# YarrowLab

Research and models for event forecasting.

## Models

- [`yarrow-forecast`](https://huggingface.co/YarrowLab/yarrow-forecast) —
  forecasts U.S. party primary and nomination outcomes. Given a question,
  its resolution criteria, and the evidence available as of a stated date,
  one forward pass returns a calibrated probability in [0, 1].
  Out of sample on 150 settled questions across 33 election nights, Brier
  score is 0.175–0.185, against 0.239 for the historical base rate.
  [CC BY-NC-4.0](https://creativecommons.org/licenses/by-nc/4.0/).

## Research

Post-training work on **bounded deviation from a statistical prior**.
Instead of asking a model to output a probability directly, we train it to
output `Δ` — a bounded deviation from a prior `p₀` computed upstream from
reference-class statistics — plus an attribution category for *why* the
prior might be wrong, and an evidence span supporting that attribution.
The final probability `p = σ(logit(p₀) + Δ)` is composed in code; the model
never sees or touches market prices.

- [`yarrow-delta-sft-v1-smoke`](https://huggingface.co/YarrowLab/yarrow-delta-sft-v1-smoke) —
  a pipeline-validation checkpoint (LoRA over a randomly initialized tiny
  model). Confirms the SFT loop runs end to end; not a trained model.