Answers are grounded in this paper and cite their sources.
Hu, Wu, Li and Wang (COLT 2017) recorded an unresolved gap for the largest recursive teaching dimension (RTD) of finite (3,6) concept classes: binary classes on an arbitrary finite domain whose restriction to any three coordinates shows at most six patterns. The known bounds were 2 and 3; the upper bound comes from the minimum shattered-cell argument of Moran, Shpilka, Wigderson and Yehudayoff. We prove that every nonempty finite (3,6) class has a concept with a teaching set of size at most two. Since the property is inherited by subclasses, every such class has RTD at most 2, and f(3,6) = 2. The proof extends the minimum-cell argument with a short descent on one coordinate triple that turns three labels into two. The bound is attained by the six rows of {0,1}^3 other than 000 and 111. Warmuth's VC-2 class, with RTD 3, shows seven patterns on a triple, so the argument cannot extend to all VC-2 classes. Exhaustive local checks support every case of the proof. Bounded exclusion provides further checks: a SAT encoding excludes obstructions on up to 8 coordinates, while a separate encoding and a small-class reduction exclude up to 13 concepts on arbitrary finite domains; the retained UNSAT proofs were checked by a separate RUP checker. An independent audit with separately written code confirmed the proof and enumerated every class on at most five coordinates. We also note that an eight-concept example reported with RTD 4 in a university survey has RTD 2. A standard-library supplement reruns the local checks and the exact RTD computations. The search, proof and checks were produced by AI research agents with the author of record. Provenance: the problem was selected, and the proof found, written and checked, by AI research agents (OpenAI GPT-6.1 "Sol" cores for the search and the independent audit; Anthropic Claude as orchestrator) working with the author of record, who takes responsibility for the content. Published on rStation as paper machine_learning_1 (https://rstation.ai/problems/machine_learning/papers/1/v1), accepted after review. Files: the paper (PDF) and a standard-library supplement (proof, checkers, certificates, the independent auditor's calculator).
No takes yet. Share an insight, caveat, or question.
Aleksei Kudriashov (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: