A World of AI product · planetAI LLP ← The atlas·GREEN‑LLM·HARMONIA

Assay

LLM claim validation · a World of AI product by planetAI LLP

Every model launch arrives with a number. Assay checks whether that number survives independent measurement.

Vendor says 40% cheaper, 30% faster, half the tokens. Assay sets each claim against third-party benchmarks and your own test tasks, then strikes a hallmark: the gain is real, partly real, or marketing.

Claims assayed to date–
Loading the ledger…

The public ledger

Each release Assay has examined, claim by claim. The dashed line is what the vendor promised; the bar is what independent measurement delivered. Tap a hallmark above or a filter below to narrow the list.

How a hallmark is struck

Assay divides the measured gain by the claimed gain. Deliver at least 80% of the promise and the claim is hallmarked; 40 to 80% is plated; under 40%, or a move in the wrong direction, is fool's gold. A gain that costs measured quality is capped at plated. Try it.

The release watch

Every Monday at 09:00 IST a GitHub Actions job lists new models from OpenAI, Anthropic and Google, reads their release notes, runs the test packs within a call budget and publishes a fresh audit report.

Last watch run

–

–

Models under watch

–

    Take it further

    Bring your own API keys to the lab and run the live packs on your own tasks. Keys stay in your browser.