Zeno Divergent

The second-order layer. Not "these models are beginning to disagree", but: these models are currently healthy, and their recent error dynamics indicate an elevated probability that their predictive validity breaks down within the horizon they advertise.

Observed fragility is DERIVED from measured upstream signals. Prospective breach probability and validity horizon are SIMULATED — the platform publishes no public second-order scoring feed, so these are a deterministic estimate anchored on the measured fragility, never presented as retrieved fact.

models scored

33

observed degrading

7

breach probability > 50%

7

simulated

median horizon deficit

70h

advertised minus estimated validity

Forecast horizon versus validity horizon

The signature view: what a model advertises, against how long we estimate it stays valid

SIMULATED
0h24h48h72h120h168h240hAir GAT candidate~2h / 72hCAMS bias adjusted~4h / 72hAirWatch calibrated NO2 head~5h / 72hAir persistence reference~6h / 72hAirWatch calibrated O3 head~7h / 72hAirWatch calibrated PM2.5 head~10h / 72hAirWatch calibrated~14h / 72hCRW SST persistence~54h / 120hYield anomaly~109h / 240hOzone damage~115h / 240hAnalytical crop~117h / 240hUrban heat island counterfact…~83h / 168hSeismic anomaly nowcast~83h / 168hClimate projection~120h / 240hSeasonal ensemble~132h / 240hSeaWatch SST LightGBM~68h / 120hUrban heat incumbent~98h / 168hSST anomaly nowcast~65h / 120hTrained yield~132h / 240hAnalytical pest~132h / 240hAnalytical disease~135h / 240hCompound risk event~96h / 168h

Grey bar: horizon the model advertises. Coloured bar: estimated validity horizon. The gap between the tick and the end of the grey bar is the interval in which a forecast is still being issued but should no longer be trusted at face value.

Fragility ledger

Ordered by estimated breach probability

DERIVED

Columns: observed fragility (derived) · breach probability within horizon (simulated).

Air GAT candidate

AIR · air-gat-candidate

SIMULATED

observed fragility

85%

derived from measured signals

breach probability

95%

within 72h · simulated

advertised horizon

72h

estimated validity

~2h

simulated

prospective fragility, next 24h

Dashed line is the breach threshold at 0.60.

why

  • interval coverage 0% vs declared 90%
  • persistent over-prediction bias 18.4
  • 5 upstream drift alerts open
  • declared status "experimental"
  • not claim-eligible — outputs cannot support a verified claim

Scored against upstream accuracy generated 51m ago. Fragility score 0.851.