Your LLM’s JSON output will drift.
Catch it before prod.

The drift nobody notices

You ship a feature on top of response_format: json_schema or tool calls. It works. Then you bump the model — gpt-4o → gpt-5, Claude Sonnet 4.5 → 4.6 — and the JSON quietly changes shape:

• •a number comes back as "0.9" — a string
• •a required field goes missing in 2 of every 10 responses
• •a new enum value appears that your parser doesn't handle
• •the model adds an extra field you never asked for

Your JSON.parse succeeds. Your app breaks downstream. Nobody noticed.

Paste your outputs. Get a per-record report.

  ✓ 8 passed   ✗ 2 failed   of 10

  ✗ output #3
      "confidence": expected number, got string ("0.92")
      → model returned a quoted number — use z.coerce.number()
  ✗ output #7
      missing required field "sources" (expected string[])
      ⚠ unexpected field "reasoning" appeared

  wrong type ×1 · missing field ×1 · extra field ×1

• No schema yet? Click Infer schema — TypeMorph builds a Zod schema from your known-good responses.
• Human-readable diagnosis, not a cryptic stack trace — with a suggested fix.
• Batch / drift mode: feed a .jsonl of logged outputs and see exactly how many drifted, and why.

In your code, or your CI

# Validate logged outputs against your schema
typemorph validate schema.ts responses.jsonl

# Bootstrap a schema from known-good responses
typemorph validate --infer good-responses.jsonl --out schema.ts

# In CI — advisory PR comment, fails the job if outputs drift
typemorph validate schema.ts fixtures/llm-outputs.jsonl --format github

Why local matters here

LLM outputs often contain your users’ data — prompts, PII, proprietary content. Most online “validators” upload it to a server. TypeMorph runs entirely in your browser tab, or in your own CI runner. No upload. No API key. No retention. Nothing to leak.

Honest limits

TypeMorph validates structure: types, required fields, nulls, shapes — and flags enum/format drift as warnings (it never hard-fails an unknown value). It’s a fast safety net and a great starting point, not a replacement for the real Zod schema you own. Inference from samples can’t know every valid value, so treat it as advisory: a signal in your PR, not a gate you blindly trust.

FAQ