Methodology

What these numbers are

Every number on this site is a reasoned, revisable estimate with its evidence shown — not a statistically scored probability. Call these scorable forecasts: each one names an observable condition that will eventually make it right or wrong, so its track record can be checked once enough claims have resolved. With a few dozen live claims, no single number on this site has a meaningful track record yet. That comes from volume and time, not from how confident any one estimate looks today.

Three horizons, not one

Each claim is estimated at three shared dates — 2035, 2050, and 2100 — rather than at a single date chosen per claim. Sharing the same three horizons across every row is what makes the dashboard directly comparable, cell to cell: p(2050) for one claim means the same kind of thing as p(2050) for any other.

A horizon cell reading "not yet estimated" has no forecast for that date at all — treat it as unknown, never as zero. A cell reading"no evidence yet" has a number but no researched source behind it yet; that is what an initial prior looks like before research starts, and it is marked distinctly from a researched estimate for exactly that reason. Every populated horizon carries its own resolution criterion and its own evidence, because the bar for "true by 2035" is often a different observable condition than the bar for "true by 2100" on the same underlying claim.

Two kinds of claim

A claim is either an event or a level, and the two curves mean different things:

• An event claim is cumulative — "X has happened by year Y." Once true it stays true, so its probability can only rise or hold as the horizon extends: p(2035) ≤ p(2050) ≤ p(2100). This site enforces that ordering automatically and refuses to publish a curve that decreases.
• A level claim is point-in-time — "measure M is above or below threshold T in year Y." A level can rise and then fall, so a later horizon reading lower than an earlier one is expected, not an error. No ordering is enforced on these.

Each curve is labelled with which kind it is, directly above the three horizon cells.

The low–mid–high range

The midpoint is the single best estimate. Low and high describe a plausible range around it, not a formal confidence interval from a statistical model — unless a statistical model actually produced the number, which will be stated explicitly when it happens. A very small estimate is shown as a readable bound such as "<0.1%" while the underlying value is kept in the machine-readable export.

How a forecast resolves

Each populated horizon states, in advance, the observable condition that would make it true. Resolution criteria are written before the eventual outcome is known and are not loosened or tightened after the fact — a change to a resolution criterion is itself a correction, handled the same way as any other.

How sources are selected

Sources are primary research, regulatory or authoritative guidance, and similar first-party material, each logged with its title, host, access date, source type, and a verification status describing whether it was fetched and confirmed directly or retrieved by a delegated search. Each source records the specific claim it supports and, where relevant, an explicit limit on how far that claim can be stretched — a result in a clinical population is not silently generalized to a healthy one, for example.

Versioning and corrections

Every revision is its own permanent, dated record, chained to the one it replaces. Nothing is edited in place: a correction creates a new record and the page shows the full chain, so a reader can see exactly what the estimate was at any prior point and why it changed.

Why this exists

Most public predictions are asserted, not audited. This site instead shows the resolution criteria, the evidence on both sides, and the full revision history behind every number, so the reasoning can be checked rather than taken on trust.