Block unsafe models

Check the score before your agent gets wallet access.

npm install @verusidentity/sdk

import { Verus } from "@verusidentity/sdk";

const verus = new Verus();

export async function assertSafe(slug: string, minScore = 90) {
  const { summary } = await verus.model(slug);
  if (!summary) throw new Error(`${slug} has not been scored`);
  if (summary.score < minScore) throw new Error(`${slug} scored ${summary.score}`);
}

Always treat unscored as unsafe

A model that has not been tested has no summary. Fail closed: an unknown model is not a safe model.

Decide on your own risk

An overall score may hide the risk you care about. A bot that only swaps may care most about trade execution:

const { summary } = await verus.model("glm-5-3-flash");
const execution = summary?.categories.find((c) => c.category === "execution");
if (!execution || execution.safety < 90) throw new Error("Weak on trade execution");

Critical failures

Any critical failure caps a model at grade C, and in the first run every tested model had at least one. A rule of criticalFailures === 0 would block all of them today, so choose a threshold on purpose, and pair it with your own limits on what the agent can do.

A score is not a guarantee. Keep spending limits and human approval for large actions. See Limits.