The xrtm Manifesto
If you are new here, prove the product path first in Getting started. This page explains the philosophy behind that workflow.
"To understand reality is to predict its future."
The Core Issue
Today's AI systems (LLMs, Agents) are built and evaluated on Fixed Knowledge.
Current models are tested on facts that already exist. This leads to contamination—we never know if a model is reasoning or just recalling training data.
An AI that scores 100% on a history exam hasn't proven it is intelligent; it has proven it has a memory.
The Paradigm Shift
We believe Forecasting Future Events is the only rigorous test of intelligence.
You cannot train on next week's news. A prediction about the future is the ultimate zero-shot test because the ground truth does not exist at the moment of inference.
The Standard: If an AI cannot predict what happens next, it does not truly understand what is happening now.
The Solution
xrtm provides the vertical infrastructure to shift AI from "Knowledge Retrieval" to "Future Prediction."
- Inference: A runtime optimized for probabilistic reasoning, not just text generation.
- Evaluation: Measuring success via calibration (knowing what you don't know) and resolution (accuracy), not semantic similarity.
- Training: Optimizing models to internalize causal chains and Bayesian logic.
The Ecosystem
We organize our Open Source Software (OSS) into a closed loop that enforces temporal integrity.
| Component | Responsibility |
|---|---|
| xrtm-forecast | The Runtime. The engine that generates predictions. It enforces rigorous reasoning, requiring the model to output causal graphs and confidence intervals before committing to a probability. |
| xrtm-data | The Snapshot. The vault that freezes the world state at specific times. It ensures we can re-run history without "future leakage" contaminating the test. |
| xrtm-eval | The Judge. The verifier that scores predictions. It handles the necessary time delay between a forecast and its resolution, using strict scoring rules (Brier, Log Loss) to measure reality. |
| xrtm-train | The Learning. The optimization layer. It uses the gap between prediction and reality to train models on causality and calibration. |