The v2 tuning, held-out, and stress phases each generate one deterministic 160-item balanced schedule, then reuse it across every cell and replicate in that phase. This gives paired, reproducible model comparisons, but the reported gate metrics are conditional on one schedule per phase. The expanded v2 article now discloses this limitation.\n\nBefore promoting a future continuous-frequency scorer:\n\n- run supported and stress gates across multiple independently seeded balanced schedules;\n- separate within-schedule replicate variation from between-schedule variation;\n- report how aggregate and worst-cell coverage, MAE ratios, and quiz length vary by schedule;\n- version the schedule seeds/count and numerical shortcut in result artifacts;\n- preserve an untouched held-out phase and do not retune from its results;\n- update the executable article and parity/runner tests with the resulting decision.\n\nThis is a validation follow-up, not a reason to reopen or retune continuous-pair-frequency-logistic-v2.
The v2 tuning, held-out, and stress phases each generate one deterministic 160-item balanced schedule, then reuse it across every cell and replicate in that phase. This gives paired, reproducible model comparisons, but the reported gate metrics are conditional on one schedule per phase. The expanded v2 article now discloses this limitation.\n\nBefore promoting a future continuous-frequency scorer:\n\n- run supported and stress gates across multiple independently seeded balanced schedules;\n- separate within-schedule replicate variation from between-schedule variation;\n- report how aggregate and worst-cell coverage, MAE ratios, and quiz length vary by schedule;\n- version the schedule seeds/count and numerical shortcut in result artifacts;\n- preserve an untouched held-out phase and do not retune from its results;\n- update the executable article and parity/runner tests with the resulting decision.\n\nThis is a validation follow-up, not a reason to reopen or retune continuous-pair-frequency-logistic-v2.