---
title: Under the bonnet: Noah Predict
description: What your AI gets when Noah is connected: a dated archive of 49 million public documents, deterministic forecasting physics that runs before any language model writes, five shapes of answer, and back-tests as of any past date.
url: https://noahpredict.com/how/
source: Noah Predict, Worldwide AI Media Ltd
last_updated: 2026-10-02
---

# Under the bonnet

## The archive

Forty-nine million public documents, read as they appear, kept with their date, source and language. 1.6 million sources, many languages, read continuously. A language model remembers the world loosely and without dates; Noah keeps it precisely, so a question about next quarter is answered against what was known, and when. The archive is not a search index: every document is tagged on arrival (who, where, what kind of event, which market, which risk), which is what lets the physics compare this week with the same week last year, or one country with its neighbours, in numbers rather than impressions.

What goes in: news and specialist media; company filings and results; regulators, courts, parliaments; markets, prices, prediction markets; public datasets and official series; fifteen months of instrumented history. Public record only.

## The physics

Deterministic Python runs before any language model writes a word. The model explains the numbers; it does not invent them.

1. Tagging: every document is classified on arrival, actors, places, event types, markets and risk areas, with its date kept.
2. Pressure: how much signal is building around a subject, compared with its own history and with its peers.
3. Regime: whether a subject is in a calm, building or breaking state, and how long such states have lasted before.
4. Couplings: which subjects move together, so a shock in one place can be read across to the others it drags.
5. Calibration: probabilities are scored against resolved questions, so a 70 means something close to 70.
6. Evidence kept: every number carries the documents it rests on, with dates, so the answer can be checked rather than believed.

## Five shapes of answer

You ask in plain words. Noah chooses the shape, runs the physics and hands the result to your AI with the evidence attached.

- Risk of: how risky is it? A place, a sector, a counterparty or a journey, this quarter and next.
- Probability of: how likely is it? A named event by a named date, as a probability with its range, drivers and counter-evidence.
- Outlook for: where is it heading? A subject month by month, with the watchpoints that would change the path.
- Edge on: is the market right? Noah's reading beside the market price, where they disagree and why.
- Opportunity in: where is it moving? A scan across a sector for subjects gaining signal before the price has noticed.

And then: back-test any of them as of a date in the past.

## As of any past date

Name a date. Noah limits itself to what was in the archive by then, runs the same physics, and gives you the forecast it would have given. Then compare it with what happened. Back-tests use the full engine, not a thinned one; the only thing that changes is the cutoff. This is how you test a theory before you trust it, and how you test Noah before you pay for it.

Your AI does the reasoning: the bigger the model, the better the forecast. Private by design: read-only, your questions stay with your AI; nothing you ask becomes Noah content.

Connect: https://noahpredict.com/connect/ (Claude) · https://chatgpt.com/plugins/plugin_asdk_app_6a44a9d050d081918636f82d3a5a5d4f (ChatGPT) · https://noahpredict.com/mcp (any MCP client)