Structured judgment, not prophecy

For the day-to-day pipeline (daily scan, weekly re-run, how new scenarios appear) see How it works. This page is the reasoning behind the rules. Nobody can predict AGI. What can be done is to lay out the branches explicitly, attach observable conditions to each, update the estimates in public as evidence arrives, and be scored on the parts that resolve. This page explains how, and where the limits are.

1. The scenario space

Every scenario is a point in a four-axis space: speed of capability growth (plateau, gradual, fast, discontinuous), concentration (one lab, few labs, open, state), alignment outcome (controlled, misaligned-contained, misaligned-uncontained, uncontrolled) and societal absorption (adaptive, disruptive, collapsed). Root scenarios are the 8 narratives people can hold in their head; branches are specific ways each plays out. Good / bad / ugly is a label on the outcome for most people, not part of the structure: the same branch can be good for some actors and ugly for others.

The tree is unbounded. Every weekly run asks whether any signal opened a branch that does not exist, and creates it. A region of the axis space with no scenario is itself a finding.

2. Probabilities

Root probabilities answer: what is the chance this is the dominant trajectory through 2035? They are meant to be roughly exhaustive and sum near 100. Branch probabilities answer: what is the chance this branch occurs by its horizon? They are not exclusive and do not sum to anything.

Every probability is a range, never a point, because the honest uncertainty is wide and a point number would be false precision. Every range has a full history; every change carries a written reason and the signals that drove it. Weekly moves are bounded (normally ±5 points on the midpoint); a larger move is allowed only with a "shock" flag and explicit justification. Estimates are anchored to outside references where they exist: Metaculus, the Forecasting Research Institute's LEAP surveys, the AI Futures Project, Manifold, and published expert estimates. Where we disagree with them, the scenario page says why.

3. Signals

A daily run scans for developments across model releases, benchmarks, agentic milestones, compute and capital, safety incidents and evaluations, governance and regulation, open-weight releases, lab governance and economic data. Each signal is dated, sourced, scored 1–5 for magnitude, tagged with direction on the three axes (timelines, concentration, safety), and linked to the scenarios it pushes. Signals are evidence, not conclusions; the weekly run decides what they mean for probabilities.

4. Tripwires and status

Each scenario carries observable thresholds. A scenario is possible by default, watch when evidence is accumulating or one tripwire has crossed, active when its preconditions are currently met, and resolved when its horizon passed. A crossed tripwire brings the weekly re-run forward and pins the scenario to the Right now page.

5. Playbooks

Every scenario has four sections: Prevent (what lowers its probability), Detect (what tells us we are in it), Respond (first actions for individuals, organizations and governments) and Recover. Good branches also carry Steering signals: what moves probability toward them. Actions that recur across many playbooks are promoted to the Standard.

6. Sources

Every link reviewed in any run is recorded in the public ledger with date, publisher, what it says and what rests on it. Claims that cannot be traced to the ledger are bugs. Press reports of unpublished documents are marked as such.

7. Calibration (from Phase 2)

Long-horizon probabilities cannot be scored soon, so we also publish near-term sub-predictions with resolution dates and score them (Brier) when they resolve. The calibration record will be public. Until it exists, treat our numbers as a structured, sourced opinion.

8. What this is not

Not a forecast of a date for AGI. Not investment, legal or security advice. Not neutral on outcomes: the site exists to make bad branches less likely and good ones more likely, and says so. Not finished: the map is wrong in places we have not found yet, which is why contributions are open.

9. How the site runs

All content is data: JSON files in a public repository. A daily scheduled task adds signals and sources; a weekly task re-evaluates scenarios, updates the trajectory forces, drafts the newsletter and creates new branches; a monthly task re-baselines against outside forecasts. Each commit rebuilds the site. The full dataset is available at /data.json.