Under the hood

Every AI service says "our AI" and shows its best examples. Underneath there may be someone else's model behind a no-code storefront, features that exist only on the landing page, a markup of three times the model's price, or results you don't own on the plan you're on. Ask an agent "is this tool any good?" and it retells the landing page or guesses.

Under the Hood makes your agent check a service the way a careful buyer would: for your task, from public data and your own account, with evidence, a date and a confidence level on every conclusion.

Seven layers, in order.

• Promise and advertising. Every claim on the site, in the terms and the ads, with its status: works, interface only, stub, missing, contradicted.
• Who is behind it. The legal entity, its age, sibling products.
• What runs underneath. Builder, backend, model vendors and the tools an agent really calls, each with a level: direct (seen in requests or stated by the company), indirect (in the code), guess.
• What works on your task. Your own inputs, measured by a profile for 3D generators, aggregators or assistants.
• Money. Your month on each plan, retries and expiring credits included, and how to leave.
• Rights and data. Ownership, training, retention and sub-processors, cited by clause.
• A verdict for your task, with the questions to ask before paying.

Is it the model they sell? The skill compares what you get with what the vendor makes, strongest evidence first: the model in the request and the response, content credentials (C2PA) from the vendor, the same seed at the vendor, the fingerprint of the result (resolution, frame rate, length, codec, sound), and for language models time to first token, speed and score against both candidate models. The conclusion is "consistent with X", with a question for the service.

What does it cost them? Each plan's credit turned into dollars, against the model vendor's public price for the same generation. You see the markup on the plan you'd buy, whether paying the vendor directly is cheaper, and where a swap would pay the service most.

It asks before it spends. The check uses public data, files you already have and your own account. It stops at two points only: before a sign-up you haven't done, and before new spending, naming the generations, the credits and the amount.

Eight scripts do the reading: an in-page stack scan, a terms reader, media metadata with C2PA, a fingerprint for video, images and language models, a 3D inspector, a cost calculator, a pre-hand-over check and an HTML report.

Checking a product you own instead? Product Audit is a read-only audit of UX, accessibility, design system and security, every finding with its level of proof.