A clinical instrument: an actuarial register of published judgments.
This site publishes an index: a number built from other people’s published numbers. Named researchers, surveys, and forecasting tournaments have written down probabilities. The site mixes those judgments, applies a small documented adjustment for current news, and stamps the result with a date.
This is an index of published judgments, not a measurement of the future. You cannot read “chance of catastrophe” off a meter. The honest product is a mix you can reproduce, with every site choice written down.
A published judgment is a number a named person or instrument stated in a citable source (a book, a paper, a survey table, a recorded interview).
A news rubric is a written list of which current events may nudge the index, and by how little. Vibes do not count.
People and instruments who published a number. On the record, dated, cited.
The site mixes eligible numbers. You still see the disagreement. A bare % is not allowed.
News can move the mix a little. Hard stops on both sides.
The site publishes it with an as-of date. No live ticker.
The site keeps three gauges. All three concern harm caused by AI. The headline mix uses the year 2100.
G1, catastrophe is the headline. It reports the probability of an AI-caused event that kills at least 10% of humanity, or irreversibly collapses industrial civilization, by 2100.
That definition is broader than extinction. A collapse that leaves people alive still counts.
G2, extinction is a sub-gauge. Literal human extinction from AI, by 2100. Shown beside G1. Kept out of the headline mix.
G3, mundane-deployment cascade is first-class. It stays on the page. It tracks premature deployment of not-ready AI into critical systems (grid, military command, health, finance, logistics) compounding faster than verification, liability, or regulation can catch it.
G1 and G3 stay separate. When the ledger has no published forecast of this exact event, the site labels G3 constructed: a site-built picture of the thesis. Constructed is not an aggregate of published numbers. The gauge is drawn differently so it cannot wear a percentage that pretends to be a mix of sources.
AI-caused. At least 10% of humanity dead, or irreversible industrial collapse. By 2100. This is the number on the front of the site.
Literal human extinction from AI, by 2100. Beside the headline. Kept out of G1.
Not-ready AI pushed into critical systems faster than oversight. Collapse-shaped. On every snapshot.
Four steps, in this order.
Collect. Take named people and instruments who published a number that matches this question (catastrophe, this horizon). Books, papers, expert surveys, forecasting tournaments. A paraphrase, an unverified quote, and a different question stay off the mix.
Mix. Turn the eligible numbers into one point. Draw a band around it for how far those sources disagree. The point without the band is not allowed.
Nudge. A news multiplier may move the mix inside tight bounds. The mix is the expert baseline. The nudge is current conditions. You can read both.
Stamp. Publish a snapshot with an as-of date.
Sources that do not qualify still sit on a public ledger, the full list of candidates. Inclusion on the list is not inclusion in the mix.
The news multiplier lets current events change the index. They sit inside a bound smaller than the disagreement already on the record.
Published experts already disagree by a wide margin. News may move the index by a fraction of that disagreement. The dial has hard stops.
Every movement is a row on a public event ledger: date, what happened, which way, how much. “Nothing moved” is also a row.
News cannot double the number or cut it in half. A vendor demo does not score. An independent, dated, checkable fact might.
Every scored number cites a named source. Every movement cites a logged event. If you cannot point at the row, it does not belong in the mix.
A point with no band is not publishable. Disagreement is part of the number. Dropping the band would be an editorial act.
A loud figure that cannot be pinned to a primary quote still sits on the ledger. Shown is not scored. The list is not the statistic.
When there are no published forecasts of that event, the site does not fake an aggregate. Constructed stays labeled constructed.
The index is not a prediction market, a personal p(doom), or a countdown. No ads.
The full protocol (eligibility, mixing, bounds, ledgers) is methodology v0.1. This page is the public whiteboard. That document is the instrument manual.