Zeno · audit
Paste a forecast.
See how it ends.
Words or numbers. Free.
Your number stays yours.
Assessment scope
- range supplied · judged
- point + range · judged
- point alone · kept, support named
Show the evidence · past misses, inside a check, review budget
Explore a field · which confident forecasts missed
01 / Explore
02 / Investigate
COVID-19
Who predicts best
14 ranked of 16 forecasters
- 1. UGA_flucast-INFLAenza+20%
- 2. CFA_Pyrenew-Pyrenew_E_COVID-4%
- 3. UMass-ar6_pooled-9%
- 4. CFA_Pyrenew-Pyrenew_HE_COVID-22%
- 5. CEPH-Rtrend_covid-10%
skill vs panel median · orange = beats the median, grey = worse
Past record · frozen 21 Oct 2026 · n 37,973 scored forecasts
evidence +−
Three skills · this field
Correct the forecast
keeps your forecast unchanged
live model's sealed exam: no measurable difference
Warn: one forecaster wrong
ahead of most rivals · +1.1%
practice test on past data · 7,983 rows · not proven
Warn: whole panel wrong
not measured yet
no validated test yet
proven only after its own judging · run receipts ↗
No market comparison, shared-miss pairs or replay cases recorded for this field yet.
03 / Live signals
Open questions
…
Next close
…
Live AI entrants
…
Next model
step 1 of 5: reading all the data · GPU training not started
reading… · SurpriseBench →
past replay · already resolved
What a check looks inside.
systems
6
distinct families
5
own ranges missed
6 of 6
still missed after correction
2 of 6
ECMWF AIFS
family ecmwf
no recordstill misses
ECMWF IFS
family ecmwf
widened ×1.9now covers
GEM · Canada
family cmc
widened ×2.1now covers
ICON · Germany
family dwd
widened ×2.0now covers
GEFS · US
family ncep
widened ×2.3now covers
UKMO · UK
family ukmo
widened ×2.3still misses
Read: 5 independent developers agreed, and all 6 ranges missed on the same side. Agreement here was one shared mistake, not five confirmations.
Thin bar: the system's own 90% range. Wide bar: the range after the correction its past record earned (none where no record exists). Teal line: observed 23.1°C. Source Offdiagonal failure-atlas-v1/weather/weather-ensemble-residuals.parquet.
review budget
How many alerts are worth it?
of 100 alerts
62 real misses · 38 false alarms
share of all confident misses caught
33%
The rest still slip through. Alerts fire above the level that flagged the riskiest 10% of training cases.
- ensemble weather only
- confirmation period · v0.57
- retrospective, not prospective