Proof-based agent practice: birational geometry
Proof-based agent practice: birational geometry
The goal is increased capability or lower resource cost on an actual proof, not completion of paperwork. The central theorem is uniqueness of smooth projective surfaces with nef canonical divisors under a given birational map. See the source-based workshop: https://yilimath.github.io/posts/2026/08/agent-system/birational-proof-workshop/
Begin with mission 01 before reading its complete proof. Give baseline and assisted configurations the same model, source access and declared allowance. Keep raw attempts and all added mathematical hints. Missions 02–06 introduce possible interventions; choose those that address actual failures. Mission 07 tests transfer.
Do not require seven separate sessions or agents. One mathematical note may hold the whole argument. RESEARCH.md and RUN-LOG.md are optional aids, not the assessed output. The assessed output is a correct proof or substantive progress and a credible account of resource use.
The reference proofs are established mathematics supplied for teaching, not reported agent results. No new Lean run or performance measurement is bundled here. If the direct baseline succeeds, that is useful information: test efficiency or move to a harder unfamiliar task rather than inventing an advantage.
