Public health · research

When will the hub ensemble be confidently wrong?

Tested on outcomes the model never saw.

24 fields · 7 areas

1 · Who forecasts it

16

forecasters · 38,610 forecasts · 6,365 with outcomes

  1. 2025-03-08 → 2026-05-09 · 90% interval not measured

  2. 2025-06-28 → 2026-05-09 · 90% interval not measured

  3. 2024-11-23 → 2026-02-07 · 90% interval not measured

past record · ranked vs panel median · as of 21 Oct 2026

2 · Which confident forecasts missed

ahead of most rivals

Confident-miss warnings vs each forecaster's own track record (tree) · +1.1% on confirmation · interval not measured

Test n 7,983 · groups 2,474

practice test · not the served model

3 · Should you trust a forecast?

Supported correction or labelled reference.

Evidence · COVID-19
Inventory source
frozen v0.80.2 build census + v0.82 practice run. Counts describe the frozen corpus, not independent samples.
Warning test period
historical: development and later confirmation slices of past data. Calendar boundaries not supplied in this summary.
Comparator and method
each forecaster's own track record (tree); Brier score of confident-miss warnings. CPU tree rehearsal, not the served neural model.
90% interval
Not measured in this export. No interval inferred from row counts or significance values.
Confirmation sample
7,983 forecast rows · independent groups: 2,474. Corpus inventory is not the test sample.
Model identity
Served model: v1.11. Research model: not yet promoted.
Dependence evidence
10 documented links across the corpus, not COVID-19-specific same-side misses. Missing links mean unknown.
Scoring boundary
Past record and practice results are not sealed live performance. Check receipts identify the answer actually served.

What runs Zeno

Member of NVIDIA Inception · H200 research training

Job records, not a total of successful runs. GPU hours not verified here. Inception membership is not an endorsement.

Evidence, not affiliation

Forecast seals carry immutable receipts. A seal proves submission timing, not forecasting skill.

On-chain compute receipts: not verified. Celo token presence does not prove model execution or forecast accuracy.

+17%

less error than the crowd · unseen data

  • COVID-19n 5,726 weeks

    +17% · 95% range 9–23%

Decision
Act on a hub forecast or wait a week
What Zeno looks for
Hub forecasts that miss together when reporting lags
Field check
Planned
You send
Hub forecasts and reported counts
Pilot measures
Interval coverage per independent week
Who it is for
Public health agencies, forecasting hubs