Forecast quality is a separate question

Frozen SKAB validation: five runs, 5,100 eligible observed timestamps, eight targets. This post-hoc descriptive audit uses genuine cached Chronos-2-small forecasts and matched block persistence. Models, scores, calibration and cutoffs are unchanged.

Chronos median normalized MAE: 0.596757; persistence: 0.798079. The macro error reduction is 25.23%. Chronos has lower MAE in 39 of 40 run×channel cells. These cells share runs and are not 40 independent deployments.

Same forecasts, separate regimes

Exactly what is averaged

Within each run×channel cell: mean over eligible observed timestamps; then unweighted mean over all 5×8=40 cells. The same hierarchy is used within each regime or horizon. Relative reduction = 1 − macro Chronos nMAE / macro persistence nMAE.

Per channel fixed healthy-fit 1.4826*MAD; if <=1e-10 use healthy-fit population SD; if <=1e-10 use 1. No validation refit.

All scales are fitted on the 2,236 healthy training observations from the original classical pipeline. No validation values fit these scales.

Per-channel errors in physical channel units

Horizon after the shared issuance time

Intervals and pinball loss

Chronos [q10,q90] nominal80% interval; report empirical coverage together with mean width/fixed channel scale. Persistence has no interval in this experiment.

Unweighted mean of normalized pinball loss at the three cached levels q=.1,.5,.9; not all13 model quantiles. Overall macro normalized pinball: 0.194682.

What these results support

The median forecast improves over matched block persistence in this validation protocol, including inside annotated abnormal states. This does not turn residual surprise into a full-state detector. Average forecast error and max-channel anomaly discrimination optimize different objectives. The audit does not establish a causal mechanism for the residual-detector failure.

• Five validation runs from one experimental rig; no uncertainty estimate of independent deployment performance.
• Forecast quality is measured separately from anomaly discrimination and does not establish why a residual detector fails.
• Classical detectors refresh every observed timestamp; forecast residuals use a fixed sixteen-step issuance schedule.
• Post-hoc descriptive forecast analysis does not modify frozen model, feature, calibration or threshold choices.
• The cached median falls outside q10/q90 in 21 of 40,800 validation values, all in the flow channel; raw quantiles are retained, not sorted.

Fixed run valve1/8

The mean ensemble issues no alarms at its frozen cutoff. Chronos residual issues ten alarms: five true and five false positives. The run has 400 annotated abnormal eligible timestamps. Scores are never smoothed before thresholding, and no interval is filled using labels.

Audit files

Summary, scales and input hashes · Per-run/channel numerators and denominators · By regime · By horizon · Validation-only cached forecasts · Actual raw-score decisions