We're building the AI stack we'd actually trust with our own business.

Every headline number about an AI model comes from the company selling it, measured on a benchmark chosen by the company selling it. That is marketing with a decimal point in it. You cannot plan a budget or pick a model from it.

We run our own stress tests nightly against live endpoints at every major provider — deterministic validators, code executed in a sandbox, JSON parsed against a strict schema. No LLM grades another LLM. Every result is dated, reproducible, and free to read without an account.

The failure everyone actually fears from AI is not that it says "I don’t know" — it is that it invents something plausible and says it firmly. An assistant that is right 95% of the time and undetectably wrong the rest is not 95% useful; it is unusable for anything that matters, because you have to check all of it.

So Groundyard answers from your own sources or it declines. A lookup cannot hallucinate: where a real answer exists in your data we fetch it rather than ask a model to recall it. The abstain gate is tuned deliberately tight — over-abstaining costs a visitor a page they could have been handed, under-abstaining invents an answer, and those two losses are not the same size. On our own traffic so far it has declined 42 of 145 grounded questions rather than guess. That is a small sample and we would rather show it than round it.

Not done yet: The rate is published here and nowhere else yet; it belongs on /measured as a standing receipt with its own method note, once there is enough traffic behind it to mean something.

The industry keeps promising autonomy, then quietly discovering that the interesting decisions are the ones a human wanted to make. An assistant that re-tiers itself, changes its own model, or edits its own guardrails is not saving you work — it is moving the work to the moment something goes wrong.

Nothing changes how your assistant runs without a person approving it. Every answer can be marked wrong by the person reading it, with a reason, and that verdict is attached to the turn. Spend caps are checked before the model is called, so the failure mode is a refusal rather than a surprise invoice.

Not done yet: The correction loop is live but early — few enough verdicts so far that we would not dress it up as a dataset. It gets more useful the more it is used, and we would rather say that than imply scale we do not have.

Most deployments pick one premium model on day one and never revisit it, because nobody can prove a cheaper one is safe. So you pay top price for every "what are your hours" for the life of the contract.

Switchyard replays your assistant’s real questions against cheaper candidates, judges how well each resolved them, and proposes a better-value route with the evidence attached. A quality floor is enforced in code: a cheaper model that gives up more than a set margin of judged answer quality is refused, so "cheaper" can never quietly become "worse". An operator approves each proposal today.

Not done yet: Tenant-controlled auto-routing: you set the bounds, and routing moves inside them without waiting on us. The measurement, the judging and the quality floor already exist — what is missing is the control surface and the policy, not the engine. It ships when the bounds are something you can set and see, because an automatic change you cannot inspect is exactly the thing the belief above objects to.