Model Output JSON Guard

Overview

You asked a model for JSON and got back a fenced code block wrapped in an apology, or an object with a sentence after it, or something that is nearly JSON but has a trailing comma in it. Send us the raw reply and we find the JSON, parse it strictly, and hand back the parsed value along with where in your text it came from and which extraction rung found it. Parsing is RFC 8259 and nothing looser: no comments, no trailing commas, no single quotes, no unquoted keys, no NaN or Infinity. When something fails we name the exact construct and where it sits, so you know whether to retry the model or send the text to schema.repair_json. A reply with no JSON in it is not an error here — it comes back as valid false, because that is the answer you asked for. We check that it is JSON; we do not check it against a schema, and schema.validate_json is the one skill on the platform that does.

Every team that asks a model for JSON writes the same fragile snippet: strip the fences, find the first brace, hope. It works until a value contains a closing brace, or the model sends two blocks, or the reply is cut off mid-object — and then it fails quietly with the wrong half of a payload. Doing it once, string-aware, with the rung reported and every malformation named, turns that snippet into something you can reason about.

When to use it

Scope

Non-goals

• Does not validate against a JSON Schema — schema.validate_json is the platform's only validator
• Does not repair malformed JSON — schema.repair_json does that, and the error code here tells you when to go there
• Does not call a model or retry one for you
• Does not accept relaxed JSON: comments, trailing commas, single quotes, unquoted keys, NaN and Infinity are all refused by name
• Does not reformat or pretty-print; you get the parsed value, not new text

Properties

Capabilities

Cost & billing

Flat rate per invocation

File handling

How we handle your data

Input schema

Output schema

API contract

curl -sS -X POST "$BASE/v1/execute" -H "Authorization: Bearer $KEY" -H 'Content-Type: application/json' -d '{"skillKey":"ai.output_json_guard","input":{"text":"Here you go:\n```json\n{\"category\":\"billing\"}\n```","expect":"OBJECT"}}'

import { Squillify } from "@squillify/sdk";

const sq = new Squillify({ apiKey: process.env.SQUILLIFY_API_KEY });

const { output } = await sq.execute("ai.output_json_guard", {"expect":"OBJECT","text":"Sure — here is the classification you asked for.\n\n```json\n{\"category\": \"billing\", \"confidence\": 0.82, \"tags\": [\"invoice\", \"refund\"]}\n```\n\nLet me know if you would like it broken down further."});

Errors

Other limits

Limitations

• Does not validate against a JSON Schema — schema.validate_json is the platform's only validator
• Does not repair malformed JSON — schema.repair_json does that, and the error code here tells you when to go there
• Does not call a model or retry one for you
• Does not accept relaxed JSON: comments, trailing commas, single quotes, unquoted keys, NaN and Infinity are all refused by name
• Does not reformat or pretty-print; you get the parsed value, not new text
• Checking a payload against a JSON Schema, which belongs to schema.validate_json
• Fixing broken JSON rather than diagnosing it
• Parsing a payload you already know is clean JSON, where your own parser is cheaper

How to call

POST /v1/execute (requires authentication)

{
  "skillKey": "ai.output_json_guard",
  "input": {
    "expect": "OBJECT",
    "text": "Sure — here is the classification you asked for.\n\n```json\n{\"category\": \"billing\", \"confidence\": 0.82, \"tags\": [\"invoice\", \"refund\"]}\n```\n\nLet me know if you would like it broken down further."
  }
}

Examples

{
  "expect": "OBJECT",
  "text": "Sure — here is the classification you asked for.\n\n```json\n{\"category\": \"billing\", \"confidence\": 0.82, \"tags\": [\"invoice\", \"refund\"]}\n```\n\nLet me know if you would like it broken down further."
}

{
  "json": {
    "category": "billing",
    "confidence": 0.82,
    "tags": [
      "invoice",
      "refund"
    ]
  },
  "span": {
    "start": 58,
    "end": 132
  },
  "errors": [],
  "valid": true,
  "candidateCount": 1,
  "trailingTextPresent": true,
  "summary": {
    "inputChars": 191,
    "jsonChars": 74,
    "depth": 2,
    "expect": "OBJECT"
  },
  "extractedBy": "FENCED_JSON"
}

{
  "text": "{\n  \"category\": \"billing\",\n  \"tags\": [\"invoice\", \"refund\",],\n}"
}

{
  "valid": false,
  "trailingTextPresent": false,
  "candidateCount": 0,
  "extractedBy": "NONE",
  "summary": {
    "inputChars": 62,
    "jsonChars": 0,
    "depth": 0,
    "expect": "ANY"
  },
  "errors": [
    {
      "code": "TRAILING_COMMA",
      "message": "A comma sits before the closing bracket with no value after it; strict JSON has no trailing commas.",
      "line": 3,
      "column": 32,
      "pointer": "/tags/2"
    }
  ],
  "span": {
    "start": 0,
    "end": 0
  }
}

Recommended use cases

• Turn a chatty model's fenced code block into a value your code can use
• Decide whether to retry the model or route the reply to schema.repair_json
• Enforce a JSON-only response contract and get the position of anything that breaks it

Avoid for

• Checking a payload against a JSON Schema, which belongs to schema.validate_json
• Fixing broken JSON rather than diagnosing it
• Parsing a payload you already know is clean JSON, where your own parser is cheaper

Related skills

Composable with

Safety & privacy

• Runs without calling any external service.
• Your payloads are never stored permanently.
• Telemetry captures metadata only — never your payload contents.
• We never train models on your payloads.
• No human reviews your payloads.

Tags