Forecast Verification & Local Calibration
Forecast verification & Glenlea local calibration
Each original forecast is archived, then verified only against sourced, quality-graded observations after the valid time passes. This stage records, measures and suggests; it never changes Scoring V2 automatically.
Model Performance by Lead Time
Cloud forecast performance
No single accuracy percentage: MAE, bias, false-clear rate, clear-hit rate and sample count are shown separately.
| Model | Lead time | Cloud MAE | Bias | False Clear | Clear Hit | Samples |
|---|---|---|---|---|---|---|
| GFSglobal_deterministic | 0_6h | 24% | -3.8% | 23% | 87% | 133 Moderate evidence |
| NBMblend | 0_6h | 25.4% | -10.2% | 24% | 84% | 137 Moderate evidence |
| ICONglobal_deterministic | 0_6h | 27.1% | 5.4% | 46% | 23% | 58 Preliminary evidence |
| HRDPScanadian_regional | 0_6h | 31.5% | -21.6% | 45% | 86% | 52 Preliminary evidence |
| Astrosphericastrospheric_primary | 0_6h | 32.5% | -26.7% | 33% | 100% | 15 Insufficient evidence |
| RDPScanadian_regional | 0_6h | 34.8% | -27.6% | 48% | 87% | 52 Preliminary evidence |
| GFSglobal_deterministic | 12_24h | 28.3% | -5.9% | 55% | 86% | 68 Preliminary evidence |
| Astrosphericastrospheric_primary | 12_24h | 31.1% | -22.9% | 46% | 85% | 42 Preliminary evidence |
| NBMblend | 12_24h | 31.3% | -20.2% | 68% | 100% | 102 Moderate evidence |
| ICONglobal_deterministic | 12_24h | 38.1% | -3.8% | 100% | 0% | 29 Insufficient evidence |
| HRDPScanadian_regional | 12_24h | 41.3% | -28.1% | 67% | 33% | 29 Insufficient evidence |
| RDPScanadian_regional | 12_24h | 44.1% | -39.8% | 67% | 100% | 33 Preliminary evidence |
| ICONglobal_deterministic | 24_48h | 10.7% | 10.7% | — | — | 7 Insufficient evidence |
| NBMblend | 24_48h | 20.6% | -6.9% | — | — | 15 Insufficient evidence |
| GFSglobal_deterministic | 24_48h | 22% | -3.2% | — | — | 8 Insufficient evidence |
| RDPScanadian_regional | 24_48h | 23% | -9.2% | — | — | 10 Insufficient evidence |
| HRDPScanadian_regional | 24_48h | 26% | -7.2% | 100% | — | 8 Insufficient evidence |
| Astrosphericastrospheric_primary | 24_48h | 38.3% | -32% | 80% | — | 27 Insufficient evidence |
| GFSglobal_deterministic | 6_12h | 27.2% | -13.2% | 42% | 96% | 79 Preliminary evidence |
| NBMblend | 6_12h | 29.1% | -16.4% | 42% | 96% | 92 Preliminary evidence |
| ICONglobal_deterministic | 6_12h | 29.5% | 11.3% | 60% | 13% | 34 Preliminary evidence |
| RDPScanadian_regional | 6_12h | 31.9% | -24.6% | 46% | 91% | 36 Preliminary evidence |
| Astrosphericastrospheric_primary | 6_12h | 32.1% | -22.6% | 41% | 80% | 33 Preliminary evidence |
| HRDPScanadian_regional | 6_12h | 36.2% | -16.5% | 52% | 54% | 39 Preliminary evidence |
Forecast vs Actual
Past-night trajectory
2026-09-02 · Common timeline; observations always retain source and quality.
AstrosphericHRDPSRDPSGFSICONNBMActual
Confidence Calibration
Is the stated confidence reliable?
| System confidence bin | Mean stated value | Observed reliability | Samples | Evidence |
|---|---|---|---|---|
| 20-30% | 25% | 45% | 31 | Preliminary evidence |
| 30-40% | 38% | 0% | 17 | Insufficient evidence |
| 40-50% | 45% | 23% | 69 | Preliminary evidence |
| 50-60% | 55% | 21% | 63 | Preliminary evidence |
| 60-70% | 65% | 38% | 79 | Preliminary evidence |
| 70-80% | 75% | 35% | 99 | Preliminary evidence |
| 80-90% | 85% | 46% | 156 | Moderate evidence |
| 90-100% | 95% | 62% | 254 | Moderate evidence |
Decision Verification
Trip-decision verification
Decision outcomes currently require an explicit user field report; automatic station data does not silently decide whether an imaging trip succeeded.
Local Calibration · Shadow Mode
Local weight suggestions
Automated evaluation continues, but exact weights, anomalous samples and experimental details appear only after Admin unlock. Production V2 cannot change automatically.
There are currently no web controls that can change Production. Automated evaluation continues in Shadow Mode.
Ground Truth Policy
Observed-condition quality levels
Forecast ≠ Observation. No verification result is produced without reliable observed conditions.
Major Forecast Misses
Major forecast misses
Major imaging-relevant misses are tracked, such as Cloud ≤10% when observed cloud is ≥70%.
HRDPS · 2026-09-03 22:00 CDT
Forecast 0% · Observed 100% · Quality CHRDPS · 2026-09-03 22:00 CDT
Forecast 0% · Observed 100% · Quality CHRDPS · 2026-09-03 22:00 CDT
Forecast 0% · Observed 100% · Quality CHRDPS · 2026-09-03 22:00 CDT
Forecast 0% · Observed 100% · Quality CHRDPS · 2026-09-03 22:00 CDT
Forecast 0% · Observed 100% · Quality CHRDPS · 2026-09-03 22:00 CDT
Forecast 0% · Observed 100% · Quality CHRDPS · 2026-09-03 22:00 CDT
Forecast 0% · Observed 100% · Quality CHRDPS · 2026-09-03 22:00 CDT
Forecast 0% · Observed 100% · Quality CActual Conditions Archive
Recent observed conditions
The public page shows automatic observations and anonymized verification information only. Complete field-report details are available to the administrator.
OvercastQuality C
ECCC SWOB cloud · Winnipeg/James Armstrong Richard · 2026-09-03 22:00 CDT
ECCC SWOB reported total cloud amount percent.Weather observationQuality C
ECCC SWOB · KLEEFELD (MAFRI) · 2026-09-03 22:00 CDT
Nearby official in-situ station; not located at Glenlea.OvercastQuality C
ECCC SWOB cloud · Winnipeg/James Armstrong Richard · 2026-09-03 21:21 CDT
Estimated total cover from the maximum ECCC SWOB total_cloud_amount layer code; the documented tenths/coverage range midpoint is used. Non-quantitative codes are ignored. This is not a direct Glenlea cloud sensor measurement.OvercastQuality C
ECCC SWOB cloud · Winnipeg/James Armstrong Richard · 2026-09-03 21:00 CDT
Estimated total cover from the maximum ECCC SWOB total_cloud_amount layer code; the documented tenths/coverage range midpoint is used. Non-quantitative codes are ignored. This is not a direct Glenlea cloud sensor measurement.Weather observationQuality C
ECCC SWOB · KLEEFELD (MAFRI) · 2026-09-03 21:00 CDT
Nearby official in-situ station; not located at Glenlea.Variable CloudQuality C
ECCC SWOB cloud · Winnipeg/James Armstrong Richard · 2026-09-03 20:00 CDT
Estimated total cover from the maximum ECCC SWOB total_cloud_amount layer code; the documented tenths/coverage range midpoint is used. Non-quantitative codes are ignored. This is not a direct Glenlea cloud sensor measurement.Weather observationQuality C
ECCC SWOB · KLEEFELD (MAFRI) · 2026-09-03 20:00 CDT
Nearby official in-situ station; not located at Glenlea.OvercastQuality C
ECCC SWOB cloud · Winnipeg/James Armstrong Richard · 2026-09-03 19:00 CDT
ECCC SWOB reported total cloud amount percent.Weather observationQuality C
ECCC SWOB · KLEEFELD (MAFRI) · 2026-09-03 19:00 CDT
Nearby official in-situ station; not located at Glenlea.Weather observationQuality C
ECCC SWOB · KLEEFELD (MAFRI) · 2026-09-03 18:00 CDT
Nearby official in-situ station; not located at Glenlea.OvercastQuality C
ECCC SWOB cloud · Winnipeg/James Armstrong Richard · 2026-09-03 18:00 CDT
Estimated total cover from the maximum ECCC SWOB total_cloud_amount layer code; the documented tenths/coverage range midpoint is used. Non-quantitative codes are ignored. This is not a direct Glenlea cloud sensor measurement.OvercastQuality C
ECCC SWOB cloud · Winnipeg/James Armstrong Richard · 2026-09-03 17:00 CDT
Estimated total cover from the maximum ECCC SWOB total_cloud_amount layer code; the documented tenths/coverage range midpoint is used. Non-quantitative codes are ignored. This is not a direct Glenlea cloud sensor measurement.Read-only Data Access
JSON / AI / MCP
Update log
What changed in each published version
v0.2.3 · Responsive layout consistency
- Stopped English and French metric names from breaking inside words; compact phone labels now keep every hourly row the same height.
- Extended priority gradients and row dividers to the true section edge, removing the blank clipped strip on iPhone layouts.
- Added a dedicated tablet layout, restored the sky timeline on larger screens and enlarged interactive controls to comfortable touch targets.
- Localized remaining mixed-language labels and retested the briefing, drawers and hourly dialog across portrait phones, tablets and desktops.
v0.2.2 · Portrait forecast readability
- Restored the phone briefing to a left-hand verdict and right-hand score ring, with proportions tuned for narrow portrait screens.
- Compressed the pinned metric column and hourly columns so more consecutive hours can be compared at once.
- Added full-height midnight separators, date labels, a current-time marker and a compact touch guide.
- Checked English, French and Chinese at multiple portrait phone sizes while retaining the past 24 hours and full future model coverage.
v0.2.1 · Mobile layout & diagnostics
- Reflowed the score ring on iPhone-sized screens so long English and French conclusions keep the full text width.
- Completed dynamic English/French interface copy and added wider device/layout checks.
- Added a privacy-safe diagnostic download and version-aware rendered-page cache.
v0.2.0 · Public Beta briefing
- Introduced the Night Mission Briefing, hourly conditions strip and future-night comparison.
- Added forecast verification in shadow mode without changing production scoring.
Website feedback
Found a translation, layout, loading or forecast-page issue? Download the privacy-safe diagnostic log and attach it to your email. It contains device, viewport, version and timing data only—never cookies, passwords, API keys or server logs.
Public Beta · Winnipeg Time · America/Winnipeg