The method Sign in

Show the work, or it doesn’t count

No black boxes.
Read every step.

Most AI answers are a single model’s first draft. A New Priors forecast is an adversarial process: independent analysts research live evidence, a red team attacks their cases, judges score what survives, and transparent math — not another model — produces the final number. All of it is in the report you get.

Run a forecast Read a real report

The pipeline, end to end

1

Resolution first

An analyst pins down exactly what counts as YES, the resolution source, and the deadline before anyone forecasts. Vague questions make dishonest forecasts.

2

Independent panels

Multiple frontier models run as separate forecasters, each researching live primary sources on its own. No forecaster sees another’s work — so agreement means something.

3

Red team & judges

Adversarial reviewers attack every case; judges score what survives. A forecast that can’t defend its reasoning doesn’t ship.

4

Deterministic aggregation

Panel views pool through published, reproducible math. Independent validators recompute the headline from its own tables — divergence gets rejected, not rounded.

What lands in your report

Every analyst’s case

Each forecaster’s full argument with the sources it actually read — cited and linkable, so you can check the load-bearing claims yourself.

The attacks it survived

The red team’s strongest objections and how each case held up. You see what almost changed the number, not just the number.

The math, reproducible

Aggregation tables you can recompute by hand: forecaster views, weights, pooled result, confidence band. The headline follows from the table or it fails validation.

Why this design

One model overcommits

Any single model anchors on its first frame. Independent forecasters with opposing briefs surface the considerations one pass misses — and the spread between them is reported as honest uncertainty.

Fresh evidence wins

Training data ends; the world doesn’t. Forecasters search the live web at run time, so the probability reflects this morning’s filings and headlines, not last year’s memory.

Models don’t do arithmetic

The final number is never “whatever the model said” — it’s computed outside the models from their structured outputs, then checked against them. Trust the process, verify the math.

Read the reasoning. Keep the probability.