Methodology

The Time Tree tracks uncertain future events across seven domains. Its credibility rests entirely on being transparent about how — so this page is a first-class part of the site, not a footnote.

How events are defined

Every tracked event has a falsifiable definition: precise enough that a reasonable person can determine in hindsight whether it happened. "AI will transform the economy" doesn't meet the standard. "Software performs the majority of analysis, drafting, and decision-support tasks in a major economy's white-collar workforce" does — it's still uncertain, but it can be confirmed or refuted.

The definition also forces honesty about what we actually mean. Vague language about "AGI" or "climate tipping points" masks fundamental disagreements about what would count as the event. Making it precise surfaces that disagreement rather than hiding it.

Domains

Events are organized into seven domains:

• AI & Computing — artificial intelligence, large-scale automation, machine reasoning
• Robotics & Automation — physical robots, autonomous systems, manufacturing
• Biotech & Health — gene editing, longevity, brain-computer interfaces, biosecurity
• Energy & Engineering — fusion power, advanced infrastructure, next-generation energy
• Climate & Environment — tipping points, decarbonization, adaptation
• Geopolitics & Governance — great-power conflict, institutional fracture, binding international agreements
• Space — off-world settlement, deep space exploration

Domains are not mutually exclusive — many events cut across several. The domain label reflects the primary driving force, not an exhaustive categorisation.

Probability estimates

Each event carries a probability model with three components:

• P10 year — the optimistic but plausible bound: there is roughly a 10% chance the event has occurred by this year.
• Median year — our current best estimate of when the event is most likely to occur.
• P90 year — the pessimistic but plausible bound: roughly a 90% chance it has occurred by this year (conditional on it happening at all).
• Probability ever occurs (p_ever) — the total probability mass: the likelihood that the event happens at all within any meaningful horizon. For events that are effectively inevitable — AGI, mass automation — this is 1.0. For contingent events — an engineered pathogen outbreak, a binding governance pact — it sits well below 1.0.

From these four numbers we derive a full log-normal probability distribution, which is displayed as two curves on every event page:

• The probability density curve (shaded area) shows where probability mass is concentrated — the years in which the event is most likely to happen if it happens.
• The cumulative probability curve (solid line) shows the probability that the event has already occurred by any given year. For contingent events this curve plateaus below 100%, at the p_ever ceiling.

These estimates are editorial judgments grounded in primary sources — forecasting aggregators (Metaculus, Manifold), peer-reviewed research, industry roadmaps, and expert assessments. They are not outputs of a calibrated forecasting model. Treat them as honest best guesses, not precise predictions.

How events get added and retired

An event qualifies for the taxonomy if it meets three criteria:

1. It's falsifiable — the definition can be confirmed or refuted in hindsight
2. It has civilizational-scale consequences — it materially changes the context for a significant portion of humanity
3. It's genuinely uncertain — neither near-certain nor negligibly unlikely over the relevant horizon

Events are retired when they occur (marked with a date), when they become near-certain enough to be more useful as context than as tracked uncertainty, or when the probability becomes low enough that the uncertainty range is no longer meaningful.

Revisions

Estimates are revisited when significant new evidence arrives — a major research paper, a forecasting shift, a real-world milestone. When a median year or p_ever changes materially, we note the revision and explain the reasoning. Forecasts that are never revised are forecasts that aren't being taken seriously.

Correction policy

When we get something wrong — a source cited incorrectly, an estimate that was clearly off — we correct it and note the correction rather than quietly editing the record. The revision history of our estimates is part of the product.

What this site doesn't do

We don't predict outcomes, recommend positions, or advise decisions. The site is a tracking tool, not an oracle.

We don't claim our probability estimates are calibrated in a rigorous statistical sense. Forecasting aggregators (Metaculus, Manifold, Good Judgment) do calibrated crowd forecasting well, and we link to them where relevant. Our estimates are honest editorial judgments, not the output of a proper forecasting infrastructure.