METHODOLOGY 0.3

Rules before conclusions.

AI Risk Trajectory is designed to make overstatement harder. The system tracks causal stages and evidence, not a single probability of catastrophe.

Assessment dimensions

Realisation records how far a pathway has actually progressed in evidence. Exposure records real-world opportunity. Consequence Envelope records the credible scale of a fully realised pathway without implying likelihood. Control Assurance records how strongly safeguards are demonstrated. Evidence Profile keeps volume, independence, consistency, directness, replication and external validity separate.

Transition rule

A new paper, model release, incident or policy can enter the evidence base without moving a classification. A state changes only when its published transition condition is met. Routine evidence reviews cannot silently rewrite the ontology or methodology.

Uncertainty

Unknown, contested and insufficient-evidence states are valid outputs. The model does not create a probability where evidence cannot support one. Supporting evidence, limiting evidence and evidence gaps remain visible.

Weekly review

Each weekly review starts at the previous evidence cutoff, identifies new candidate evidence, maps it to existing objects, actively searches for contrary evidence, tests transition criteria, records accepted and rejected changes, validates the new snapshot and publishes an auditable changelog. A valid weekly result can be “no material change”.

AI use

AI can assist evidence discovery, classification and explanation. The public Ask layer is read-only. Material scientific state changes should require human review in the mature system.

Current limitation

This research product has not yet undergone external peer review. Its classifications are intended to be inspectable and challengeable rather than presented as institutional consensus.

See the current assessment and dataset summary.