Hillclimb

The maker never grades its own work. Everything else is bookkeeping that keeps the grades honest.

0. Set up (once, before any change)

Write hillclimb/<slug>/PLAN.md with:

Target: the artifact and the one sentence it should live up to.
Rubric: 3–7 rows. Each row gets anchors for 6, 8 and 10, citing real references ("10 = the Nintendo Switch eShop page", "8 = a senior reviewer would merge it unchanged"). Frozen for the run.
Evidence command: one command that produces what evaluators see (screenshots, a rendered page, sample outputs on a fixed input set, a recording + av-verdict). Same seed and inputs every round. Evaluators see evidence, never code or the author's reasoning.
Gate: what must stay green regardless (tests, typecheck, verify skill, budget).
Stop rule: a target (every row ≥ 8), a floor (≥ 3 rounds), a cap (rounds or hours), and the spend cap if anything paid runs.