Point it at your evidence, ask something consequential, and it keeps thinking — re-running when the evidence moves, and taking a score when the world settles the matter.
Will Q4 demand exceed 12,000 units?
a filing lands · a number moves · a threshold breaks
It re-thinks when something changes. Not when a cron says so.
31 Dec · the world resolves
Every question resolves eventually. When it does, the engine is scored against what actually happened — and that record is on this page.
In most tools a question is a message: you send it, something comes back, it scrolls away. Here it’s an object you own — evidence attached, conditions that wake it, a history of what it used to think, and a score once the world resolves it.
Will Q4 demand exceed 12,000 units?
Sample question · resolving 31 Dec
A Postgres table. A folder of contracts. A vendor API nobody documented. Point at it, say what it is and how far you trust it, and it becomes evidence — traceable back to the row, the page, the paragraph it came from.
Your warehouse records what already happened to you. It has no opinion on a rate decision, a competitor’s filing, a tariff schedule or a wet October. Prolepsia reads those too — filings and registries, price series, official statistics, and what the market already thinks — and labels them the same way your own sources are: what it is, when it was published, how far to trust it.
And only what was knowable at the time. A figure revised later, fed backwards, makes the score further down this page a lie.
Set the conditions that would actually change your mind — a filing appears, a rate moves twenty basis points, your pipeline shifts five percent. A system that re-runs every morning regardless is one you stop reading by the second week.
Every run keeps what it broke the question into, what it retrieved and what it discarded, which models ran, and where they disagreed. Open any answer and walk backwards through it.
The disagreement is usually the interesting part.
The answers eventually come true, or they don’t — so the engine is measured. Brier score, log loss, calibration curve, performance against market consensus and against the dumbest baseline that works.
If the numbers get worse, they’re still here.
An agent will happily execute on a guess, at speed, and tell you it went well. Over MCP it gets a probability with an error bar instead — and the option to say the question isn’t answerable yet.
One tool call. Read the MCP docs →
A message to the people it affects, a webhook into your own services, a tool call your agent can act on, or a queue where a person signs off before anything moves.
Nobody has to log in and go look.
An LLM with a search tool gives you a confident paragraph, once, unscored. This returns a number rather than a mood, keeps the evidence that produced it, re-runs when that evidence changes, and is graded against reality afterwards.
You can check whether that is worth anything. You cannot check a paragraph.
Prolepsia is a forecasting engine for teams. You attach your own evidence — a database, a document store, an API — ask a question with a date it resolves on, and Prolepsia returns a probability and an expected value rather than a paragraph. The question stays open: it re-runs when the evidence moves, and it is scored against what actually happened once the date passes.
An LLM with a search tool answers once, in prose, and is never graded. Prolepsia returns a number with an error bar, keeps the evidence that produced it, re-runs when that evidence changes, and publishes its accuracy — Brier score, log loss and a calibration curve — against questions that have since resolved. A probability can be checked after the fact. A confident paragraph cannot.
Prolepsia does not require a clean warehouse. It takes a Postgres table, a folder of contracts, an undocumented vendor API or a hand-kept spreadsheet. You label what each source is and how far you trust it; answers stay traceable back to the row, page or paragraph they came from. Your own sources are only half of it — Prolepsia also reads the public record around them, because the evidence that moves a forecast is usually evidence you do not own.
Yes, and that is where most of the movement comes from. Internal data records what already happened to you; it has no opinion on a rate decision, a competitor's filing, a tariff schedule or the weather. Prolepsia reads outside sources alongside your own — filings and registries, price series, official statistics, and what the market already thinks — and labels each one the same way: what it is, when it was published, and how far to trust it.
Every source is read as of the date it was published, not as it reads today. Official statistics get revised, filings get restated and indices get backfilled, so a forecast re-run over history would otherwise score itself using figures nobody could have seen at the time. Prolepsia uses the vintage that existed on the day. Without that discipline a calibration record measures hindsight rather than accuracy.
On conditions you define, not on a schedule. A filing appears, a rate moves twenty basis points, a pipeline number shifts five percent — those wake the question. A system that re-runs every morning regardless is one people stop reading by the second week.
Every question has a resolution date. When it passes, the forecast is scored against the outcome using Brier score and log loss, plotted as a calibration curve, and compared against market consensus and against a naive baseline. The record is published, including the periods where it gets worse.
Yes. Prolepsia exposes its forecasts to agents over MCP (Model Context Protocol) as a single tool call. The agent receives a probability with an error bar, and can be told that a question is not answerable yet — instead of acting at speed on an unmarked guess.
In your systems. A message to the people a forecast affects, a webhook into your own services, a tool call an agent can act on, or a sign-off queue where a person approves before anything moves. Nobody has to log in and go look.