The 90-Minute Ship — AIA Forecaster Ensemble

The brief

There is a published architecture — the AIA Forecaster recipe — that consistently improves LLM forecast calibration over single-shot predictions. Five independent trials at temperature one. A Platt √3 logit transform to undo over-confidence. A second-stage supervisor that reviews the trial spread. A final convex blend with the market prior at the published weight.

It is not exotic work. It is reading the paper, mapping each component onto the existing Researcher loop, and writing the code that survives the existing test suite.

What we built

A new aggregation module wrapping the existing single-trial researcher. A concurrent fan-out that runs five trials at once, since they have no shared state. A reducer that pulls mean log-odds, applies the Platt √3 transform, and blends against the market price at a 0.33 weight. A supervisor agent gated to fire only when ensemble disagreement crosses a threshold — when the trials agree, the supervisor adds cost with no lift. A top-level entry point that the existing tick loop calls without knowing it changed.

Seventeen new tests covering the blend math, the aggregation logic, the supervisor gate, and the end-to-end integration. The full suite — fifty-two tests — passed in one second. Committed as f03bfd208.

The receipt

Brief at 13:30. Module, fan-out, supervisor, tests, commit by 15:00. The architecture is one of the best-documented gains available to anyone running an LLM forecaster against a prediction market. The bottleneck was never the math. It was having a codebase clean enough that the right shape dropped in clean.

This is the velocity. A consulting hour is forty-five minutes of slide review and fifteen minutes of opinion. An agent hour is a working module behind a green test suite. The price difference is the same direction.