A probability from a screenshot — and a record of whether it was right.
Drop in a screenshot of a market page, a headline or a chart. You get a calibrated
probability in about thirty seconds, with nothing to fill in.
What makes it different from asking a chatbot
Anyone can produce a confident number. The question is whether numbers called 70% happen about
70% of the time. This is built so that question has an answer.
Every forecast is stored before the outcome is known, in an append-only
record. Nothing is edited later; a revision is a new row that keeps the old one.
It is scored when the question resolves — Brier score, log loss and
calibration error, broken down by category, never only pooled.
You see the disagreement. Seven independent samples produce the number,
and the spread between them is shown rather than averaged away.
It tells you when a question can never be scored instead of pretending
otherwise, and suggests a version that could be.
It does not copy the market. Shown the same page priced anywhere from 8%
to 92%, the forecast moves by about five percentage points — less than its own sampling
noise.
Sources are real. Ask for them and it searches the web, quotes the passages
it found, and forecasts again — so you can see what the evidence was worth.
What it will not do
Take odds or prices as an input, or convert them into a probability.
Calculate stakes, expected value, or tell you what to bet.
Forecast whether a named person dies, is harmed, or is arrested.
Answer questions about private individuals or their personal data.
Probabilities are the product. A recommendation layer on top of an unproven forecast just
launders a guess into advice.
Pricing
Free
€0 / month
One read every three days
Every caveat and the sample spread
Your calibration page in full
Standard
€44.99 / month
Ten reads a day
Web search with real citations
Multi-outcome markets
Four coloured themes
Premium
€79.99 / month
Unlimited reads
Deeper searches per question
Every theme
Export a read as JSON
Monthly, cancel any time from your account. Prices include VAT where it applies.
Said plainly: there is no track record yet.
The forecasts come from a language model reasoning from what it knows. The scoring apparatus
is built, tested and running — but it measures accuracy over time, and this is new, so
it has not measured much yet. Every forecast says so on its face. If you want proof of
accuracy before paying for it, wait; the calibration page is public to every account and it
will show the record honestly, including when the record is bad.
Contact
Questions, refunds or anything else: REPLACE@EXAMPLE.COM.
Refunds within 14 days of purchase, no reason needed.