Averment

Safety and execution control infrastructure for AI automated systems.

Beta — This SDK is under active development. APIs may change between releases.

Install

npm install averment

Quick Start

import Averment from "averment";

const averment = new Averment({
  apiKey: process.env.AVERMENT_API_KEY,
  engine: "dde-test",
});

// Safe execution — only runs if allowed
await averment.decide(
  { summary: "AI agent deleting production records" },
  async () => executeAction(),
);

What It Does

Averment sits between intent and execution. It evaluates proposed actions in real-time using hybrid (both LLM and deterministic) reasoning, behavioral pattern detection and compiled policy evaluation.

Every decision returns an enforcement action:

| Action | Meaning | | --------- | ------------------------------------- | | allow | Safe to execute | | caution | Moderate risk; proceed with awareness | | review | High risk; requires active attention | | block | Must not execute |

Usage Patterns

With handlers

await averment.decide({
  summary: "User initiating payment",
  context: { amount: 50000, device_new: true },
  handlers: {
    allow: async () => processPayment(),
    review: async () => requireOTP(),
    caution: async () => queueForReview(),
    block: async () => denyAction(),
  },
});

With strategies

await averment.decide(
  { summary: "User login from new device", context: { device_new: true } },
  async () => createSession(),
  {
    strategy: "confirm_review",
    confirmReview: async () => await verifyOTP(),
  },
);

BYOLLM

const averment = new Averment({
  apiKey: process.env.AVERMENT_API_KEY,
  engine: "dde-test",
  llm: {
    provider: "openai",
    model: "gpt-4o",
    apiKey: process.env.OPENAI_API_KEY,
  },
});

Your LLM credentials are never stored or logged.

License

MIT