No fake logos / No invented quotes / Every claim on this page has a receipt

Assay / spell-check for AI behavior

Break this prompt.We’ll catch it.

Assay re-tests your AI prompts every time anything changes, and blocks the deploy when answers get worse. Pick an edit below. It’s the Tuesday-passes, Friday-breaks bug, caught live.

Free for your first 3 projects. No card needed.

Step 1 / Edit the prompt like a teammate would

invoice-extractor / prompt.txt
You extract invoice data.Return ONLY JSON matching the schema:invoice_id, total (number), currency, due (ISO date)model: flagship-large

What the AI answered

{  "invoice_id": "INV-2041",  "total": 42.50,  "currency": "USD",  "due": "2026-11-20"}

Step 2 / Assay’s receipt

Assay inspection

receipt no. 0042 / 50 saved traces

  • Valid JSONPASS
  • total is a numberPASS
  • currency presentPASS
  • due is ISO 8601PASS
TRACES PASSED
50 / 50
DEPLOY STATUS
SAFE TO SHIP
Approved

Baseline. Now try an edit on the left.

In plain words

A driving test for your AI, every time anything changes.

Your AI feature is like a new employee who can change how they work without telling you. Assay makes them retake the test before they’re allowed to touch customers.

  • 01

    Save real examples

    Real inputs from production, each with the right answer attached. That’s your test set.

  • 02

    Re-test on every change

    Edit the prompt, update the model or tweak the schema. All examples replay automatically.

  • 03

    Block what got worse

    If answers broke, the deploy stops and you see the exact examples that failed.

The claim ledger

0 / 4 verified by you

Don’t take our word. Run the test.

Every promise comes with the test behind it. Open a receipt and run it yourself. Figures come from the sample project in the demo above.

  • RECEIPT No. 001

    NOT YET RUN

    TEST: Swap the model, replay saved traces, validate every field against the schema.

    DATA: 50 production-style traces, 4 schema fields each.

    YOUR RUN
    Not yet run. This stays unverified until you run it.

Same change, two timelines

The Friday you avoid.

Without Assay

  1. TUE 14:02Prompt tweaked. “Looks fine” in chat.
  2. TUE 14:10Shipped.
  3. THU 09:30Finance notices totals stored as text.
  4. FRI 17:45Three engineers, one weekend, one rollback.

With Assay

  1. TUE 14:02Prompt tweaked. Pull request opened.
  2. TUE 14:0550 traces replay. 9 fail. Deploy blocked.
  3. TUE 14:20Fix pushed. 50 / 50 pass.
  4. FRI 17:45Nobody is paged. Weekend intact.

Assay is

  • A test runner and deploy gate for AI features that must return strict, structured answers.
  • Model-agnostic. It works with any model you can call from code.
  • Honest about limits. AI-judge scores are estimates, and we show the confidence.

Assay is not

  • A proxy sitting in your request path.
  • A prompt playground. Edit prompts wherever you already do.
  • A replacement for human review on high-stakes outputs. It tells you where to look.

Questions engineers ask first

Does it slow down production?
No. Capture is asynchronous and tests run in CI, never in the request path.
Where does our data live?
Sensitive fields are redacted in your own code before anything is sent. Traces stay in the region you pick, and you choose how long we keep them, from 7 to 90 days.
What does it cost later?
Free for your first 3 projects, then $49 per project each month. Early access members keep that rate for life.

Find out on Tuesday, not on Friday.

Admit one / early access

Free for your first 3 projects. No card needed.

Built with