A flashy benchmark score without a simple baseline is theater. Run a dumb-but-honest baseline on the same split — keyword rules, prior model, human template — then report the delta. Buyers who demand the floor number waste less money on ceiling fireworks.

Simple baseline, same split, delta, cost/latency — Eval hygiene: improvement is a subtraction problem.

Make the floor visible

Define the baseline in writing before tuning the fancy system.
Keep data splits identical when you compare.
Report cost and latency beside accuracy — free accuracy is rare.

Public habit: refuse a leaderboard screenshot that never names the baseline.

Sources: Public ML evaluation methodology notes; buyer RFP eval checklists. Soft-lander essay — no product claim.