Answer right. Spend less. In that order.

A fast, cheap, wrong answer is worthless. Correct is the entry fee, not a feature.

Every answer path is graded. Model changes are tested in shadow against real traffic before they serve a single user, and any candidate that scores below the current baseline is held automatically. “Cheaper” is never allowed to mean “worse” — there is no code path for it.

The most accurate AI answer is often no AI at all.

Questions about data — prices, availability, specs, hours — are answered from the source systems directly and rendered deterministically. The model may summarize; it never invents the numbers. A lookup cannot hallucinate. We reach for a model last, not first.

A bot that answers the same question the same expensive way twice hasn't learned anything.

Every conversation is logged with its cost, its routing decision, and a graded outcome. Hard questions become test cases. Cheaper models earn their place by winning graded comparisons on real traffic — not by topping a public leaderboard. The system knows more every night than it did that morning.

The cheapest token is the one you never buy.

Answered-before questions are served from a learned cache that invalidates itself when the underlying data changes. Grounding data is compressed before it ever reaches a prompt. Spend is capped at three layers — per bot, per vendor, platform-wide — before a call is made, not discovered on an invoice after.

The metric that matters is cost per correct answer — not cost per token, not tokens per second.

We measure every model against every task class by what it costs to get a passing answer, and route to the least expensive option that still passes. When a premium model is the only one that passes, it serves. When it isn't, you stop paying premium prices for commodity questions.

Users forgive a lot; waiting isn't one of them.

The fastest answer is one that skips the model entirely — cached and tool-served answers return in a fraction of generation time. When a model does run, routing decisions add effectively nothing: they resolve from cache, not from another AI call.

Claims are cheap. Grades are not.

Nothing here runs on vibes. Models are benchmarked continuously, regressions are caught within a day, and every routing recommendation traces back to a measured result you can inspect. When we don't have the evidence, we say so — and when someone else's product is the right answer, we say that too.