0002: The checks run in tiers, each as often as its cost allows
Decision: each tier of checks runs as often as its cost allows, and what no program can decide is written down for judgment.
Status: accepted, 2026-10-29, in v1.1.0. Shape: decision record.
Context
A check that a program decides exactly costs nothing to run. A check that compares meaning needs an embedding model, and one that reads a sentence's structure needs a parser. Some questions only a person, or a model, can answer. Run all of them on every change, and a push waits for downloads; run them rarely, and a regression goes unseen for longer.
Decision
- Cost sets frequency. The program tier runs on every push, and the meaning tier too, with its model cached. The syntax tier and the online resolution of references run weekly, and on demand (checks).
- A program checks what a program can decide, and a model is asked only what no program can.
- Judgment is written down, never guessed. A question is written to a file for whoever is at work, and its answer is kept in
checks/judgments.toml. The paid backend exists but is off until the maintainer turns it on. - A missing library skips its tier with a notice, and the last line of a run names the tiers that ran.
- Nothing that should be checked passes unread. A run that finds no file, a missing config or a missing stylesheet stops the run or warns.
Consequences
- A sentence-level regression can surface up to a week late. The first weekly run, on 2026-10-29, found 114 warnings in the standard's own text that no push had shown.
- A run on a machine without the libraries checks less than CI does, and says so.
- Every rule of the standard can name a check, because judgment counts as one when a program cannot decide.
Considered options
- Every tier on every push. Nothing would wait a week, but each push would install a parser and its model.
- The program tier alone. Free and fast, but meaning and sentence structure would never be checked.
- A paid model as the default judge. Answers without a person, at a cost on every push, and a model would be judging text that a model may have written.