return observations called signals, each with a severity score. The engine combines those signals into a probability that approaches but never reaches certainty, so it never labels something simply "safe." Absence of warning signals means no signals were found, not a guarantee. The engine evaluates artifacts rather than identities. It looks at addresses, domains, links, contracts, and message patterns instead of profiling who is on the other end. The authors argue that recognizing a scam pattern is what stops a transfer, while identifying the person adds privacy risk without much protective value. Every evaluation produces an audit record showing what was proposed, which rules fired, and what was decided. For agent-proposed actions, a spending envelope always applies. Higher autonomy means fewer confirmations for small things, never unlimited spending. The library ships in Python and JavaScript with no build step. The JavaScript version exists so analysis can run on a user's device, avoiding transmission of private messages. Both implementations share a single conformance test suite, so a divergence between languages fails a test rather than surprising a user. Adapters translate between the engine and the outside world, attaching external facts like domain age or address history. The project is early, with a small rule library and settled interfaces.