Sealed 2026-10-26T00:15:07Z. The visual form of the MANDATE 4.7 checks.
10 of 10 bins occupied; empty bins are absent, not interpolated. Marker size is the number of observations in the bin.
| Brier (multiclass, 0 to 2) | 0.63636 |
| base-rate reference | 0.46455 |
| improvement over reference | -0.17181 |
| persistence reference | 0.97633 |
| improvement over persistence | 0.33997 |
| Brier (per state, 0 to 1) | 0.31818 |
| base-rate reference | 0.23227 |
| improvement over reference | -0.08590 |
| persistence reference | 0.48817 |
| improvement over persistence | 0.16999 |
Scored against the outcome rule forward_window_state over a 21-bar forward window — this detector's own declared exam, overriding the campaign default (next_bar_state). Bars scored: 1352 of 1394 emitted. The 42 unscored bars are the forward window's cost: the earliest have no closed reference window and the last 21 have no completed forward window, so their outcome has not happened yet.
While every detector emits a degenerate distribution, the per-bar multiclass Brier is 0 if the state persists to the next bar and 2 if it flips — so brier_multiclass is exactly twice the state-change rate, and the skill over the base rate is just the state's autocorrelation restated. A detector that never changed state at all would score near-perfectly here while saying nothing whatever. That is not a leak (no future value reaches any emission; the emissions are bit-identical to the pre-E.4 ones) and it is not a reason to withhold the number, but it does mean this metric only begins to measure information content once a detector puts real mass on more than one state. The machinery is being proven now precisely so that it is already trusted when that happens.
Consecutive-bar run lengths per state, mean marked. The same occupancy can be one long stretch or a hundred single-bar twitches.