OpenThomas

An autonomous agent that forecasts station temperatures and trades them on Kalshi and Polymarket. Every claim it makes is published before it settles, with the market price it was formed against and the observation that would prove it wrong.

Build in public, trade in public. You can watch it work at openthomas.com — open positions, live theses, realized P&L, Brier score, and what the thinking cost in tokens. The harness, the risk engine, and every strategy parameter are MIT-licensed at github.com/PredictionMarketTrader/openthomas.

What it thinks with

GLM-5.2 (NVFP4), self-hosted on our own GPUs. Open weights, so an agent whose entire argument is that the crowd is mispricing something can be checked all the way down to the model that formed the view.

The LLM is not the edge. It adjusts a statistical baseline — a seven-model NWP consensus, corrected for each station's learned bias — and it is clamped to that baseline ± a bounded delta. The model may never replace the statistics, only move them, and only for information it can name.

What lives here

openthomas/journal — every settled market the agent has traded: the first forecast it made, the price the market was offering at the time, and how the world resolved. Regenerated and pushed as it trades, so each commit is a timestamped record of what was known when.

Three rails are baked into the data rather than left to whoever trains on it: settled markets only, first forecast per market, and a temporal split carried in the rows. Each of them, forgotten once, produces a model that looks brilliant and is worthless.

What does not live here yet

Trained adapters. The journal is still filling. Fine-tuning on a few hundred settlements memorizes noise, so the agent uses Platt scaling until there is enough history for a weight update to mean anything. Publishing an adapter now would be a claim we cannot support, and this organization does not make claims it cannot support.

When they ship, each will carry the dataset revision it was fit on and its Brier score on days it never saw — including when it lost. The pipeline refuses to write a model card without them.

The self-improvement loop that will produce them is described in docs/RSI.md: OpenThomas improves along two artifacts. The harness — gate, risk engine, decision rules — ships to GitHub, where a diff is reviewable. The model ships here, because a weight change is not reviewable as text; the only honest way to publish one is alongside its training set and its held-out score.

Both face the same gate. A trained adapter is a candidate, not an authority.

Paper trading. Fills are simulated at the real bid/ask on live venue data. Prediction market trading can lose all the money you allocate; none of this is financial advice.