Four kinds of claim
Observed findings describe measured performance or behavior under specified conditions. Provider reports describe what a company says it observed; they are labeled as such. Causal analyses connect mechanisms to possible outcomes. Policy proposals argue for particular actions.
Keeping these categories visible makes the argument easier to challenge. A laboratory demonstration can establish that a behavior is possible in that setting without establishing its frequency in the world. A severe possible consequence does not by itself supply its probability.
How sources are selected
The evidence library prioritizes original research, evaluation reports, and first-party incident accounts. Each entry records the conditions, limits, significance, and evidence that could change the assessment. The collection is curated, not an exhaustive systematic review.
Provider observations remain distinguished from independent evaluations. Findings that complicate the argument belong in the collection too. Dates refer to the source or experiment, while the library separately identifies the review date.
What Rollback Lab calculates
The technical model compares assumed actor resources with assumed costs of rebuilding capability through new training, retained The learned numerical parameters of a model. A copy can support another deployment when software and hardware are available., Training another model from a stronger model’s outputs or behavior, transferring some of its capabilities., or system orchestration. It runs a central trajectory and a 27-case stress envelope.
The frontier reference is an index, not an empirically established ASI threshold. Model cycles are not years. Stress-case counts are not probabilities. Resource costs, policy effectiveness, efficiency gains, and capability coefficients are scenario assumptions. Passing the screen identifies a conditional candidate, not proof of safety.
What the human-outcome model calculates
The story distinguishes survival, authority, and reciprocal necessity, using conditional state transitions. The opening hazards are explicit illustrative premises. They are not estimated from incident counts or surveys.
The long-run limit extends the specified process indefinitely. A persistent positive loss hazard can drive retention toward zero. A sufficiently fast vanishing hazard or a genuine end to exposure changes the result. Extinction after loss of authority requires its own additional hazard; authority loss alone does not mathematically establish extinction.
What the visualizations mean
The story’s geographic displays illustrate conditional pathways. Geography is based on Natural Earth, while progression, timing, and regional effects are scenario assumptions. The map supplies no additional probability estimate.
The dependency explorer is a qualitative guide to physical and institutional requirements. The handover tool isolates authority retention under three explicit risk patterns. Its transitions are neither calendar years nor model calls.
Updating the argument
When new evidence changes a claim, the affected explanation and its source entry should change together. Substantive changes belong in the dated updates and corrections log, with the prior claim and reason for revision.
The “What would change our assessment?” page identifies empirical and policy developments that could weaken the argument or require a different remedy.