Simple on the surface.
Careful underneath.

The question comes first.

What policy does this person advocate for advancing frontier AI capabilities? We first classify individual recommendations by activity, capability, actors, conditions and time. ASI restrictions, company commitments and deployment policies do not automatically establish a general frontier position.

Methodology 2.0 guides the current map. Earlier assessments retain their original wording and dates; corrections to our interpretation are recorded separately from changes in a person’s views.

Five categories.

Stop

Advocate an enduring prohibition on frontier development.

Pause

Suspend development for a fixed period or until evidenced conditions are met.

Pace

Continue development under advocated limits on pace or progression.

Continue

Affirmatively support proceeding without a proposed pace change.

Accelerate

Advocate increasing the pace of frontier capability development.

Placement within a category carries no extra meaning. We do not infer a pace recommendation from a prediction, a concern about risk, support for evaluation, or a person’s employer.

The qualification travels with the label.

Each position includes the context needed to understand it. We record what kind of AI development the person is talking about, any conditions attached to their recommendation, and the evidence supporting our classification. Each reviewed assessment identifies a representative passage and any additional supporting evidence. Source records preserve original links, publishers, passage locations, and provenance. We distinguish direct recommendations from reported or organizational statements. Completed Mixed, Unclear and Insufficient evidence outcomes appear separately under Unclear or mixed; Awaiting review means research or editorial review is unfinished. We revisit completed outcomes when new evidence, a methodology review, or an editorial correction warrants it. A category is used when the headline and its visible qualification are a fair, publicly defensible summary. Different scopes alone do not require Mixed; genuine competing prescriptions or unresolved material evidence can.

Scopes and conditions determine the headline.

An indefinite halt with evidenced reopening conditions is Pause. Stop needs affirmative support for an enduring prohibition; a missing restart clause is insufficient. A currently closed development gate is distinct from continuing under constraints. Future phases and contingent alternatives are evaluated separately from the operative recommendation.

A category is a fair summary of the evidence with its qualification, not a universal policy. Strong attributable company or capability-specific evidence can support it. Compatible research support and contingent safeguards remain visible without automatically creating Mixed. Important narrower exceptions can remain explicit under a best-fit headline: for example, a broad frontier Pause with a permanent ASI prohibition. Mixed applies when the prescriptions genuinely compete and choosing one headline would misrepresent the person even with a qualification. Substantial but unresolved relevant evidence is Unclear; absent qualifying evidence after research is Insufficient. None of these outcomes is a failure of research.

Shared bills, letters and proposals are compared clause by clause. Different headlines require supported differences in endorsement, scope or additional evidence. Independent review and explicit human approval remain separate from automated validation.

A coalition prohibition with necessary conditions for considering reopening can coexist with a stricter personal permanent ban. Those prerequisites do not by themselves recommend resuming development. Both positions remain visible in the qualification; an affirmative recommendation to resume the same development would still conflict with a permanent ban.

A new statement is not always a new position.

We only record a change in position when the evidence supports one. A later statement may clarify an earlier view, apply to a different scope, or reflect changed circumstances rather than a reversal. Each source keeps its original date and context so the history can be read accurately. Editorial reinterpretations are identified separately from changes in the person’s views. Unknown dates retain explicit bounds. Statement or publication dates are distinct from when we discovered the source and reviewed our interpretation. Historical assessments retain their own links; a newer statement does not automatically replace the current assessment.

A record should be correctable.

If a source, quote, date, interpretation, or classification looks wrong, send a correction with the relevant evidence. Submissions are reviewed before any change is made.