schedopt_calib

The run returned a usable operating point, but not the one the hypothesis named. Reaching 95 unseen_mismatch within 250 steps is something almost nothing in the smooth three-parameter family under study does: six of the seven in-family probes crossed on none of their 30 seeds, the seventh on 0.10, while the off-family burst schedule carried over from the earlier work crossed on 0.93. An objective that reads zero everywhere the search is about to look offers the optimizer nothing to climb, which is the failure mode this phase exists to catch. Lowering the threshold to 92.5 and extending the budget to 400 recovers a graded response across the same probes, and the campaign is frozen there.

The pre-registration left open whether the crossing test needs a dwell guard, a requirement that the threshold be met at two consecutive waypoints rather than one. It does, and the guard turns out to do more than suppress noise. Requiring the level to hold for 25 steps removes every one of down_floor's apparent crossings and roughly a third of lin_up's and fixed_16's, while leaving the incumbent untouched. Schedules therefore differ not only in whether they reach the band but in whether they stay there once they do, and a test without the guard would score a graze and a durable crossing alike. Smoothing was rejected on the opposite evidence: a three-waypoint average shifts crossing times and shaves a few points off most arms without separating any pair that the raw curve did not already separate.

The probe ordering itself says something about what the search is likely to find. The two arms that never cross are the two that hold α above the productive region for most of the run: down_floor descends from 5 only as far as 1.3 and reaches that floor at the final waypoint, and late_fast stays near 5 until late. Neither is incapable of the band so much as out of time, since their best unseen_mismatch averages 89.7 and 84.7 across seeds and every late_fast seed and 0.77 of down_floor's are still rising at 400 steps, which reads as undertraining rather than a ceiling. What separates the arms is how much training happens once the schedule has brought α into the region where the rule is learnable, and both a descent through that region and an ascent into it beat holding above it.

This does not overturn the finding that a schedule's endpoint governs its fate, because the two measurements answer different questions. Mirror_v1 scored accuracy at the end of a 1200-step run, where a long uniform tail has time to undo the rule through erosion. Here the objective is only whether a run has crossed the band by step 400, and lin_down spends just its final 50 steps below α 0.5, too brief for the tail to do damage before the budget is up. The objective is the probability of crossing by 400, not final accuracy at a longer horizon, so it will favor schedules whose endpoints would be poor choices under a longer run.

Two limits on how far this generalizes are worth stating. The calibration seeds are disjoint from every later phase's pool, so the frozen threshold carries no advantage into the search, but it was chosen to make these eight shapes separate and a different family might have wanted a different band. And no seed in the pool is hopeless: each crossed between two and five of the eight arms, so the achievable probability is not capped by a handful of broken initializations, which seed bimodality in the fixed-α arms might otherwise have suggested.