Post-trained a 4B Qwen distill to emit pg_hint_plan hints that beat Postgres's default planner on JOB: 44.7% latency reduction across 113 join-heavy queries, starting from a model that couldn't produce a plan for 99 of them. The pipeline is practical, not just a benchmark flex: CEB (~13.6k queries, 12 topologies) for training, JOB for validation, with topology overlap checked and found to be zero. Hints steer Postgres without touching its cost model, so this deploys on stock Postgres. The engineering details practitioners will actually steal are the measurement ones. Cache-warmness via shared hit/read block counters within 2%, interleaved candidate/default pairs, a 5% tie zone, and — the part everyone gets wrong — minimizing "fooling rate" rather than coefficient of variation, because job-13b's bimodal timing distribution hands a no-op candidate a phantom speedup roughly 40% of the time. Split across a rented 2x H100 node and four Postgres containers on a desk. Distillation from GPT-6 Astra trajectories, then a custom GRPO variant for the noisy reward. Notable that the honest framing is workload-amortized: don't use this for one-off queries, use it for queries run thousands of times. Larry Ellison is presumably drafting the license for the join tree as we speak.