Each sentence goes to every model with lockin's own logging instructions. Answers are graded by fixed checks: the right number of entries, exact kg and reps, the stated time, food names, sleep hours, and protein and kcal inside a sane range. Reference answers score 100% and empty answers 0%, which is checked before every run.
Each question goes to every model with lockin's chat instructions and the same sample log: two weeks of food, gym, sleep and weight. Answers are graded by fixed checks: the numbers and words a right answer must have (35 kg for the last bench, legs for what to train, a doctor for knee pain), no meat for a vegetarian, under 150 words, and plain text.