Your agent's cheap model just confidently produced a wrong number — and nothing errored.

Assay checks every step against reality — it runs the test, it recomputes the arithmetic — not another AI's opinion. It catches the silently-wrong step, pinpoints it, and fixes only that step: retry the cheap model with the reason, then escalate to a stronger model on another provider. You keep your model and your price.

👇 A cheap agent settles a trade in 3 steps. Watch the left wire the wrong amount — and Assay catch it.

⚡ caught & escalated⛔ runaway killed⚐ flagged for a human✓ verified against reality

● Cheap model alone

one cheap model, no safety net — what most agents do today

● Cheap model + Assay

same cheap model — Assay verifies each step against reality and recovers the failures

How you connect — one line, keep your own key

# You pick your model. Assay verifies it. Any provider.
client = OpenAI(
    base_url="http://localhost:8129/v1",   # <- the only change (runs today)
    api_key=YOUR_KEY,                       # your key, passes through
)

Not a router. Routers pick a model up front and lock it. Assay keeps the model YOU chose, checks each step against reality, and escalates only the failures — across providers a single lab won't reach.