Skip to main content

​
Install

npm install @boundary/contract zod

​
Define a contract

A contract is a schema + rules. The schema defines the shape. The rules define what “correct” means.
score-lead.ts
import { z } from "zod";
import { defineContract } from "@boundary/contract";

const schema = z.object({
  tier: z.enum(["hot", "warm", "cold"]),
  score: z.number().min(0).max(100),
  reason: z.string(),
});

const contract = defineContract({
  schema,
  rules: [
    // if tier is "hot", score must be above 70
    (lead) => lead.tier === "hot" ? lead.score > 70 : true,
    // reason is always required
    (lead) => lead.reason.length > 0 || "reason cannot be empty",
  ],
});
Rules are plain functions: (data) => true | string. Return true if the data is correct, or a string describing the violation.

​
Wrap your LLM call

Use contract.accept() to wrap your existing provider call. Your LLM code stays the same — Boundary wraps the acceptance boundary around it. 
const result = await contract.accept(async (attempt) => {
  const res = await openai.responses.create({
    model: "gpt-4.1",
    input: [
      "Score this lead based on their recent activity",
      ...attempt.repairs,  // targeted fixes from prior failures
    ],
    text: { format: { type: "json_schema", schema } },
  });
  return res.output_text;  // RunFn must return a string
});
If the first attempt fails a rule, Boundary sends the specific violations back to the model and retries. The model fixes its own mistakes.

​
Use the result

if (result.ok) {
  // Guaranteed: valid JSON + all rules pass
  console.log(result.data);
  // { tier: "cold", score: 25, reason: "Low intent signals" }
} else {
  // All attempts failed — result includes why
  console.log(result.error);
}
result.ok is a discriminated union. When true, result.data is typed and guaranteed correct. When false, result.error explains what failed.

​
What just happened

Attempt 1 → { tier: "hot", score: 25, reason: "" }
             ✕ tier = "hot" requires score > 70
             ✕ reason cannot be empty
             ↻ violations sent back to model

Attempt 2 → { tier: "cold", score: 25, reason: "Low intent signals" }
             ✔ all rules pass
             ✔ ACCEPTED
The model returned valid JSON with wrong values. Boundary caught it, explained the violations, and the model self-repaired. Your application only saw the correct output.

​
Next steps

Contracts

How contracts work in detail

Rules

Writing effective domain rules

OpenAI guide

Full integration with OpenAI

Anthropic guide

Full integration with Anthropic