zeno-divergent-v1

Zeno Divergent

A model of models: it learns which forecast to trust, and when.

Every important decision already has forecasts: tomorrow's wind, next month's inflation, an asset's volatility, the odds on a market question. What's missing is a way to know which of them deserves weight right now. Zeno is built for that. It reads dated forecasts, what was knowable at the time and what actually happened, and learns when to trust a forecast, when a small correction helps, and when to leave the best answer alone.

Download

from huggingface_hub import snapshot_download
path = snapshot_download("mbarbosa1/zeno-divergent-v1", revision="v0.28-preview")
# then run load_example.py inside `path`

What v0.28 shows

• The anchored design is the stable one. Anchored models scored a dev loss of about 6.64 on every seed. The unanchored version collapsed on 2 of 3 seeds (8.73), which is direct evidence that starting from a strong forecast is the right foundation.
• Fully reproducible. Every checkpoint ships with a replay fixture and expected hashes, so anyone can confirm they get the same predictions.
• What it doesn't show yet: v0.28 was not scored against a crowd or market reference (0 matched cases; 34 of 754 dev cases were evidence-eligible). 6.638 is the average across seeds. It is a development-selected research candidate, not a proven forecaster.

Where this is going

• More of the world: v0.29 adds weather, earthquakes, wildfires and economic data alongside markets, so the model learns trust across very different domains.
• A public track record: a live scoreboard at zenodivergent.dev logs a forecast every day, in public and append-only, and scores it against standard baselines.
• Open competition: entering FutureEval and other forecasting benchmarks, so results are scored by independent organisers rather than by us.
• Useful where mistakes are costly: Earth observation, financial risk and critical infrastructure, anywhere teams already run several models and need to know which one is failing.

How it works, in one line

prediction = trusted forecast + gate × bounded correction. The gate can stay closed, so if a correction doesn't earn its place on held-out data, the trusted forecast stays the answer.

Explore

Scoreboard · Evidence · Arena · zenodivergent.dev