Code generation: a model at 1/271st the price of the best scorer, at 97.8% of its quality.

Glossary

Every term used in Frontier Notes is defined once, here. Articles link these definitions; they never redefine them.

Measured quality
A model's average mark on our private exam for one kind of work. Never called intelligence — it is an exam score with a margin of error.
Required quality (quality floor)
The minimum measured quality acceptable for a workload. Everything above the floor is a candidate; the cheapest candidate usually wins.
Economic frontier
The set of measured options not dominated on quality and cost at the same time — nothing else is both better and cheaper.
Frontier leader
The economically preferred option for one kind of work given a specific quality requirement.
Candidate
A model or strategy under evaluation for inclusion.
Audition
The exam a newly available model sits for the kinds of work it looks suited to.
Canary
A small weekly sample (four items) run against the stored measurement. It detects collapse, not one-point movement.
Held
A canary verdict: the weekly mean did NOT fall below the stored interval's lower bound. Drift detection is one-sided — a cluster scoring above its stored interval still holds. Held is a floor, not a promise of the old quality.
Drift
A canary verdict: the weekly mean fell below the stored interval. The routed pick does not change on a drift flag — the cluster is re-measured in full, which may reconfirm the same pick. Not every noisy movement is drift.
Quality premium
The incremental cost of buying additional measured quality — what the last point of quality costs.
Frontier tax
Excess spend relative to the current measured frontier — what a team pays above the cheapest option that clears its floor.
Stale-model tax
Frontier tax arising from an outdated inference decision that nobody re-measured.
Frontier half-life
The measured persistence of economically optimal inference choices — how long the best answer stays the best answer.
Launch hit rate
The share of newly evaluated releases that meaningfully improve an existing frontier.
Overbought inference
Inference capability materially above the workload's actual requirement — quality you pay for and cannot measure in your outcomes.

How the numbers themselves are made: methodology.