Methodology

Ask a single large language model what will happen in the next round of US-China trade talks and you will get one answer. It will be fluent, plausible, and the weighted average of everything the model has read about trade negotiations. It will have no adversaries, no skeptics, and no surprises.

This is the failure mode of single-agent forecasting. The model is often directionally right, but it is smooth where reality is rough. The hard parts of prediction live at the edges of distributions, in the places where different credible observers would disagree.

Multi-agent systems are built to preserve cognitive diversity on purpose. Each agent is given its own role, incentives, and style of reasoning. They argue against each other, and from that argument something emerges that no individual agent held at the start.

The Probability Lab's Advocate / Skeptic / Judge pattern is borrowed from peer review, adversarial legal proceedings, and scientific falsification. Each tradition shares a conviction that truth is best approached by forcing a claim to survive an attack rather than by asking whether it sounds plausible.

In practice, the Advocate builds the strongest possible case that a scenario will happen. The Skeptic dismantles it. Only after both have been given their best shot does the Judge render a conditional probability.

Large language models are not deterministic in practice. Tiny changes in context ordering mean that the same prompt can walk different paths and arrive at different conclusions. For a single-answer question, this is annoying. For a negotiation, where outcomes are path-dependent and small early shifts cascade, it is diagnostic.

Run a scenario once and you have an anecdote. Run it fifty times and you have data: a deal rate, a distribution of rounds to resolution, a cluster of consensus terms, and a taxonomy of failure modes.

Aurelon does not claim to predict the future. What it produces are calibrated probability estimates, numbers that are meant to move when new evidence arrives, and that are meant to be judged by their track record across many questions rather than by whether any single call came out right.

• LLMs have training cutoffs.
• Genuinely novel scenarios are harder to reason about.
• Rare events are systematically underweighted.
• Human judgment should always remain in the loop.

Aurelon is a tool for clearer thinking, not an oracle. If it ever starts to feel like an oracle, that is a signal to be more skeptical, not less.