Methodology
Predictomat builds forecasting infrastructure in which AI performs structured subtasks within a deterministic framework. The framework itself - versioning, stage separation, factor economy, resolution tracking — is not model-generated. It is constructed and auditable.
Approach
Common approaches treat the language model as the analyst and rely on broad prompts to produce reliable output. Our approach inverts this. The model receives narrowly defined subtasks within an overall structure that is designed independently of the model. What changes as models improve are the individual results - not the methodology that governs them. Which specific LLM is used is, in this sense, a secondary question. The model is a tool within the framework, not its foundation.
Calibration
Every forecast is a verifiable event with a defined resolution point. The history of resolved forecasts yields a measurable calibration - a quantitative statement about how reliable the published probabilities actually are. This calibration is published on an ongoing basis.
Epistemic position
Forecasts are not statements about the future. They are disciplined probability estimates that must be tested against reality. AI extends the toolkit. It does not replace judgment about how a question is structured, nor does it carry the responsibility for the methodology in which it is used.
Responsibility for methodology, interpretation, and publication rests with the publisher, not with the model.
Current state of disclosure
The calibration record is being built. Individual components of the methodology are versioned internally and will be reflected in published forecasts as the resolved-forecast base expands. This page will be updated as that record grows.